Signal SentrySearchUpdated Nov 20, 05:30 UTCAM Drop
Models · open + commercial

New models, and whether they matter.

Which models lead, what they cost and what changed this week: charts and use-case tables from Epoch AI, LMArena and OpenRouter data, vendor claims kept apart, a dated changelog and every model story the Drops covered. Each model has a scorecard.

At a glance

How the models compare

Built from openly licensed data: Epoch AI, LMArena, OpenRouter, τ²-bench and Vectara. Every chart names its source and date, and when the newest model it has scored came out; model names link to their scorecards.

Capability

Epoch Capabilities Index

The top 10 models on Epoch AI's index, which rolls many benchmark results into one score. Higher is better.

Epoch AI · Epoch Capabilities Index: top models
RankModelIndex
1GPT-6.2 AstraOpenAI174.17
2Claude Fable 5.2Anthropic173.5
3Gemini 4 CarbonGoogle169.78
4GPT-6 AstraOpenAI169.2
5Claude Opus 5.5Anthropic168.7
6GPT-6.1 SolOpenAI168.63
7GPT-6.1 TerraOpenAI167.72
8Claude Fable 5.1Anthropic167.09
9Claude Sonnet 5.5Anthropic166.59
10Gemini 4 ArgonGoogle166.57
Colour shows the maker:
  • OpenAI
  • Anthropic
  • Google

The axis starts at 164, not zero: this score has no true zero, so the dots show position, not size. Each model's best entry; the rank is the source's.

Source: Epoch AI (CC BY), as of Nov 20, 2026 · newest model scored came out Nov 17, 2026Source

CostOur calculation

Price of a sample task

The sample task: a prompt of 10,000 tokens (about 7,500 words, like a long report or a few code files) and an answer of 2,000 tokens (about 1,500 words). What that costs each model in the Capability chart at its list price. Lower is better.

Price of a sample task: top models
ModelSample task
GPT-6.1 TerraOpenAI · $1 in · $5 out per 1M$0.02
Claude Sonnet 5.5Anthropic · $2 in · $10 out per 1M$0.04
GPT-6.1 SolOpenAI · $2 in · $10 out per 1M$0.04
Gemini 4 ArgonGoogle · $2 in · $10 out per 1M$0.04
Gemini 4 CarbonGoogle · $2 in · $10 out per 1M$0.04
Claude Opus 5.5Anthropic · $4 in · $20 out per 1M$0.08
Claude Fable 5.2Anthropic · $8 in · $40 out per 1M$0.16
GPT-6.2 AstraOpenAI · $8 in · $40 out per 1M$0.16
Claude Fable 5.1Anthropic · $10 in · $50 out per 1M$0.20
GPT-6 AstraOpenAI · $10 in · $50 out per 1M$0.20
Colour shows the maker:
  • OpenAI
  • Anthropic
  • Google

How we calculate it: (input price per 1M tokens × 10,000 + output price per 1M tokens × 2,000) ÷ 1,000,000, with each model's list price on OpenRouter. Real tasks vary: a short chat costs less, an agent run that reads many files costs far more, and reasoning models also bill their thinking as output tokens.

Example: GPT-6.2 Astra lists $8 in and $40 out per 1M tokens, so 10,000 × $8 ÷ 1M = $0.08, plus 2,000 × $40 ÷ 1M = $0.08: $0.16 a task.

Source: OpenRouter, as of Nov 20, 06:00 UTC

Source: Epoch AI (CC BY), as of Nov 20, 2026 · newest model scored came out Nov 17, 2026Source

Preference

LMArena text score

The top 10 models on LMArena's text leaderboard (style control), ranked from people's votes between two anonymous answers. Higher is better.

