pattern,provider,model_class,basis,size_or_tier_note,own_factor_model_id,reasoning_default,source,assigned,notes gpt-4o-mini,OpenAI,A,tier,Mini = efficiency tier,,no,"main.tex Table S20 (Supplementary S6), v3",2026-09,Table S4 row 1 gpt-4.1-mini,OpenAI,A,tier,Mini tier of 4.1,,no,"main.tex Table S20 (Supplementary S6), v3",2026-09,Table S4 row 2 gpt-4.1-nano,OpenAI,A,tier,Nano tier,,no,"main.tex Table S20 (Supplementary S6), v3",2026-09,Table S4 row 3 gpt-5-mini,OpenAI,A,tier,Mini tier of GPT-5,,yes,"main.tex Table S20 (Supplementary S6), v3",2026-09,"Table S4 row 4; the GPT-5 family reasons by default, so a Tier 2a billed output count already includes the hidden tokens and must not be multiplied" gpt-5-nano,OpenAI,A,tier,Nano tier of GPT-5,,yes,"main.tex Table S20 (Supplementary S6), v3",2026-09,Table S4 row 5 o1-mini,OpenAI,A,tier,Mini tier; reasoning raises tokens per interaction and not the per-token class as for o3-mini,,yes,"main.tex Table S20 (Supplementary S6), v3",2026-09,Table S4 row 6 o3-mini;o4-mini,OpenAI,A,tier,"Mini tier; reasoning raises tokens per interaction (multiplier 10-50x, mid 25, author assumption), not the per-token class",,yes,"main.tex Table S20 (Supplementary S6), v3",2026-09,Table S4 row 7 gpt-4o;chatgpt-4o-latest,OpenAI,B,tier,"Mid price tier (USD 2.50 in / 10.00 out per M); size undisclosed; energy reported by Jegham et al. would place it at or above Class C, so B is a low-side assignment",,no,"main.tex Table S20 (Supplementary S6), v3",2026-09,Table S4 row 8 gpt-4.1,OpenAI,B,tier,Successor of 4o at the same tier (USD 2.00 / 8.00 per M),,no,"main.tex Table S20 (Supplementary S6), v3",2026-09,Table S4 row 9 gpt-4;gpt-4-turbo,OpenAI,C,tier,"Frontier closed model, no disclosed size",,no,"main.tex Table S20 (Supplementary S6), v3",2026-09,Table S4 row 10 gpt-5,OpenAI,C,tier,Frontier closed model (USD 1.25 / 10.00 per M),,yes,"main.tex Table S20 (Supplementary S6), v3",2026-09,"Table S4 row 11; as a prefix this also covers gpt-5.1, 5.2, 5.4, 5.5 and 5.6 and their -chat, -codex and -pro variants, which sit at or above the flagship tier of their generation" o1;o3;o3-pro,OpenAI,C,tier,"Full-size o-series reasoning models (o1 USD 15.00 / 60.00, o3 USD 2.00 / 8.00 per M); hidden tokens 10-50x visible, mid multiplier 25 (author assumption)",,yes,"main.tex Table S20 (Supplementary S6), v3",2026-09,Table S4 row 12 claude-haiku-*;claude-3-haiku*;claude-3-5-haiku*,Anthropic,A,tier,"Haiku = small tier (claude-haiku-4-5 USD 1.00 / 5.00 per M, the cheapest Anthropic tier)",,no,"main.tex Table S20 (Supplementary S6), v3",2026-09,Table S4 row 13; price confirmed 2026-10-02 against the Anthropic pricing page claude-sonnet-*;claude-3-sonnet*;claude-3-5-sonnet*;claude-3-7-sonnet*,Anthropic,B,tier,"Sonnet = mid tier (claude-sonnet-5 USD 2.00 / 10.00 per M); energy reported by Jegham et al. would place it at or above Class C, so B is a low-side assignment",,no,"main.tex Table S20 (Supplementary S6), v3",2026-09,Table S4 row 14 claude-opus-*;claude-3-opus*,Anthropic,C,tier,"Opus = frontier tier (claude-opus-5 USD 5.00 / 25.00, claude-opus-5-5 USD 4.00 / 20.00 per M)",,no,"main.tex Table S20 (Supplementary S6), v3",2026-09,Table S4 row 15; Opus is no longer the TOP Anthropic tier: Fable and Mythos are priced above it (see the claude-fable-* row) gemini-*-flash;gemini-*-flash-lite,Google,A,tier,Flash = efficiency tier; Flash-Lite below it,,,"main.tex Table S20 (Supplementary S6), v3",2026-09,"Table S4 row 16. FLAGGED: later-generation Flash models are priced at the level of an earlier Pro (gemini-3.5-flash USD 1.50 / 9.00 per M against gemini-2.5-pro USD 1.25 / 10.00), so the tier word may understate them" gemini-*-pro,Google,B,tier,"Pro = mid tier (gemini-2.5-pro USD 1.25 / 10.00, gemini-3.1-pro USD 2.00 / 12.00 per M)",,,"main.tex Table S20 (Supplementary S6), v3",2026-09,"Table S4 row 17; a text Pro exists only for generations 2.5 and 3.1 as of 2026-10-02, generations 3.5 to 3.8 being Flash-only" gemini-ultra,Google,C,tier,Ultra = frontier tier,,,"main.tex Table S20 (Supplementary S6), v3",2026-09,Table S4 row 18. FLAGGED: Google's current model list has no tier above Pro and no Ultra model; this row covers historical identifiers only llama-3.1-8b;llama-3.2-1b;llama-3.2-3b;llama-3.2-11b*,Meta,A,size,Open 1-11B dense (<=15B),meta-llama/Llama-3.1-8B-Instruct,no,"main.tex Table S20 (Supplementary S6), v3",2026-09,Table S4 row 19; the 8B variant is the Class A calibration model gemma-*-2b;gemma-*-4b;gemma-*-9b;gemma-*-12b,Google,A,size,Open 2-12B dense (<=15B),google/gemma-3-12b-it,no,"main.tex Table S20 (Supplementary S6), v3",2026-09,"Table S4 row 20; own factor applies to gemma-3-12b-it, the measured member. FLAGGED: there is no Gemma 3 2B release (the sizes are 270M, 1B, 4B, 12B and 27B), so the 2b alternative is vestigial" llama-3.2-90b*;qwen2.5-72b,Meta/Alibaba,C,size,"90B dense (vision) and 72B dense, above 70B",,no,"main.tex Table S20 (Supplementary S6), v3",2026-09,Table S4 row 21 gemma-*-27b,Google,B,size,Open 27B dense,google/gemma-3-27b-it,no,"main.tex Table S20 (Supplementary S6), v3",2026-09,"Table S4 row 22; own factor applies to gemma-3-27b-it, the measured member" llama-3.1-70b;llama-3.3-70b,Meta,B,size,Open 70B dense (upper edge of B),meta-llama/Llama-3.3-70B-Instruct,no,"main.tex Table S20 (Supplementary S6), v3",2026-09,Table S4 row 23; model_factors.csv carries both Llama-3.1-70B-Instruct (the Class B calibration model) and Llama-3.3-70B-Instruct llama-4-scout,Meta,B,size,MoE 109B total / 17B active (Meta's own model-information table),meta-llama/Llama-4-Scout-17B-16E-Instruct,no,"main.tex Table S20 (Supplementary S6), v3",2026-09,Table S4 row 24 llama-3.1-405b;llama-4-maverick,Meta,C,size,405B dense; Llama 4 Maverick MoE 400B total / 17B active (Meta's own model-information table),meta-llama/Llama-3.1-405B-Instruct,no,"main.tex Table S20 (Supplementary S6), v3",2026-09,Table S4 row 25; model_factors.csv also carries Llama-3.1-405B-Instruct-FP8 and Llama-4-Maverick-17B-128E-Instruct-FP8 deepseek-v3;deepseek-chat,DeepSeek,C,size,MoE 671B total / 37B active,deepseek-ai/DeepSeek-V3.1,no,"main.tex Table S20 (Supplementary S6), v3",2026-09,Table S4 row 26; as a prefix this also covers deepseek-v3.1 and deepseek-v3.2. deepseek-chat was retired from DeepSeek's API docs by 2026-10-02 but still appears in historical billing exports deepseek-r1,DeepSeek,C,size,Reasoning variant of V3; hidden tokens as for the OpenAI o-series,deepseek-ai/DeepSeek-V3.1,yes,"main.tex Table S20 (Supplementary S6), v3",2026-09,Table S4 row 27 qwen3-235b-a22b,Alibaba,C,size,"MoE 235B total / 22B active, the Class C calibration model",Qwen/Qwen3-235B-A22B-Instruct-2507,no,"main.tex Table S20 (Supplementary S6), v3",2026-09,Table S4 row 28 ministral-*;mixtral-8x7b,Mistral,A,size,"Ministral 3-14B dense (Ministral 3 ships at 3B, 8B and 14B, all <=15B); Mixtral 8x7B MoE 47B total / 13B active; energy approx. dense 8B per TokenPowerBench",,no,"main.tex Table S20 (Supplementary S6), v3",2026-09,Table S4 row 29 mistral-small,Mistral,B,size,Mistral Small 3 is 24B dense,,no,"main.tex Table S20 (Supplementary S6), v3",2026-09,"Table S4 row 30; Mistral Small 4 is MoE 119B total / 6.5B active and so still Class B, by the MoE rather than the dense rule (see the mistral-small-4 row)" mistral-medium;mixtral-8x22b,Mistral,B,tier,Medium: standard tier; 8x22B MoE 141B total / 39B active,,no,"main.tex Table S20 (Supplementary S6), v3",2026-09,"Table S4 row 31; mixed basis (tier for Medium, size for 8x22B) as printed in S4. SUPERSEDED for Medium 3.5, whose 128B dense size Mistral now discloses and which is therefore Class C by the size rule (see the mistral-medium-3 row)" mistral-large,Mistral,C,size,"123B dense (Mistral's own lifecycle assessment: 1.14 g per 400-token response); Mistral Large 3 is MoE 675B total / 41B active, also Class C",,no,"main.tex Table S20 (Supplementary S6), v3",2026-09,Table S4 row 32 sonar,Perplexity,B,size,"Perplexity states Sonar is built on Llama 3.3 70B, a Class B derivative",meta-llama/Llama-3.3-70B-Instruct,no,"main.tex Table S20 (Supplementary S6), v3",2026-09,"Table S4 row 