LMArena · Text (style control): top models
RankModelArena score
1Gemini 4 CarbonGoogleas listed: Gemini 4 Carbon (High)1,526
2Gemini 4 ArgonGoogleas listed: Gemini 4 Argon (High)1,521
3Gemini 4 NeonGoogleas listed: Gemini 4 Neon (High)1,518
4Muse WonderMetaas listed: Muse Wonder (Max)1,511
5Claude Opus 5.5Anthropicas listed: Claude Opus 5.5 (High)1,507
6Claude Fable 5Anthropicas listed: Claude Fable 5 (High)1,504
7Claude Opus 4.6Anthropicas listed: Claude Opus 4.6 (High)1,503
8Claude Fable 5.1Anthropicas listed: Claude Fable 5.1 (Max)1,502
9Claude Opus 4.7Anthropicas listed: Claude Opus 4.7 (High)1,501
10Muse Spark 1.3Metaas listed: Muse Spark 1.3 (Max)1,498
Colour shows the maker:
  • Anthropic
  • Google
  • Meta

The axis starts at 1,490, not zero: this score has no true zero, so the dots show position, not size. Each model's best entry; the rank is the source's.

LMArena (CC BY 4.0), as of Nov 18, 2026 · newest model scored came out Nov 18, 2026Source

Coding

MirrorCode

The top 10 models on Epoch AI's long-horizon coding test (built with METR): rebuild a whole program without its source code, matching the original's output exactly. Share of tasks solved; higher is better.

Epoch AI · MirrorCode: top models
RankModelAccuracy
1Claude Sonnet 5.5Anthropicas listed: claude-sonnet-5-5_max97.8%
2Claude Opus 5.5Anthropicas listed: claude-opus-5-5_max95.6%
3Claude Fable 5.1Anthropicas listed: claude-fable-5-1_max93.3%
4Claude Opus 5Anthropicas listed: claude-opus-5_max87.8%
6Claude Haiku 5.5Anthropicas listed: claude-haiku-5-5_max78.9%
7GPT-6.1 TerraOpenAIas listed: gpt-6.1-terra_max76.7%
9Claude Fable 5Anthropicas listed: claude-fable-5_max70.0%
10GPT-6.2 AstraOpenAIas listed: gpt-6.2-astra_max67.8%
11Claude Opus 4.8Anthropicas listed: claude-opus-4-8_max65.6%
12DeepSeek V4.1 FlashDeepSeekas listed: deepseek-v4.1-flash_max64.4%
Colour shows the maker:
  • OpenAI
  • Anthropic
  • DeepSeek

Each model's best entry; the rank is the source's.

Source: Epoch AI (CC BY), as of Nov 20, 2026 · newest model scored came out Nov 17, 2026Source

ValueOur calculation

Capability vs price

Every model on the Epoch Capabilities Index that OpenRouter prices. Up and to the left is better value; the dashed line joins the models nothing cheaper beats.