33. FLAGGED: Perplexity's current model and model-card pages state no base model and no size, so this assignment rests on an earlier Perplexity statement that cannot be re-verified" sonar-pro,Perplexity,B,tier,Pro tier (USD 3.00 / 15.00 per M plus a per-request search fee),,no,"main.tex Table S20 (Supplementary S6), v3",2026-09,Table S4 row 34 sonar-reasoning*,Perplexity,C,tier,Reasoning tier (sonar-reasoning-pro USD 2.00 / 8.00 per M plus fees),,yes,"main.tex Table S20 (Supplementary S6), v3",2026-09,Table S4 row 35; the non-Pro sonar-reasoning id is no longer on Perplexity's model list grok-*,xAI,C,tier,Frontier closed model; grok-4.7 is 'the most capable model we've built' (xAI),,,"main.tex Table S20 (Supplementary S6), v3",2026-09,"Table S4 row 36; xAI's current list has no named mini or fast text tier, so every current grok text identifier is Class C. the provider publishes no definition of its tier words, so the ordering is an inference from the prices on the page cited in source" gpt-3.5-turbo;gpt-3.5,OpenAI,A,tier,"Turbo was the cheap tier of its generation, priced an order of magnitude below gpt-4",,no,https://developers.openai.com/api/docs/models ; https://developers.openai.com/api/docs/pricing (fetched 2026-10-02; platform.openai.com/docs/models now 301-redirects here and openai.com/api/pricing returns 403),2026-10,"Added for the five gpt-3.5-turbo-* identifiers in v4_inputs/model_service_life.csv; without it they fell to the Class C default, a 21x overstatement" gpt-4-32k,OpenAI,C,tier,Long-context variant of the gpt-4 frontier tier,,no,https://developers.openai.com/api/docs/models ; https://developers.openai.com/api/docs/pricing (fetched 2026-10-02; platform.openai.com/docs/models now 301-redirects here and openai.com/api/pricing returns 403),2026-10, gpt-5*-mini,OpenAI,A,tier,Mini tier of the GPT-5 generations after 5.0 (gpt-5.4-mini USD 0.75 / 4.50 per M),,yes,https://developers.openai.com/api/docs/models ; https://developers.openai.com/api/docs/pricing (fetched 2026-10-02; platform.openai.com/docs/models now 301-redirects here and openai.com/api/pricing returns 403),2026-10,Needed because gpt-5.4-mini does not begin with the literal gpt-5-mini; see the tie-break rule in DETAILED_TABLES_README.md gpt-5*-nano,OpenAI,A,tier,Nano tier (gpt-5.4-nano USD 0.20 / 1.25 per M),,yes,https://developers.openai.com/api/docs/models ; https://developers.openai.com/api/docs/pricing (fetched 2026-10-02; platform.openai.com/docs/models now 301-redirects here and openai.com/api/pricing returns 403),2026-10, gpt-5*-terra;gpt-6*-terra,OpenAI,A,tier,OpenAI states that -terra 'roughly corresponds to the mini model tier used in earlier GPT-5 families',,yes,https://developers.openai.com/api/docs/models ; https://developers.openai.com/api/docs/pricing (fetched 2026-10-02; platform.openai.com/docs/models now 301-redirects here and openai.com/api/pricing returns 403),2026-10,"FLAGGED: gpt-5.6-terra is USD 2.00 / 12.00 per M, which is a standard-tier price, so a preparer who classes by price rather than by the vendor's own tier word would put it in B. The provider's statement is followed here, per Section 3.5" gpt-5*-luna;gpt-6*-luna,OpenAI,A,tier,"OpenAI states that -luna 'roughly corresponds to the nano model tier used in earlier GPT-5 families'; gpt-6-luna is USD 0.10 / 0.50 per M, the cheapest model on the list",,yes,https://developers.openai.com/api/docs/models ; https://developers.openai.com/api/docs/pricing (fetched 2026-10-02; platform.openai.com/docs/models now 301-redirects here and openai.com/api/pricing returns 403),2026-10,"Covers the -luna-pro variants too: -pro raises reasoning effort, which changes tokens per interaction and not the per-token class" gpt-5*-sol;gpt-6*-sol,OpenAI,C,tier,"OpenAI calls -sol the flagship model of its generation (gpt-5.6-sol USD 4.00 / 20.00, gpt-6-sol USD 2.00 / 10.00 per M)",,yes,https://developers.openai.com/api/docs/models ; https://developers.openai.com/api/docs/pricing (fetched 2026-10-02; platform.openai.com/docs/models now 301-redirects here and openai.com/api/pricing returns 403),2026-10, gpt-6*-astra,OpenAI,C,tier,"OpenAI calls gpt-6-astra 'our most capable model for the most demanding work' and prices it at USD 10.00 / 50.00 per M, above every other listed model except gpt-5-pro",,yes,https://developers.openai.com/api/docs/models ; https://developers.openai.com/api/docs/pricing (fetched 2026-10-02; platform.openai.com/docs/models now 301-redirects here and openai.com/api/pricing returns 403),2026-10,"Covers gpt-6-astra-pro, which appears in the OpenRouter top 150 but not on OpenAI's own pages" gpt-5*-codex-mini;gpt-6*-codex-mini,OpenAI,A,tier,Mini tier of the Codex line,,yes,https://developers.openai.com/api/docs/models ; https://developers.openai.com/api/docs/pricing (fetched 2026-10-02; platform.openai.com/docs/models now 301-redirects here and openai.com/api/pricing returns 403),2026-10, codex-mini,OpenAI,A,tier,Mini tier of the Codex line,,yes,https://developers.openai.com/api/docs/models ; https://developers.openai.com/api/docs/pricing (fetched 2026-10-02; platform.openai.com/docs/models now 301-redirects here and openai.com/api/pricing returns 403),2026-10,Added for codex-mini and codex-mini-latest in v4_inputs/model_service_life.csv: prefix matching could not reach the mini tier because codex- precedes it. Neither id is on OpenAI's current pages gpt-5*-codex;gpt-6*-codex,OpenAI,C,tier,"Codex variants are priced at the flagship tier of their generation (gpt-5.3-codex USD 1.75 / 14.00 per M, the same as gpt-5.2)",,yes,https://developers.openai.com/api/docs/models ; https://developers.openai.com/api/docs/pricing (fetched 2026-10-02; platform.openai.com/docs/models now 301-redirects here and openai.com/api/pricing returns 403),2026-10, gpt-5*-codex-max;gpt-6*-codex-max,OpenAI,C,tier,Priced above the flagship of its generation,,yes,https://developers.openai.com/api/docs/models ; https://developers.openai.com/api/docs/pricing (fetched 2026-10-02; platform.openai.com/docs/models now 301-redirects here and openai.com/api/pricing returns 403),2026-10,In v4_inputs/model_service_life.csv as gpt-5.1-codex-max; not on OpenAI's current pages gpt-5*-pro;gpt-6*-pro,OpenAI,C,tier,"For OpenAI, -pro is the top tier of a generation: gpt-5-pro is USD 15.00 / 120.00 per M, twelve times gpt-5",,yes,https://developers.openai.com/api/docs/models ; https://developers.openai.com/api/docs/pricing (fetched 2026-10-02; platform.openai.com/docs/models now 301-redirects here and openai.com/api/pricing returns 403),2026-10,"Note that -pro means the MID tier for Google Gemini and Perplexity and the TOP tier for OpenAI, so the patterns are provider-specific. Where -pro follows a named small tier (gpt-6-luna-pro) the small-tier pattern wins, because -pro raises reasoning effort and not per-token energy" gpt-5*-chat;gpt-6*-chat;gpt-chat-latest;chat-latest,OpenAI,C,tier,"Non-reasoning chat variant of the flagship, at or above the flagship price (chat-latest USD 5.00 / 30.00 per M)",,no,https://developers.openai.com/api/docs/models ; https://developers.openai.com/api/docs/pricing (fetched 2026-10-02; platform.openai.com/docs/models now 301-redirects here and openai.com/api/pricing returns 403),2026-10,"Added for gpt-chat-latest, gpt-5-chat-latest and the 5.1 to 5.3 -chat-latest ids in v4_inputs/model_service_life.csv, which matched no pattern or only the generic gpt-5 row. On OpenAI's current pricing page the id is the bare chat-latest" gpt-6,OpenAI,C,tier,Frontier closed model of the GPT-6 generation,,yes,https://developers.openai.com/api/docs/models ; https://developers.openai.com/api/docs/pricing (fetched 2026-10-02; platform.openai.com/docs/models now 301-redirects here and openai.com/api/pricing returns 403),2026-10,"Covers gpt-6 and, as a prefix, gpt-6.1; the tier-suffixed GPT-6 ids are handled by the -astra, -sol and -luna rows" gpt-6*-mini,OpenAI,A,tier,Mini tier of the GPT-6 generation,,yes,https://developers.openai.com/api/docs/models ; https://developers.openai.com/api/docs/pricing (fetched 2026-10-02; platform.openai.com/docs/models now 301-redirects here and