Scatter chart of 29 models, capability against the price of a sample task. Best value at its price: Claude Haiku 5.5, GPT-6.1 Terra, Gemini 4 Carbon, GPT-6.2 Astra.155160165170175$0.001$0.01$0.10$1.00Price of the sample task (log scale)Epoch Capabilities IndexClaude Fable 5.2 (Anthropic): index 173.5, $0.16 per sample taskGPT-6 Astra (OpenAI): index 169.2, $0.20 per sample taskClaude Opus 5.5 (Anthropic): index 168.7, $0.08 per sample taskGPT-6.1 Sol (OpenAI): index 168.63, $0.04 per sample taskClaude Fable 5.1 (Anthropic): index 167.09, $0.20 per sample taskClaude Sonnet 5.5 (Anthropic): index 166.59, $0.04 per sample taskGemini 4 Argon (Google): index 166.57, $0.04 per sample taskClaude Opus 5 (Anthropic): index 164.74, $0.10 per sample taskClaude Fable 5 (Anthropic): index 164.21, $0.20 per sample taskGPT-6 Sol (OpenAI): index 163.82, $0.04 per sample taskGPT-5.6 Sol (OpenAI): index 163.05, $0.08 per sample taskGPT-5.5 Pro (OpenAI): index 162.77, $0.66 per sample taskMuse Wonder (Meta): index 161.74, $0.08 per sample taskKimi K3.1 Pro (MoonshotAI): index 161.36, $0.02 per sample taskGPT-5.6 Terra (OpenAI): index 160.78, $0.044 per sample taskGPT-5.4 Pro (OpenAI): index 160.32, $0.66 per sample taskGPT-5.5 (OpenAI): index 160.19, $0.11 per sample taskClaude Opus 4.8 (Anthropic): index 159.91, $0.10 per sample taskDeepSeek V4.1 Flash 1017 (DeepSeek): index 159.4, $0.0027 per sample taskGemini 3.8 Flash (Google): index 158.83, $0.017 per sample taskKimi K3 (MoonshotAI): index 158.6, $0.034 per sample taskGemini 3.7 Flash (Google): index 158.38, $0.015 per sample taskMuse Spark 1.3 (Meta): index 158.25, $0.021 per sample taskGPT-5.4 (OpenAI): index 157.87, $0.055 per sample taskGrok 4.6 (xAI): index 157.82, $0.032 per sample taskGPT-6.2 Astra (OpenAI): index 174.17, $0.16 per sample taskGemini 4 Carbon (Google): index 169.78, $0.04 per sample taskGPT-6.1 Terra (OpenAI): index 167.72, $0.02 per sample taskClaude Haiku 5.5 (Anthropic): index 160.29, $0.002 per sample taskGPT-6.2 AstraGemini 4 CarbonGPT-6.1 TerraClaude Haiku 5.5Scatter chart of 29 models, capability against the price of a sample task. Best value at its price: Claude Haiku 5.5, GPT-6.1 Terra, Gemini 4 Carbon, GPT-6.2 Astra.155160165170175$0.001$0.01$0.10$1.00Price of the sample task (log scale)Epoch Capabilities IndexClaude Fable 5.2 (Anthropic): index 173.5, $0.16 per sample taskGPT-6 Astra (OpenAI): index 169.2, $0.20 per sample taskClaude Opus 5.5 (Anthropic): index 168.7, $0.08 per sample taskGPT-6.1 Sol (OpenAI): index 168.63, $0.04 per sample taskClaude Fable 5.1 (Anthropic): index 167.09, $0.20 per sample taskClaude Sonnet 5.5 (Anthropic): index 166.59, $0.04 per sample taskGemini 4 Argon (Google): index 166.57, $0.04 per sample taskClaude Opus 5 (Anthropic): index 164.74, $0.10 per sample taskClaude Fable 5 (Anthropic): index 164.21, $0.20 per sample taskGPT-6 Sol (OpenAI): index 163.82, $0.04 per sample taskGPT-5.6 Sol (OpenAI): index 163.05, $0.08 per sample taskGPT-5.5 Pro (OpenAI): index 162.77, $0.66 per sample taskMuse Wonder (Meta): index 161.74, $0.08 per sample taskKimi K3.1 Pro (MoonshotAI): index 161.36, $0.02 per sample taskGPT-5.6 Terra (OpenAI): index 160.78, $0.044 per sample taskGPT-5.4 Pro (OpenAI): index 160.32, $0.66 per sample taskGPT-5.5 (OpenAI): index 160.19, $0.11 per sample taskClaude Opus 4.8 (Anthropic): index 159.91, $0.10 per sample taskDeepSeek V4.1 Flash 1017 (DeepSeek): index 159.4, $0.0027 per sample taskGemini 3.8 Flash (Google): index 158.83, $0.017 per sample taskKimi K3 (MoonshotAI): index 158.6, $0.034 per sample taskGemini 3.7 Flash (Google): index 158.38, $0.015 per sample taskMuse Spark 1.3 (Meta): index 158.25, $0.021 per sample taskGPT-5.4 (OpenAI): index 157.87, $0.055 per sample taskGrok 4.6 (xAI): index 157.82, $0.032 per sample taskGPT-6.2 Astra (OpenAI): index 174.17, $0.16 per sample taskGemini 4 Carbon (Google): index 169.78, $0.04 per sample taskGPT-6.1 Terra (OpenAI): index 167.72, $0.02 per sample taskClaude Haiku 5.5 (Anthropic): index 160.29, $0.002 per sample taskGPT-6.2 AstraGemini 4 CarbonGPT-6.1 TerraClaude Haiku 5.5
Colour shows the maker:
  • OpenAI
  • Anthropic
  • Google
  • Other makers
  • Best value at its price

1 model on the index has no OpenRouter price and is not plotted.