openai.com/api/pricing returns 403),2026-10, gpt-6*-nano,OpenAI,A,tier,Nano tier of the GPT-6 generation,,yes,https://developers.openai.com/api/docs/models ; https://developers.openai.com/api/docs/pricing (fetched 2026-10-02; platform.openai.com/docs/models now 301-redirects here and openai.com/api/pricing returns 403),2026-10, gpt-oss-120b,OpenAI,B,size,Open-weight MoE 117B total / 5.1B active (50B to below 200B),,,https://developers.openai.com/api/docs/models ; https://developers.openai.com/api/docs/pricing (fetched 2026-10-02; platform.openai.com/docs/models now 301-redirects here and openai.com/api/pricing returns 403),2026-10,"Added from the OpenRouter top 150 (rank 54); the Class C default was a 3.1x overstatement. OpenAI's current pricing page has no open-weight section, so the size comes from the model card" gpt-oss-20b,OpenAI,A,size,Open-weight MoE 21B total / 3.6B active (below 50B),,,https://developers.openai.com/api/docs/models ; https://developers.openai.com/api/docs/pricing (fetched 2026-10-02; platform.openai.com/docs/models now 301-redirects here and openai.com/api/pricing returns 403),2026-10,Added from the OpenRouter top 150 (rank 113); the Class C default was a 21x overstatement o3-deep-research;o4-mini-deep-research,OpenAI,C,tier,Deep-research variants run long agentic reasoning loops and are the frontier product of their line,,yes,https://developers.openai.com/api/docs/models ; https://developers.openai.com/api/docs/pricing (fetched 2026-10-02; platform.openai.com/docs/models now 301-redirects here and openai.com/api/pricing returns 403),2026-10,"FLAGGED: o4-mini-deep-research is placed in C rather than with o4-mini (A). If the per-token class is read strictly off the base model it belongs in A, which would be a 21x difference; C is the conservative reading" claude-fable-*;claude-*-fable;claude-mythos-*;claude-*-mythos,Anthropic,C,tier,"A new top tier priced ABOVE Opus: Fable 5.1 is USD 10.00 / 50.00 per M against Opus 5.5 at 4.00 / 20.00, and Anthropic calls claude-fable-5-1 'Anthropic's most capable model open to all customers'",,no,https://platform.claude.com/docs/en/about-claude/models/overview ; https://platform.claude.com/docs/en/about-claude/pricing ; https://platform.claude.com/docs/en/about-claude/models/choosing-a-model (fetched 2026-10-02),2026-10,"Added for claude-fable-5, claude-fable-5-1, claude-mythos-5 and claude-mythos-5-1 in v4_inputs/model_service_life.csv, the whole current Anthropic top generation, which Table S4 does not know. Anthropic states that Mythos 5.1 'shares Claude Fable 5.1's specifications and pricing' and is the same model offered by invitation only, so the two collapse into one tier. Both naming orders are covered because OpenRouter writes claude-5-fable-20260609" claude-*-haiku,Anthropic,A,tier,Haiku = small tier,,no,https://platform.claude.com/docs/en/about-claude/models/overview ; https://platform.claude.com/docs/en/about-claude/pricing ; https://platform.claude.com/docs/en/about-claude/models/choosing-a-model (fetched 2026-10-02),2026-10,Second naming order: OpenRouter and some provider pages write claude-4.5-haiku-20251001 rather than claude-haiku-4-5 claude-*-sonnet,Anthropic,B,tier,Sonnet = mid tier,,no,https://platform.claude.com/docs/en/about-claude/models/overview ; https://platform.claude.com/docs/en/about-claude/pricing ; https://platform.claude.com/docs/en/about-claude/models/choosing-a-model (fetched 2026-10-02),2026-10,"Second naming order, as above" claude-*-opus,Anthropic,C,tier,Opus = frontier tier,,no,https://platform.claude.com/docs/en/about-claude/models/overview ; https://platform.claude.com/docs/en/about-claude/pricing ; https://platform.claude.com/docs/en/about-claude/models/choosing-a-model (fetched 2026-10-02),2026-10,"Second naming order, as above" claude-instant,Anthropic,A,tier,Instant was the small tier of the Claude 1 and 2 generations,,no,https://platform.claude.com/docs/en/about-claude/models/overview ; https://platform.claude.com/docs/en/about-claude/pricing ; https://platform.claude.com/docs/en/about-claude/models/choosing-a-model (fetched 2026-10-02),2026-10, claude-2;claude-v2;claude-1;claude-v1,Anthropic,C,tier,"Flagship of its generation, no disclosed size and no published tier split other than Instant",,no,https://platform.claude.com/docs/en/about-claude/models/overview ; https://platform.claude.com/docs/en/about-claude/pricing ; https://platform.claude.com/docs/en/about-claude/models/choosing-a-model (fetched 2026-10-02),2026-10, gemini-1.0-pro;gemini-1.5-pro;gemini-pro,Google,B,tier,Pro = mid tier,,,https://ai.google.dev/gemini-api/docs/models ; https://ai.google.dev/gemini-api/docs/pricing (fetched 2026-10-02),2026-10,"gemini-pro (the unversioned alias) does not match gemini-*-pro, which requires a version segment between the two literals" gemini-1.5-flash;gemini-flash,Google,A,tier,Flash = efficiency tier,,,https://ai.google.dev/gemini-api/docs/models ; https://ai.google.dev/gemini-api/docs/pricing (fetched 2026-10-02),2026-10,gemini-flash (the unversioned alias) does not match gemini-*-flash gemini-*-nano;gemini-nano,Google,A,tier,Nano = on-device efficiency tier,,,https://ai.google.dev/gemini-api/docs/models ; https://ai.google.dev/gemini-api/docs/pricing (fetched 2026-10-02),2026-10, gemini-omni,Google,B,tier,Omni is a multimodal line with no published tier above or below Pro,,,https://ai.google.dev/gemini-api/docs/models ; https://ai.google.dev/gemini-api/docs/pricing (fetched 2026-10-02),2026-10,"gemini-omni-1.1-flash also matches gemini-*-flash (A), which wins on the longer literal count; this row catches any unsuffixed omni id. tier could not be determined from the provider's own model list or pricing page, so Class B here, because the only published Omni model is a Flash is assigned as the conservative default (Section 3.5)" gemma-*-e2b;gemma-*-e4b;gemma-*-270m;gemma-*-1b;gemma-*-3b;gemma-*-7b;gemma-*-8b,Google,A,size,"Open 0.27-8B dense; Gemma 4 E2B is 2.3B effective (5.1B with embeddings) and E4B 4.5B effective (8B with embeddings), where Google's 'E' means effective parameters",,no,https://huggingface.co/google/gemma-4-26B-A4B-it ; https://huggingface.co/google (fetched 2026-10-02),2026-10,Completes S4's Gemma size ladder with the sizes Google has actually released gemma-*-26b-a4b,Google,A,size,Open MoE 25.2B total / 3.8B active (below 50B); Google's 'A' means active parameters,,,https://huggingface.co/google/gemma-4-26B-A4B-it ; https://huggingface.co/google (fetched 2026-10-02),2026-10,"Added from the OpenRouter top 150 (gemma-4-26b-a4b-it, rank 75); S4's Gemma rows are dense-only" gemma-*-31b,Google,B,size,Open 30.7B dense (above 15B to 70B),,,https://huggingface.co/google/gemma-4-26B-A4B-it ; https://huggingface.co/google (fetched 2026-10-02),2026-10,"Added from the OpenRouter top 150 (gemma-4-31b-it, rank 46)" gemma-,Google,B,size,Gemma release whose size is not encoded in the identifier,,,https://huggingface.co/google/gemma-4-26B-A4B-it ; https://huggingface.co/google (fetched 2026-10-02),2026-10,"Catch-all below every sized gemma pattern: every Gemma release to date is 270M to 31B, so B rather than the Class C default. An unsized Gemma identifier is flagged in the inventory" grok-*-mini;grok-mini,xAI,A,tier,Mini = efficiency tier,,,https://docs.x.ai/docs/models (fetched 2026-10-02; prices are on the same page),2026-10,"Covers the historical grok-3-mini. FLAGGED: xAI's current model list (fetched 2026-10-02) contains no mini and no fast text tier, so this row is for historical billing exports only" mistral-tiny,Mistral,A,tier,Tiny = smallest tier,,no,https://docs.mistral.ai/getting-started/models/models_overview/ ; https://mistral.ai/pricing/api (fetched 2026-10-02),2026-10, mistral-nemo,Mistral,A,size,12B dense (<=15B),,no,https://huggingface.co/mistralai (model cards; fetched 2026-10-02),2026-10,"The audit found this falling to the Class C default, a 21x overstatement. Mistral's current model list no longer carries it" mistral-small-4,Mistral,B,size,"MoE 119B total / 6.5B active, 128 experts with 4 active (50B to below 200B)",,no,https://docs.mistral.ai/getting-started/models/models_overview/ ; https://mistral.ai/pricing/api (fetched 2026-10-02),2026-10,Same class as S4's mistral-small row but on the MoE rather than the dense rule; recorded separately because the size basis changed mistral-medium-3,Mistral,C,size,"128B dense, above the 70B dense cut",,no,https://docs.mistral.ai/getting-started/models/models_overview/ ; https://mistral.ai/pricing/api (fetched 2026-10-02),2026-10,"CORRECTS Table S4 for Mistral Medium 3.5. S4 placed mistral-medium in B on the vendor tier; Mistral now discloses 128B dense for Medium 3.5, and Section 3.5 gives the disclosed size precedence over the tier, so Medium 3.5 is Class C. The longer literal prefix makes this row win over S4's mistral-medium for every medium-3 identifier" mistral-large-3,Mistral,C,size,MoE 675B total / 41B active (200B and above),,no,https://docs.mistral.ai/getting-started/models/models_overview/ ; https://mistral.ai/pricing/api (fetched 2026-10-02),2026-10, pixtral-12b,Mistral,A,size,12B dense multimodal (<=15B),,no,https://huggingface.co/mistralai (model cards; fetched 2026-10-02),2026-10,Not on Mistral's current model list; retained for historical billing exports pixtral-large,Mistral,C,size,124B dense multimodal,,no,https://huggingface.co/mistralai (model cards; fetched 2026-10-02),2026-10,Not on Mistral's current model list codestral;devstral,Mistral,B,size,22-24B dense (above 15B to 70B),,no,https://docs.mistral.ai/getting-started/models/models_overview/ ; https://mistral.ai/pricing/api (fetched 2026-10-02),2026-10,"FLAGGED: Mistral's current page lists Codestral with no disclosed size and does not list Devstral at all (the Medium 3.5 card says Medium supersedes Devstral 2). The 22-24B figures are from the earlier releases, so this row is an assumption carried forward" magistral,Mistral,B,size,Magistral Small is 24B dense; Magistral Medium is the standard reasoning tier,,yes,https://huggingface.co/mistralai (model cards; fetched 2026-10-02),2026-10,Not on Mistral's current model list; retained for historical billing exports mixtral-8x,Mistral,B,size,Mixtral MoE release whose size is not one of the two in S4,,no,https://docs.mistral.ai/getting-started/models/models_overview/ ; https://mistral.ai/pricing/api (fetched 2026-10-02),2026-10,"Catch-all between mixtral-8x7b (A) and mixtral-8x22b (B); 8x7B is the only Mixtral below the 50B MoE cut, so B is the conservative landing for any other 8xN" leanstral;shieldstral;classifier-,Mistral,A,size,Classifier and guard models of 3-8B dense,,no,https://docs.mistral.ai/getting-started/models/models_overview/ ; https://mistral.ai/pricing/api (fetched 2026-10-02),2026-10, mistral-7b;open-mistral-7b,Mistral,A,size,"Mistral 7B dense, 7.3B parameters (<=15B); covers Mistral-7B-Instruct v0.1 to v0.3 and the hosted ids open-mistral-7b and mistral.mistral-7b-instruct (prefix match)",,no,"Jiang et al., Mistral 7B, arXiv:2310.06825 (7.3B dense); https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.3 (model card)",2026-10,"Added 2026-10-03 after the supplement review: an identifier such as mistral-7b-instruct matched no pattern and fell to the Class C default, a 21x overstatement for a 7B dense model. Retired from Mistral's own API list; retained for historical billing exports and third-party hosts" llama-4-behemoth,Meta,C,size,MoE announced at about 2T total / 288B active (200B and above),,no,"https://huggingface.co/meta-llama (Llama 4 model card, Meta's own model-information table; fetched 2026-10-02)",2026-10,"FLAGGED: Meta's Llama 4 model card does not state Behemoth's parameter counts and Meta's current developer site no longer lists Llama at all, so the figures above are from the announcement rather than a provider table. C either way under the MoE rule" llama3-8b;llama3-1-8b;llama-3-8b,Meta,A,size,8B dense,meta-llama/Llama-3.1-8B-Instruct,no,https://docs.aws.amazon.com/bedrock/latest/userguide/models-supported.html,2026-10,"AWS Bedrock writes meta.llama3-1-8b-instruct-v1:0 with no dot in the version, which S4's llama-3.1-8b cannot match" llama3-70b;llama3-1-70b;llama3-3-70b;llama-3-70b,Meta,B,size,70B dense,meta-llama/Llama-3.3-70B-Instruct,no,https://docs.aws.amazon.com/bedrock/latest/userguide/models-supported.html,2026-10,"Bedrock spelling, as above" llama3-2-1b;llama3-2-3b;llama3-2-11b,Meta,A,size,1-11B dense,,no,https://docs.aws.amazon.com/bedrock/latest/userguide/models-supported.html,2026-10,Bedrock spelling llama3-2-90b,Meta,C,size,90B dense (vision),,no,https://docs.aws.amazon.com/bedrock/latest/userguide/models-supported.html,2026-10,Bedrock spelling llama3-1-405b;llama3-405b,Meta,C,size,405B dense,meta-llama/Llama-3.1-405B-Instruct,no,https://docs.aws.amazon.com/bedrock/latest/userguide/models-supported.html,2026-10,Bedrock spelling llama4-scout,Meta,B,size,MoE 109B total / 17B active,meta-llama/Llama-4-Scout-17B-16E-Instruct,no,https://docs.aws.amazon.com/bedrock/latest/userguide/models-supported.html,2026-10,Bedrock writes meta.llama4-scout-17b-16e-instruct-v1:0 llama4-maverick,Meta,C,size,MoE 400B total / 17B active,meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8,no,https://docs.aws.amazon.com/bedrock/latest/userguide/models-supported.html,2026-10,Bedrock spelling llama-guard;llama-prompt-guard;prompt-guard,Meta,A,size,0.07-12B dense guard models,,no,"https://huggingface.co/meta-llama (Llama 4 model card, Meta's own model-information table; fetched 2026-10-02)",2026-10, muse-spark;muse-glimmer;muse-code,Meta,C,tier,"tier could not be determined from the provider's own model list or pricing page, so Class C is assigned as the conservative default (Section 3.5)",,,https://dev.meta.ai/models/ (fetched 2026-10-02; llama.com and developer.meta.com now redirect here),2026-10,"From the OpenRouter top 150 (meta/muse-spark-1.3 at rank 35, 1.3-contributor at 63 and 1.2 at 91). Meta's current developer site lists the Muse family with no parameter counts and no prices at all, so neither branch of Section 3.5 can be applied. tier could not be determined from the provider's own model list or pricing page, so Class C is assigned as the conservative default (Section 3.5)" deepseek-reasoner,DeepSeek,C,tier,The reasoning endpoint of the DeepSeek flagship,deepseek-ai/DeepSeek-V3.1,yes,https://api-docs.deepseek.com/quick_start/pricing ; https://api-docs.deepseek.com/quick_start/models ; https://huggingface.co/deepseek-ai (model cards; fetched 2026-10-02),2026-10,Retired from DeepSeek's API docs by 2026-10-02; retained for historical billing exports deepseek-v4;deepseek-v4*-pro,DeepSeek,C,size,DeepSeek-V4-Pro is MoE 1.6T total / 49B active and the -0813 release 1.7T total (200B and above),,,https://api-docs.deepseek.com/quick_start/pricing ; https://api-docs.deepseek.com/quick_start/models ; https://huggingface.co/deepseek-ai (model cards; fetched 2026-10-02),2026-10,As a prefix deepseek-v4 also covers deepseek-v4.1 deepseek-v4*-flash;deepseek-flash,DeepSeek,C,size,"DeepSeek-V4.1-Flash is MoE 552B total / 8B active in prefill and 16B in decode, and DeepSeek-V4-Flash 284B total / 13B active; both are at or above the 200B MoE cut",,,https://api-docs.deepseek.com/quick_start/pricing ; https://api-docs.deepseek.com/quick_start/models ; https://huggingface.co/deepseek-ai (model cards; fetched 2026-10-02),2026-10,"IMPORTANT, AND IT REVERSES THE AUDIT'S EXPECTATION. The audit recorded deepseek-v4-flash falling to Class C as an error, because 'flash' is a Class A tier word. But DeepSeek discloses the parameter counts, and Section 3.5 assigns class by total parameters WHERE THE SIZE IS DISCLOSED and only falls back to the vendor tier where it is not. At 284-552B total these are Class C on the MoE rule. The tier word and the disclosed size disagree, and the rule gives the size precedence" deepseek-v2;deepseek-coder,DeepSeek,C,size,DeepSeek-V2 is MoE 236B total / 21B active,,no,https://api-docs.deepseek.com/quick_start/pricing ; https://api-docs.deepseek.com/quick_start/models ; https://huggingface.co/deepseek-ai (model cards; fetched 2026-10-02),2026-10, deepseek-r1-distill-llama-8b;deepseek-r1-distill-qwen-1.5b;deepseek-r1-distill-qwen-7b,DeepSeek,A,size,Distilled 