Show the numbers
Each model's capability index and sample-task price
ModelEpoch Capabilities IndexSample taskBest value
GPT-6.2 AstraOpenAI174.17$0.16Yes
Claude Fable 5.2Anthropic173.5$0.16
Gemini 4 CarbonGoogle169.78$0.04Yes
GPT-6 AstraOpenAI169.2$0.20
Claude Opus 5.5Anthropic168.7$0.08
GPT-6.1 SolOpenAI168.63$0.04
GPT-6.1 TerraOpenAI167.72$0.02Yes
Claude Fable 5.1Anthropic167.09$0.20
Claude Sonnet 5.5Anthropic166.59$0.04
Gemini 4 ArgonGoogle166.57$0.04
Claude Opus 5Anthropic164.74$0.10
Claude Fable 5Anthropic164.21$0.20
GPT-6 SolOpenAI163.82$0.04
GPT-5.6 SolOpenAI163.05$0.08
GPT-5.5 ProOpenAI162.77$0.66
Muse WonderMeta161.74$0.08
Kimi K3.1 ProMoonshotAI161.36$0.02
GPT-5.6 TerraOpenAI160.78$0.044
GPT-5.4 ProOpenAI160.32$0.66
Claude Haiku 5.5Anthropic160.29$0.002Yes
GPT-5.5OpenAI160.19$0.11
Claude Opus 4.8Anthropic159.91$0.10
DeepSeek V4.1 Flash 1017DeepSeek159.4$0.0027
Gemini 3.8 FlashGoogle158.83$0.017
Kimi K3MoonshotAI158.6$0.034
Gemini 3.7 FlashGoogle158.38$0.015
Muse Spark 1.3Meta158.25$0.021
GPT-5.4OpenAI157.87$0.055
Grok 4.6xAI157.82$0.032

Our calculation: Epoch AI's index against the sample-task price (10,000 tokens in, 2,000 out, at each model's OpenRouter list price; the formula is under the price chart).

Source: Epoch AI (CC BY), as of Nov 20, 2026 · newest model scored came out Nov 17, 2026Source

Source: OpenRouter, as of Nov 20, 06:00 UTC

No speed chart: OpenRouter shares speed only with signed-in API callers, and no open source publishes it. We will not estimate it.

Left out until their sources catch up (we show a benchmark only if it has scored a model that came out in the last 60 days):

  • Epoch AI · swe_bench_verified: newest model scored came out Jun 16, 2026
  • τ²-bench · Banking (pass^1): newest model scored came out Jul 24, 2026
  • τ²-bench · Airline (pass^1): newest model scored came out Feb 16, 2026
  • τ²-bench · Retail (pass^1): newest model scored came out Feb 16, 2026
  • τ²-bench · Telecom (pass^1): newest model scored came out Feb 16, 2026
By use case

Benchmarks by use case

Independent results grouped the way launch posts compare models. Rows are models we can match to OpenRouter; each cell is the model's best entry, and the strongest tint marks the best in its column.

Coding

MirrorCode (Epoch AI's long-horizon test: rebuild a whole program without its source), web apps people vote on (LMArena WebDev), and real GitHub issues (SWE-bench Verified) while Epoch keeps it current.

Coding: independent results for the models we can match
ModelMirrorCode↑ higher is betterLMArena WebDev↑ higher is better
Claude Sonnet 5.5Anthropic97.8%best1,773
Claude Opus 5.5Anthropic95.6%1,818
Claude Fable 5.1Anthropic93.3%1,741
Claude Opus 5Anthropic87.8%1,691
Claude Haiku 5.5Anthropic78.9%–
GPT-6.2 AstraOpenAI67.8%1,781
Claude Fable 5.2Anthropic–1,824best
Gemini 4 CarbonGoogle–1,769

Sorted by MirrorCode. A dash: that source has no result for the model among the rows we keep, or we could not match its name to the model.

Left out as out of date: SWE-bench Verified: newest model scored came out Jun 16, 2026.

  • MirrorCode: Source: Epoch AI (CC BY), as of Nov 20, 2026 · newest model scored came out Nov 17, 2026Source
  • LMArena WebDev: LMArena (CC BY 4.0), as of Nov 18, 2026 · newest model scored came out Nov 18, 2026Source

Agents and tool use

τ²-bench's customer-service tasks with tools (airline on OpenRouter's own runs, banking on Sierra's leaderboard) and LMArena's agent votes. No computer-use benchmark is openly licensed yet: vendors' claims are under Vendor-reported.

Agents and tool use: independent results for the models we can match
Modelτ²-bench airline (OpenRouter)↑ higher is betterLMArena Agent↑ higher is better
DeepSeek V4 Flash 0423DeepSeek81.7%best–
DeepSeek V4 Flash 0731DeepSeek79.8%–
Qwen3.5 397B A17BQwen79.4%–
Ling 3.1 FlashinclusionAI79.3%–
Claude Opus 5.5Anthropic74.9%0.186
Claude Fable 5.2Anthropic–0.274best
Gemini 4 NeonGoogle–0.184
Gemini 4 CarbonGoogle–0.161

Sorted by τ²-bench airline (OpenRouter). A dash: that source has no result for the model among the rows we keep, or we could not match its name to the model.

Left out as out of date: τ²-bench banking: newest model scored came out Jul 24, 2026.

  • τ²-bench airline (OpenRouter): Source: OpenRouter (openrouter.ai), licensed under CC BY 4.0, as of Nov 20, 2026 · newest model scored came out Oct 23, 2026Source
  • LMArena Agent: LMArena (CC BY 4.0), as of Nov 18, 2026 · newest model scored came out Nov 18, 2026Source

Reasoning and knowledge

Graduate-level science questions (run by Epoch AI and by OpenRouter), and competition and research maths.

Reasoning and knowledge: independent results for the models we can match
ModelGPQA Diamond (Epoch)↑ higher is betterGPQA Diamond (OpenRouter)↑ higher is betterFrontierMath 1–3↑ higher is betterFrontierMath 4↑ higher is betterOTIS Mock AIME↑ higher is better
GPT-6.2 AstraOpenAI96.8%best–92.6%100.0%best100.0%best
Gemini 4 HydrogenGoogle96.8%–88.1%87.8%100.0%best
Gemini 4 CarbonGoogle96.5%–90.5%95.1%100.0%best
Gemini 4 ArgonGoogle96.4%96.5%best86.0%78.0%100.0%best
Claude Sonnet 5.5Anthropic95.6%92.9%88.8%80.5%100.0%best
Gemini 3.8 FlashGoogle95.4%95.3%––98.9%
GPT-6.1 SolOpenAI95.4%94.4%93.7%100.0%best100.0%best
Claude Fable 5.2Anthropic93.9%95.1%95.1%best100.0%best99.7%
  • GPT-6.2 AstraOpenAI
    GPQA Diamond (Epoch)
    96.8%best
    GPQA Diamond (OpenRouter)
    –
    FrontierMath 1–3
    92.6%
    FrontierMath 4
    100.0%best
    OTIS Mock AIME
    100.0%best
  • Gemini 4 HydrogenGoogle
    GPQA Diamond (Epoch)
    96.8%
    GPQA Diamond (OpenRouter)
    –
    FrontierMath 1–3
    88.1%
    FrontierMath 4
    87.8%
    OTIS Mock AIME
    100.0%best
  • Gemini 4 CarbonGoogle
    GPQA Diamond (Epoch)
    96.5%
    GPQA Diamond (OpenRouter)
    –
    FrontierMath 1–3
    90.5%
    FrontierMath 4
    95.1%
    OTIS Mock AIME
    100.0%best
  • Gemini 4 ArgonGoogle
    GPQA Diamond (Epoch)
    96.4%
    GPQA Diamond (OpenRouter)
    96.5%best
    FrontierMath 1–3
    86.0%
    FrontierMath 4
    78.0%
    OTIS Mock AIME
    100.0%best
  • Claude Sonnet 5.5Anthropic
    GPQA Diamond (Epoch)
    95.6%
    GPQA Diamond (OpenRouter)
    92.9%
    FrontierMath 1–3
    88.8%
    FrontierMath 4
    80.5%
    OTIS Mock AIME
    100.0%best
  • Gemini 3.8 FlashGoogle
    GPQA Diamond (Epoch)
    95.4%
    GPQA Diamond (OpenRouter)
    95.3%
    FrontierMath 1–3
    –
    FrontierMath 4
    –
    OTIS Mock AIME
    98.9%
  • GPT-6.1 SolOpenAI
    GPQA Diamond (Epoch)
    95.4%
    GPQA Diamond (OpenRouter)
    94.4%
    FrontierMath 1–3
    93.7%
    FrontierMath 4
    100.0%best
    OTIS Mock AIME
    100.0%best
  • Claude Fable 5.2Anthropic
    GPQA Diamond (Epoch)
    93.9%
    GPQA Diamond (OpenRouter)
    95.1%
    FrontierMath 1–3
    95.1%best
    FrontierMath 4
    100.0%best
    OTIS Mock AIME
    99.7%

Sorted by GPQA Diamond (Epoch). A dash: that source has no result for the model among the rows we keep, or we could not match its name to the model.