1.5-8B dense,,yes,https://api-docs.deepseek.com/quick_start/pricing ; https://api-docs.deepseek.com/quick_start/models ; https://huggingface.co/deepseek-ai (model cards; fetched 2026-10-02),2026-10, deepseek-r1-distill-qwen-14b,DeepSeek,A,size,Distilled 14B dense (<=15B),,yes,https://api-docs.deepseek.com/quick_start/pricing ; https://api-docs.deepseek.com/quick_start/models ; https://huggingface.co/deepseek-ai (model cards; fetched 2026-10-02),2026-10, deepseek-r1-distill-llama-70b;deepseek-r1-distill-qwen-32b,DeepSeek,B,size,Distilled 32-70B dense,,yes,https://api-docs.deepseek.com/quick_start/pricing ; https://api-docs.deepseek.com/quick_start/models ; https://huggingface.co/deepseek-ai (model cards; fetched 2026-10-02),2026-10, qwen*-max,Alibaba,C,tier,Max is the most expensive Qwen API tier (qwen3.8-max USD 2.00 / 6.00 per M),,,https://www.alibabacloud.com/help/en/model-studio/models ; https://www.alibabacloud.com/help/en/model-studio/model-pricing (fetched 2026-10-02),2026-10,"Covers qwen3.7-max, qwen3.8-max and qwen3.8-max-prime. FLAGGED: the provider publishes no definition of its tier words, so the ordering is an inference from the prices on the page cited in source; Alibaba's model-studio pages give no description for any of max, plus, flash or turbo" qwen*-plus,Alibaba,B,tier,Plus sits between max and flash (qwen3.7-plus USD 0.40 / 1.60 per M),,,https://www.alibabacloud.com/help/en/model-studio/models ; https://www.alibabacloud.com/help/en/model-studio/model-pricing (fetched 2026-10-02),2026-10,"the provider publishes no definition of its tier words, so the ordering is an inference from the prices on the page cited in source" qwen*-flash;qwen*-turbo,Alibaba,A,tier,"Flash and turbo are the cheapest Qwen API tiers (qwen3.8-flash USD 0.15 / 0.47, qwen-turbo USD 0.05 / 0.20 per M)",,,https://www.alibabacloud.com/help/en/model-studio/models ; https://www.alibabacloud.com/help/en/model-studio/model-pricing (fetched 2026-10-02),2026-10,"the provider publishes no definition of its tier words, so the ordering is an inference from the prices on the page cited in source" qwen3-8b,Alibaba,A,size,8B dense (<=15B),Qwen/Qwen3-8B,,ML.ENERGY Leaderboard v3 snapshot (model_factors.csv in this deposit),2026-10,Measured in ML.ENERGY v3 qwen3-14b,Alibaba,A,size,14B dense (<=15B),Qwen/Qwen3-14B,,ML.ENERGY Leaderboard v3 snapshot (model_factors.csv in this deposit),2026-10,Measured in ML.ENERGY v3 qwen3-32b,Alibaba,B,size,32B dense (above 15B to 70B),Qwen/Qwen3-32B,,ML.ENERGY Leaderboard v3 snapshot (model_factors.csv in this deposit),2026-10,Measured in ML.ENERGY v3 qwen3-30b-a3b,Alibaba,A,size,MoE 30B total / 3B active (below 50B),Qwen/Qwen3-30B-A3B-Instruct-2507,,ML.ENERGY Leaderboard v3 snapshot (model_factors.csv in this deposit),2026-10,Measured in ML.ENERGY v3; the 50B MoE cut was set to keep this model with the Class A models it measures alongside qwen3-vl-235b-a22b;qwen3-omni-235b-a22b,Alibaba,C,size,MoE 235B total / 22B active (200B and above),Qwen/Qwen3-235B-A22B-Instruct-2507,,v4_inputs/raw/openrouter_rankings_month.json (top 150 by tokens),2026-10,From the OpenRouter top 150 (rank 148); does not match S4's qwen3-235b-a22b because the -vl- segment intervenes qwen*-2.4t-a95b,Alibaba,C,size,MoE 2.4T total / 95B active (200B and above),,,https://huggingface.co/Qwen (model cards; fetched 2026-10-02),2026-10,"From the OpenRouter top 150 (qwen3.8-2.4t-a95b, rank 87); size confirmed on the Qwen model card" qwen*-397b-a17b,Alibaba,C,size,MoE 397B total / 17B active (200B and above),,,v4_inputs/raw/openrouter_rankings_month.json (top 150 by tokens),2026-10,"From the OpenRouter top 150 (qwen3.5-397b-a17b, rank 76); the size is read off the identifier, Alibaba's card not being reachable for this release" qwen*-122b-a10b,Alibaba,B,size,MoE 122B total / 10B active (50B to below 200B),,,v4_inputs/raw/openrouter_rankings_month.json (top 150 by tokens),2026-10,"From the OpenRouter top 150 (qwen3.5-122b-a10b, rank 146); size read off the identifier" qwen*-35b-a3b,Alibaba,A,size,MoE 35B total / 3B active (below 50B),,,v4_inputs/raw/openrouter_rankings_month.json (top 150 by tokens),2026-10,From the OpenRouter top 150 (qwen3.5-35b-a3b and qwen3.6-35b-a3b); size read off the identifier qwen*-27b,Alibaba,B,size,27-28B dense (above 15B to 70B),,,https://huggingface.co/Qwen (model cards; fetched 2026-10-02),2026-10,"From the OpenRouter top 150 (qwen3.5-27b, qwen3.6-27b, qwen3.8-27b); Qwen3.8-27B is confirmed on the Qwen model card" qwen*-72b;qwen*-110b;qwen*-180b,Alibaba,C,size,"72-180B dense, above the 70B dense cut",,,https://www.alibabacloud.com/help/en/model-studio/models ; https://www.alibabacloud.com/help/en/model-studio/model-pricing (fetched 2026-10-02),2026-10, qwen*-0.5b;qwen*-1.5b;qwen*-1.8b;qwen*-3b;qwen*-4b;qwen*-7b;qwen*-8b;qwen*-14b,Alibaba,A,size,0.5-14B dense (<=15B),,,https://www.alibabacloud.com/help/en/model-studio/models ; https://www.alibabacloud.com/help/en/model-studio/model-pricing (fetched 2026-10-02),2026-10, qwen*-32b,Alibaba,B,size,32B dense (above 15B to 70B),,,https://www.alibabacloud.com/help/en/model-studio/models ; https://www.alibabacloud.com/help/en/model-studio/model-pricing (fetched 2026-10-02),2026-10, qwen,Alibaba,C,tier,Qwen release whose size and tier are not encoded in the identifier,,,https://www.alibabacloud.com/help/en/model-studio/models ; https://www.alibabacloud.com/help/en/model-studio/model-pricing (fetched 2026-10-02),2026-10,"Catch-all below every sized and tiered qwen pattern; Class C per the unmapped default, and flagged" command-r7b,Cohere,A,size,7B dense (<=15B),,no,https://docs.cohere.com/docs/models ; https://cohere.com/pricing (fetched 2026-10-02),2026-10,FLAGGED: the 7B figure is read off the identifier. Cohere's model page states no parameter count for any Command model command-light;command-nightly-light,Cohere,A,tier,Light = the cheap tier of the legacy Command line (USD 0.30 / 0.60 per M),,no,https://docs.cohere.com/docs/models ; https://cohere.com/pricing (fetched 2026-10-02),2026-10, command-r,Cohere,B,size,35B dense (above 15B to 70B),,no,https://docs.cohere.com/docs/models ; https://cohere.com/pricing (fetched 2026-10-02),2026-10,"FLAGGED: Cohere's current page states no size; 35B is from the original Command R release. Beaten by command-r-plus and command-r7b, which have longer literal prefixes" command-r-plus,Cohere,C,size,"104B dense, above 70B",,no,https://docs.cohere.com/docs/models ; https://cohere.com/pricing (fetched 2026-10-02),2026-10,"FLAGGED: size from the original Command R+ release, not from Cohere's current page" command-a,Cohere,C,tier,Cohere calls command-a-03-2025 its 'most performant model to date' and command-a-plus its first mixture-of-experts model; no size is published and the current generation has no published price,,no,https://docs.cohere.com/docs/models ; https://cohere.com/pricing (fetched 2026-10-02),2026-10,"Covers command-a-plus, command-a-reasoning, command-a-vision and command-a-translate. tier could not be determined from the provider's own model list or pricing page, so Class C is assigned as the conservative default (Section 3.5)" command,Cohere,C,tier,Legacy Cohere Command (USD 1.00 / 2.00 per M); no published size,,no,https://docs.cohere.com/docs/models ; https://cohere.com/pricing (fetched 2026-10-02),2026-10,"Catch-all below the sized and tiered command patterns. tier could not be determined from the provider's own model list or pricing page, so Class C is assigned as the conservative default (Section 3.5)" north-mini,Cohere,A,size,MoE 30B total / 3B active (below 50B),,,https://docs.cohere.com/docs/models ; https://cohere.com/pricing (fetched 2026-10-02),2026-10,"Cohere does disclose sizes for its North models, unlike Command" north-small,Cohere,C,size,MoE 218B total / 25B active (200B and above),,,https://docs.cohere.com/docs/models ; https://cohere.com/pricing (fetched 2026-10-02),2026-10,"FLAGGED: Cohere names this model 'small', but the disclosed 218B total puts it in Class C on the MoE rule, and Section 3.5 gives the disclosed size precedence over the vendor's