  • GPQA Diamond (Epoch), FrontierMath 1–3, FrontierMath 4, OTIS Mock AIME: Source: Epoch AI (CC BY), as of Nov 20, 2026 · newest model scored came out Nov 17, 2026Source
  • GPQA Diamond (OpenRouter): Source: OpenRouter (openrouter.ai), licensed under CC BY 4.0, as of Nov 20, 2026 · newest model scored came out Nov 9, 2026Source

Puzzles and spatial reasoning

Chess puzzles, mystery-game deduction and furniture assembly, as run by Epoch AI.

Puzzles and spatial reasoning: independent results for the models we can match
ModelChess Puzzles↑ higher is betterMystery Game Puzzles↑ higher is betterFurniture Assembly↑ higher is better
Claude Fable 5.2Anthropic79.0%best90.0%90.0%best
GPT-6.2 AstraOpenAI78.0%95.0%best89.2%
Gemini 4 CarbonGoogle78.0%73.0%77.5%
GPT-6 AstraOpenAI76.0%84.0%80.0%
GPT-6.1 SolOpenAI68.0%80.0%80.0%
GPT-5.6 SolOpenAI68.0%58.0%56.7%
Claude Opus 5.5Anthropic67.0%71.0%83.3%
GPT-6.1 TerraOpenAI–74.0%79.2%
  • Claude Fable 5.2Anthropic
    Chess Puzzles
    79.0%best
    Mystery Game Puzzles
    90.0%
    Furniture Assembly
    90.0%best
  • GPT-6.2 AstraOpenAI
    Chess Puzzles
    78.0%
    Mystery Game Puzzles
    95.0%best
    Furniture Assembly
    89.2%
  • Gemini 4 CarbonGoogle
    Chess Puzzles
    78.0%
    Mystery Game Puzzles
    73.0%
    Furniture Assembly
    77.5%
  • GPT-6 AstraOpenAI
    Chess Puzzles
    76.0%
    Mystery Game Puzzles
    84.0%
    Furniture Assembly
    80.0%
  • GPT-6.1 SolOpenAI
    Chess Puzzles
    68.0%
    Mystery Game Puzzles
    80.0%
    Furniture Assembly
    80.0%
  • GPT-5.6 SolOpenAI
    Chess Puzzles
    68.0%
    Mystery Game Puzzles
    58.0%
    Furniture Assembly
    56.7%
  • Claude Opus 5.5Anthropic
    Chess Puzzles
    67.0%
    Mystery Game Puzzles
    71.0%
    Furniture Assembly
    83.3%
  • GPT-6.1 TerraOpenAI
    Chess Puzzles
    –
    Mystery Game Puzzles
    74.0%
    Furniture Assembly
    79.2%

Sorted by Chess Puzzles. A dash: that source has no result for the model among the rows we keep, or we could not match its name to the model.

  • Chess Puzzles, Mystery Game Puzzles, Furniture Assembly: Source: Epoch AI (CC BY), as of Nov 20, 2026 · newest model scored came out Nov 17, 2026Source

Factual accuracy

Short factual answers (higher is better), and how often a model adds something a document does not say when it summarizes it (lower is better).