word" c4ai-aya;aya-expanse-8b;aya-23-8b,Cohere,A,size,8B dense,,no,https://docs.cohere.com/docs/models ; https://cohere.com/pricing (fetched 2026-10-02),2026-10, aya-expanse-32b;aya-23-35b,Cohere,B,size,32-35B dense,,no,https://docs.cohere.com/docs/models ; https://cohere.com/pricing (fetched 2026-10-02),2026-10, nova-micro;nova-*-micro;nova-lite;nova-*-lite,Amazon,A,tier,"Micro ('lowest latency, very low cost', AWS) and Lite are the two small tiers of the Nova family",,no,https://docs.aws.amazon.com/nova/latest/userguide/what-is-nova.html ; https://docs.aws.amazon.com/nova/latest/nova2-userguide/ (fetched 2026-10-02),2026-10,"Bedrock ids are amazon.nova-micro-v1:0, amazon.nova-lite-v1:0 and amazon.nova-2-lite-v1:0, optionally with a us./eu./jp./global. region partition in front; the normalisation strips both. FLAGGED: AWS's Bedrock pricing page is JavaScript-rendered and would not yield per-token prices to a plain fetch, so the tier order here rests on AWS's capability wording, not on published prices" nova-pro;nova-*-pro,Amazon,B,tier,Pro = standard tier of the Nova family,,no,https://docs.aws.amazon.com/nova/latest/userguide/what-is-nova.html ; https://docs.aws.amazon.com/nova/latest/nova2-userguide/ (fetched 2026-10-02),2026-10,"A Nova 2 Pro preview id is in circulation but is not on AWS's own Nova 2 page, which lists Nova 2 Lite, Nova 2 Sonic and Nova Multimodal Embeddings only" nova-premier;nova-*-premier,Amazon,C,tier,AWS calls Nova Premier its 'most capable' Nova model,,no,https://docs.aws.amazon.com/nova/latest/userguide/what-is-nova.html ; https://docs.aws.amazon.com/nova/latest/nova2-userguide/ (fetched 2026-10-02),2026-10,Nova Premier's AWS page marks it Legacy with a model end-of-life date that has passed; retained for historical billing exports titan-text-lite,Amazon,A,tier,Lite = small tier of the Titan Text family (about USD 0.30 / 0.40 per M),,no,https://docs.aws.amazon.com/bedrock/latest/userguide/model-cards.html (fetched 2026-10-02),2026-10,"FLAGGED: Bedrock's current model list carries Titan embedding and image models only, so the Titan text models appear to be withdrawn; these rows are for historical billing exports" titan-text-express,Amazon,B,tier,Express = standard tier of the Titan Text family,,no,https://docs.aws.amazon.com/bedrock/latest/userguide/model-cards.html (fetched 2026-10-02),2026-10, titan-text-premier,Amazon,C,tier,Premier = frontier tier of the Titan Text family,,no,https://docs.aws.amazon.com/bedrock/latest/userguide/model-cards.html (fetched 2026-10-02),2026-10, phi-3-mini;phi-3-small;phi-3-medium;phi-3.5-mini;phi-4-mini;phi-4-multimodal,Microsoft,A,size,"3.8B (mini), 7B (small) and 14B (medium) dense, all <=15B, from the model cards' own text",,no,https://azure.microsoft.com/en-us/products/phi ; https://huggingface.co/microsoft (model cards; fetched 2026-10-02),2026-10,"Every Phi release is at or below 15B dense, so the whole family is Class A on the size rule, not on a tier" phi-3.5-moe;phi-3-moe,Microsoft,A,size,"MoE 16x3.8B with 6.6B active, 42B total (below 50B)",,no,https://azure.microsoft.com/en-us/products/phi ; https://huggingface.co/microsoft (model cards; fetched 2026-10-02),2026-10, phi-4,Microsoft,A,size,14B dense (Microsoft's own figure on the Azure Phi page; the Hugging Face Hub displays 15B),,no,https://azure.microsoft.com/en-us/products/phi ; https://huggingface.co/microsoft (model cards; fetched 2026-10-02),2026-10,"As a prefix this covers phi-4-reasoning and phi-4-reasoning-plus, which are the same 14B weights" phi-,Microsoft,A,size,Phi release whose size is not encoded in the identifier,,,https://azure.microsoft.com/en-us/products/phi ; https://huggingface.co/microsoft (model cards; fetched 2026-10-02),2026-10,"Catch-all below the sized phi patterns: every Phi release read here is <=15B dense or <50B MoE, so A rather than the Class C default. An unsized Phi identifier is flagged in the inventory" glm-5*-flash,Z.ai (Zhipu),C,size,GLM-5.3-Flash is MoE 320B total / 18B active (200B and above),,,https://docs.z.ai/guides/overview/pricing ; https://huggingface.co/zai-org/GLM-5.3-Flash (fetched 2026-10-02),2026-10,"SAME REVERSAL AS DEEPSEEK FLASH. Z.ai's own model card discloses 320B total for GLM-5.3-Flash, and Section 3.5 gives the disclosed size precedence over the 'flash' tier word, so this is Class C and not Class A. From the OpenRouter top 150 (glm-5.3-flash, rank 8)" glm-*-flashx,Z.ai (Zhipu),A,tier,FlashX is priced between Flash and the base model (GLM-5.3-FlashX USD 0.37 / 1.25 against the base at 1.40 / 4.40 per M),,,https://docs.z.ai/guides/overview/pricing ; https://huggingface.co/zai-org/GLM-5.3-Flash (fetched 2026-10-02),2026-10,"FLAGGED: Z.ai defines none of its tier words and publishes no size for the FlashX variants, so this rests on price alone. On GLM-4.5 the X and AirX variants are priced ABOVE the base model, so 'X' is not a uniform discount marker. From the OpenRouter top 150 (glm-5.3-flashx, rank 116)" glm-4*-flash;glm-4*-air,Z.ai (Zhipu),A,tier,"GLM-4.7-Flash is complimentary and GLM-4.7-FlashX USD 0.07 / 0.40 per M, the cheapest tiers on the list",,,https://docs.z.ai/guides/overview/pricing ; https://huggingface.co/zai-org/GLM-5.3-Flash (fetched 2026-10-02),2026-10,FLAGGED: GLM-4.5-Air is MoE 106B total / 12B active and so Class B on the size rule. It is grouped with the flash tiers here because the three share one cheap pricing band; split -air out if the distinction matters glm-;zai-glm,Z.ai (Zhipu),C,size,"GLM-4.5 and GLM-4.6 are MoE 355B total / 32B active; GLM-5, 5.2 and 5.3 are about 753B total with no published active count (all 200B and above)",,,https://docs.z.ai/guides/overview/pricing ; https://huggingface.co/zai-org/GLM-5.3-Flash (fetched 2026-10-02),2026-10,"From the OpenRouter top 150 (glm-5.3 rank 5, glm-5.2 rank 10, glm-5.1 rank 65, glm-5 rank 97, glm-4.7 rank 98, glm-5.3-prime rank 130). The -prime suffix is not on Z.ai's pricing page and lands here. The zai-glm alternative covers Mistral's resale of GLM on its own platform" kimi-k2;kimi-k3;kimi-k1;moonshot-v1,Moonshot AI,C,size,"Kimi K3 is MoE 2.8T total / 104B active and K2, K2.5, K2.6 and K2.7-Code are 1T total / 32B active (all 200B and above)",,,https://platform.kimi.ai/docs/pricing/chat ; https://huggingface.co/moonshotai (model cards; fetched 2026-10-02),2026-10,"From the OpenRouter top 150 (kimi-k3 rank 4, kimi-k2.6 rank 45, kimi-k2.5 rank 84, kimi-k2.7-code rank 93)" minimax-m;minimax-h;minimax-text;abab,MiniMax,C,size,MiniMax-M3 is MoE about 428B total / 23B active and M2 230B total / 10B active (both 200B and above),,,https://huggingface.co/MiniMaxAI (model cards; fetched 2026-10-02),2026-10,"From the OpenRouter top 150 (minimax-m3 rank 58, minimax-m2.7 rank 127)" nemotron-nano;nvidia-nemotron-nano;nemotron-*-nano,NVIDIA,A,size,"Nemotron Nano is 9B and 12B dense and Nemotron-3-Nano 4B dense and MoE 30B total / 3B active, all Class A",nvidia/NVIDIA-Nemotron-Nano-9B-v2,,ML.ENERGY Leaderboard v3 snapshot (model_factors.csv in this deposit),2026-10,"Both Nemotron Nano sizes are measured in ML.ENERGY v3; the 12B row of model_factors.csv is nvidia/NVIDIA-Nemotron-Nano-12B-v2. The nvidia-nemotron-nano alternative is needed because the repository name itself begins with NVIDIA-, so stripping the nvidia/ path segment still leaves it" nemotron-*-super;nemotron-super,NVIDIA,B,size,Nemotron-3-Super is MoE 120B total / 12B active (50B to below 200B),,,https://huggingface.co/models?author=nvidia (model cards; fetched 2026-10-02),2026-10, nemotron-*-ultra;nemotron-ultra;nemotron-4-340b,NVIDIA,C,size,Nemotron-3-Ultra is MoE 550B total / 55B active and Nemotron-4 340B dense (both above the Class C cut),,,https://huggingface.co/models?author=nvidia (model cards; fetched 2026-10-02),2026-10,"From the OpenRouter top 150 (nemotron-3-ultra-550b-a55b, rank 106); the card states 550B total with 55B active" hy4;hunyuan-large,Tencent,C,size,Hy4-preview is MoE 770B total / 49B active; Hunyuan-Large 389B total / 52B active,,,https://huggingface.co/tencent (model cards; fetched 2026-10-02),2026-10,"From the OpenRouter top 150 (hy4-preview, rank 36)" hy3,Tencent,C,size,"MoE 295B total / 21B active (200B and above), 192 experts with top-8 routing",,,https://huggingface.co/tencent (model cards; fetched 2026-10-02),2026-10,From the OpenRouter top 150 (rank 114) hunyuan-a13b,Tencent,B,size,MoE 80B total / 13B active (50B to below 200B),,,https://huggingface.co/tencent (model cards; fetched 2026-10-02),2026-10, hunyuan,Tencent,C,tier,Hunyuan release whose size is not published,,,https://huggingface.co/tencent (model cards; fetched 2026-10-02),2026-10,"Catch-all below the sized hunyuan patterns. tier could not be determined from the provider's own model list or pricing page, so Class C is assigned as the conservative default (Section 3.5)" mimo-v2;mimo-*-flash;mimo-*-pro,Xiaomi,C,size,"MiMo-V2.6-Pro is MoE 1.02T total / 42B active, V2.6-Flash 309B total / 15B active and V2.5 310B total / 15B active (all 200B and above)",,,https://huggingface.co/XiaomiMiMo (model cards; fetched 2026-10-02),2026-10,"SAME REVERSAL AS DEEPSEEK AND GLM FLASH: Xiaomi's cards disclose 309B total even for the Flash variant, so the size rule puts every MiMo V2.x in Class C although 'flash' is a Class A tier word. From the OpenRouter top 150 (mimo-v2.5-pro rank 85, mimo-v2.6-pro rank 89, mimo-v2.5 rank 101, mimo-v2.6-flash rank 134, mimo-v2.6-pro-ultraspeed rank 145)" mimo-*-distill-qwen-9b,Xiaomi,A,size,9B dense distill,,,https://huggingface.co/XiaomiMiMo (model cards; fetched 2026-10-02),2026-10, mimo,Xiaomi,C,tier,MiMo release whose size is not published,,,https://huggingface.co/XiaomiMiMo (model cards; fetched 2026-10-02),2026-10,"Catch-all below the sized mimo patterns; MiMo-7B was 7B dense and so Class A, but Xiaomi publishes no size for releases other than V2.x. tier could not be determined from the provider's own model list or pricing page, so Class C is assigned as the conservative default (Section 3.5)" jev-,Typesafe,C,tier,"tier could not be determined from the provider's own model list or pricing page, so Class C is assigned as the conservative default (Section 3.5)",,,v4_inputs/raw/openrouter_rankings_month.json (top 150 by tokens),2026-10,"From the OpenRouter top 150 (typesafe/jev-1.13, rank 64). No provider documentation was reachable. tier could not be determined from the provider's own model list or pricing page, so Class C is assigned as the conservative default (Section 3.5)" inkling,Thinking Machines,C,tier,"tier could not be determined from the provider's own model list or pricing page, so Class C is assigned as the conservative default (Section 3.5)",,,v4_inputs/raw/openrouter_rankings_month.json (top 150 by tokens),2026-10,"From the OpenRouter top 150 (thinkingmachines/inkling, rank 92). tier could not be determined from the provider's own model list or pricing page, so Class C is assigned as the conservative default (Section 3.5)" aion-,AionLabs,C,tier,"tier could not be determined from the provider's own model list or pricing page, so Class C is assigned as the conservative default (Section 3.5)",,,v4_inputs/raw/openrouter_rankings_month.json (top 150 by tokens),2026-10,"From the OpenRouter top 150 (aion-labs/aion-3.0, rank 109). tier could not be determined from the provider's own model list or pricing page, so Class C is assigned as the conservative default (Section 3.5)" jamba-mini;jamba-1.5-mini,AI21,B,size,MoE 52B total / 12B active (50B to below 200B),,no,https://docs.ai21.com/docs/jamba-models,2026-10,"FLAGGED: AI21 calls this its mini tier, which would be Class A, but the disclosed 52B total is just above the 50B MoE cut and Section 3.5 gives the size precedence" jamba-large;jamba-1.5-large;jamba-instruct,AI21,C,size,MoE 398B total / 94B active (200B and above),,no,https://docs.ai21.com/docs/jamba-models,2026-10, palmyra,Writer,C,size,Palmyra X5 is above 70B dense with no published exact size,,no,https://dev.writer.com/home/models,2026-10,"tier could not be determined from the provider's own model list or pricing page, so Class C is assigned as the conservative default (Section 3.5)" sonar-deep-research,Perplexity,C,tier,"Deep research tier (USD 2.00 / 8.00 per M plus citation, reasoning and search fees)",,yes,https://docs.perplexity.ai/getting-started/models ; https://docs.perplexity.ai/models/model-cards (fetched 2026-10-02),2026-10, text-embedding;text-similarity;text-search;ada-embedding,OpenAI,out-of-scope,out-of-scope,Embedding models,,,main.tex Section 3.5 (assignment rule) and Section 3.1 (scope),2026-10,"Out of scope (Section 3.1): embedding, speech, image, video, moderation and reranking models carry no measurement under the protocol used here. Disclose as excluded with the spend involved. The calculator returns a warning and no class for these patterns" gpt-image;dall-e;gpt-5*-image;gpt-6*-image;gpt-4o*-image,OpenAI,out-of-scope,out-of-scope,Image generation and editing models,,,main.tex Section 3.5 (assignment rule) and Section 3.1 (scope),2026-10,"Out of scope (Section 3.1): embedding, speech, image, video, moderation and reranking models carry no measurement under the protocol used here. Disclose as excluded with the spend involved. The calculator returns a warning and no class for these patterns" whisper;tts-1;gpt-transcribe;gpt-live;gpt-realtime;gpt-audio;gpt-4o*-transcribe;gpt-4o-mini-tts;gpt-4o*-audio;gpt-4o*-realtime,OpenAI,out-of-scope,out-of-scope,"Speech, audio, realtime and transcription models",,,main.tex Section 3.5 (assignment rule) and Section 3.1 (scope),2026-10,"Out of scope (Section 3.1): embedding, speech, image, video, moderation and reranking models carry no measurement under the protocol used here. Disclose as excluded with the spend involved. The calculator returns a warning and no class for these patterns" omni-moderation;text-moderation,OpenAI,out-of-scope,out-of-scope,Moderation models,,,main.tex Section 3.5 (assignment rule) and Section 3.1 (scope),2026-10,"Out of scope (Section 3.1): embedding, speech, image, video, moderation and reranking models carry no measurement under the protocol used here. Disclose as excluded with the spend involved. The calculator returns a warning and no class for these patterns" sora,OpenAI,out-of-scope,out-of-scope,Video generation models,,,main.tex Section 3.5 (assignment rule) and Section 3.1 (scope),2026-10,"Out of scope (Section 3.1): embedding, speech, image, video, moderation and reranking models carry no measurement under the protocol used here. Disclose as excluded with the spend involved. The calculator returns a warning and no class for these patterns" gemini-*-flash-image;gemini-*-flash-lite-image;gemini-*-pro-image;gemini-*-image,Google,out-of-scope,out-of-scope,Image generation and editing models,,,main.tex Section 3.5 (assignment rule) and Section 3.1 (scope),2026-10,"Out of scope (Section 3.1): embedding, speech, image, video, moderation and reranking models carry no measurement under the protocol used here. Disclose as excluded with the spend involved. The calculator returns a warning and no class for these patterns" gemini-*-tts;gemini-*-live;gemini-*-transcribe;gemini-*-live-translate,Google,out-of-scope,out-of-scope,"Text-to-speech, live-audio and transcription models",,,main.tex Section 3.5 (assignment rule) and Section 3.1 (scope),2026-10,"Out of scope (Section 3.1): embedding, speech, image, video, moderation and reranking models carry no measurement under the protocol used here. Disclose as excluded with the spend involved. The calculator returns a warning and no class for these patterns" imagen;veo;lyria;chirp;diffusiongemma,Google,out-of-scope,out-of-scope,"Image, video and audio generation models",,,main.tex Section 3.5 (assignment rule) and Section 3.1 (scope),2026-10,"Out of scope (Section 3.1): embedding, speech, image, video, moderation and reranking models carry no measurement under the protocol used here. Disclose as excluded with the spend involved. The calculator returns a warning and no class for these patterns" text-embedding-gecko;gemini-embedding;embedding-001;text-multilingual-embedding,Google,out-of-scope,out-of-scope,Embedding models,,,main.tex Section 3.5 (assignment rule) and Section 3.1 (scope),2026-10,"Out of scope (Section 3.1): embedding, speech, image, video, moderation and reranking models carry no measurement under the protocol used here. Disclose as excluded with the spend involved. The calculator returns a warning and no class for these patterns" embed-english;embed-multilingual;embed-v;embed-4;rerank-,Cohere,out-of-scope,out-of-scope,Embedding and reranking models,,,main.tex Section 3.5 (assignment rule) and Section 3.1 (scope),2026-10,"Out of scope (Section 3.1): embedding, speech, image, video, moderation and reranking models carry no measurement under the protocol used here. Disclose as excluded with the spend involved. The calculator returns a warning and no class for these patterns" titan-embed;titan-image;titan-multimodal,Amazon,out-of-scope,out-of-scope,Embedding and image models,,,main.tex Section 3.5 (assignment rule) and Section 3.1 (scope),2026-10,"Out of scope (Section 3.1): embedding, speech, image, video, moderation and reranking models carry no measurement under the protocol used here. Disclose as excluded with the spend involved. The calculator returns a warning and no class for these patterns" nova-sonic;nova-*-sonic;nova-canvas;nova-reel;nova-multimodal-embeddings,Amazon,out-of-scope,out-of-scope,"Speech, image, video and embedding models",,,main.tex Section 3.5 (assignment rule) and Section 3.1 (scope),2026-10,"Out of scope (Section 3.1): embedding, speech, image, video, moderation and reranking models carry no measurement under the protocol used here. Disclose as excluded with the spend involved. The calculator returns a warning and no class for these patterns" grok-imagine;grok-*-image;grok-voice,xAI,out-of-scope,out-of-scope,"Image, video and voice models",,,main.tex Section 3.5 (assignment rule) and Section 3.1 (scope),2026-10,"Out of scope (Section 3.1): embedding, speech, image, video, moderation and reranking models carry no measurement under the protocol used here. Disclose as excluded with the spend involved. The calculator returns a warning and no class for these patterns" voxtral;mistral-ocr;mistral-embed;codestral-embed;mistral-moderation,Mistral,out-of-scope,out-of-scope,"Speech, OCR, embedding and moderation models",,,main.tex Section 3.5 (assignment rule) and Section 3.1 (scope),2026-10,"Out of scope (Section 3.1): embedding, speech, image, video, moderation and reranking models carry no measurement under the protocol used here. Disclose as excluded with the spend involved. The calculator returns a warning and no class for these patterns" muse-image;muse-voice;sam-3;sam3,Meta,out-of-scope,out-of-scope,"Image, voice and segmentation models",,,main.tex Section 3.5 (assignment rule) and Section 3.1 (scope),2026-10,"Out of scope (Section 3.1): embedding, speech, image, video, moderation and reranking models carry no measurement under the protocol used here. Disclose as excluded with the spend involved. The calculator returns a warning and no class for these patterns" seedance;seedream,ByteDance,out-of-scope,out-of-scope,Video and image generation models,,,main.tex Section 3.5 (assignment rule) and Section 3.1 (scope),2026-10,"Out of scope (Section 3.1): embedding, speech, image, video, moderation and reranking models carry no measurement under the protocol used here. Disclose as excluded with the spend involved. The calculator returns a warning and no class for these patterns" hailuo;minimax-video;minimax-speech;minimax-music;music3;speech-0,MiniMax,out-of-scope,out-of-scope,"Video, speech and music generation models",,,main.tex Section 3.5 (assignment rule) and Section 3.1 (scope),2026-10,"Out of scope (Section 3.1): embedding, speech, image, video, moderation and reranking models carry no measurement under the protocol used here. Disclose as excluded with the spend involved. The calculator returns a warning and no class for these patterns" kling-,Kuaishou,out-of-scope,out-of-scope,Video generation models,,,main.tex Section 3.5 (assignment rule) and Section 3.1 (scope),2026-10,"Out of scope (Section 3.1): embedding, speech, image, video, moderation and reranking models carry no measurement under the protocol used here. Disclose as excluded with the spend involved. The calculator returns a warning and no class for these patterns" wan-,Alibaba,out-of-scope,out-of-scope,Video generation models,,,main.tex Section 3.5 (assignment rule) and Section 3.1 (scope),2026-10,"Out of scope (Section 3.1): embedding, speech, image, video, moderation and reranking models carry no measurement under the protocol used here. Disclose as excluded with the spend involved. The calculator returns a warning and no class for these patterns" glm-ocr;glm-*-voice,Z.ai (Zhipu),out-of-scope,out-of-scope,OCR and voice models,,,main.tex Section 3.5 (assignment rule) and Section 3.1 (scope),2026-10,"Out of scope (Section 3.1): embedding, speech, image, video, moderation and reranking models carry no measurement under the protocol used here. Disclose as excluded with the spend involved. The calculator returns a warning and no class for these patterns" stable-diffusion;sdxl;stable-image;stable-video,Stability AI,out-of-scope,out-of-scope,Image and video generation models,,,main.tex Section 3.5 (assignment rule) and Section 3.1 (scope),2026-10,"Out of scope (Section 3.1): embedding, speech, image, video, moderation and reranking models carry no measurement under the protocol used here. Disclose as excluded with the spend involved. The calculator returns a warning and no class for these patterns" flux-,Black Forest Labs,out-of-scope,out-of-scope,Image generation models,,,main.tex Section 3.5 (assignment rule) and Section 3.1 (scope),2026-10,"Out of scope (Section 3.1): embedding, speech, image, video, moderation and reranking models carry no measurement under the protocol used here. Disclose as excluded with the spend involved. The calculator returns a warning and no class for these patterns" voyage-,Voyage AI,out-of-scope,out-of-scope,Embedding and reranking models,,,main.tex Section 3.5 (assignment rule) and Section 3.1 (scope),2026-10,"Out of scope (Section 3.1): embedding, speech, image, video, moderation and reranking models carry no measurement under the protocol used here. Disclose as excluded with the spend involved. The calculator returns a warning and no class for these patterns" nomic-embed;bge-;gte-;e5-;jina-embed,various,out-of-scope,out-of-scope,Open-weight embedding models,,,main.tex Section 3.5 (assignment rule) and Section 3.1 (scope),2026-10,"Out of scope (Section 3.1): embedding, speech, image, video, moderation and reranking models carry no measurement under the protocol used here. Disclose as excluded with the spend involved. The calculator returns a warning and no class for these patterns" photon;ray-,Luma,out-of-scope,out-of-scope,Image and video generation models,,,main.tex Section 3.5 (assignment rule) and Section 3.1 (scope),2026-10,"Out of scope (Section 3.1): embedding, speech, image, video, moderation and reranking models carry no measurement under the protocol used here. Disclose as excluded with the spend involved. The calculator returns a warning and no class for these patterns" marengo;pegasus,TwelveLabs,out-of-scope,out-of-scope,Video understanding models,,,main.tex Section 3.5 (assignment rule) and Section 3.1 (scope),2026-10,"Out of scope (Section 3.1): embedding, speech, image, video, moderation and reranking models carry no measurement under the protocol used here. Disclose as excluded with the spend involved. The calculator returns a warning and no class for these patterns" *,any,C,default,No pattern matched: Class C is the highest-intensity class and therefore the conservative fallback,,,"main.tex Table of decision rules (Section 3.5, row 'Unmapped model identifier') and Supplementary S6",2026-09,"THE FALLBACK RULE AS DATA. An identifier matching no other pattern is assigned Class C and flagged in the inventory. This row has a literal prefix of length zero and a literal character count of zero, so every other pattern beats it. Measure the share of activity that lands here and disclose it (main.tex Table of decision rules asks for the fallback to be flagged)"