Factual accuracy: independent results for the models we can match
ModelSimpleQA Verified↑ higher is betterHallucination in summaries↓ lower is better
Claude Fable 5.2Anthropic94.0%best–
GPT-6.2 AstraOpenAI78.5%–
Gemini 4 CarbonGoogle77.7%–
Gemini 4 ArgonGoogle76.0%12.1%
GPT-5.4 NanoOpenAI–3.1%best
Gemini 2.5 Flash LiteGoogle–3.3%
Phi 4Microsoft–3.7%
DeepSeek V4.1 FlashDeepSeek–4.2%
  • Claude Fable 5.2Anthropic
    SimpleQA Verified
    94.0%best
    Hallucination in summaries
    –
  • GPT-6.2 AstraOpenAI
    SimpleQA Verified
    78.5%
    Hallucination in summaries
    –
  • Gemini 4 CarbonGoogle
    SimpleQA Verified
    77.7%
    Hallucination in summaries
    –
  • Gemini 4 ArgonGoogle
    SimpleQA Verified
    76.0%
    Hallucination in summaries
    12.1%
  • GPT-5.4 NanoOpenAI
    SimpleQA Verified
    –
    Hallucination in summaries
    3.1%best
  • Gemini 2.5 Flash LiteGoogle
    SimpleQA Verified
    –
    Hallucination in summaries
    3.3%
  • Phi 4Microsoft
    SimpleQA Verified
    –
    Hallucination in summaries
    3.7%
  • DeepSeek V4.1 FlashDeepSeek
    SimpleQA Verified
    –
    Hallucination in summaries
    4.2%

Sorted by SimpleQA Verified. A dash: that source has no result for the model among the rows we keep, or we could not match its name to the model.

  • SimpleQA Verified: Source: Epoch AI (CC BY), as of Nov 20, 2026 · newest model scored came out Nov 17, 2026Source
  • Hallucination in summaries: Vectara Hallucination Leaderboard (HHEM-2.3), Apache-2.0, as of Oct 22, 2026 · newest model scored came out Oct 13, 2026Source

Long context

How much text each model accepts at once, for the models on the Epoch Capabilities Index: a listed limit, not a measured score. No openly licensed long-context benchmark is available yet; vendors' own long-context claims are under Vendor-reported.

Long context: independent results for the models we can match
ModelContext window↑ more text at once
GPT-6 AstraOpenAI1.05Mbest
GPT-6.1 SolOpenAI1.05Mbest
GPT-6.1 TerraOpenAI1.05Mbest
GPT-6.2 AstraOpenAI1.05Mbest
Gemini 4 CarbonGoogle1.05M
Claude Fable 5.1Anthropic1M
Claude Fable 5.2Anthropic1M
Claude Opus 5.5Anthropic1M

Sorted by Context window. A dash: that source has no result for the model among the rows we keep, or we could not match its name to the model.

  • Context window: Source: OpenRouter, as of Nov 20, 06:00 UTC

Price and speed

For the models on the Epoch Capabilities Index: our sample-task price (10,000 tokens in and 2,000 out at the model's OpenRouter list price; the formula is under the price chart), and OpenRouter's measured cost and time per task on its own GPQA Diamond and τ²-bench runs. Time per task is generation time, not a speed; no openly licensed tokens-per-second source exists yet.

Price and speed: independent results for the models we can match
ModelSample task↓ lower is betterCost per task (GPQA)↓ lower is betterCost per task (τ² airline)↓ lower is betterTime per task (GPQA)↓ lower is betterTime per task (τ² airline)↓ lower is better
GPT-6.1 TerraOpenAI$0.02best$0.011best–28 s–
GPT-6.1 SolOpenAI$0.04$0.021–48 s–
Gemini 4 CarbonGoogle$0.04––––
Claude Opus 5.5Anthropic$0.08$0.054$0.35best27 s98 sbest
Claude Fable 5.2Anthropic$0.16$0.087–26 sbest–
GPT-6.2 AstraOpenAI$0.16––––
GPT-6 AstraOpenAI$0.20$0.11–48 s–
Claude Fable 5.1Anthropic$0.20$0.14$0.8140 s182 s
  • GPT-6.1 TerraOpenAI
    Sample task
    $0.02best
    Cost per task (GPQA)
    $0.011best
    Cost per task (τ² airline)
    –
    Time per task (GPQA)
    28 s
    Time per task (τ² airline)
    –
  • GPT-6.1 SolOpenAI
    Sample task
    $0.04
    Cost per task (GPQA)
    $0.021
    Cost per task (τ² airline)
    –
    Time per task (GPQA)
    48 s
    Time per task (τ² airline)
    –
  • Gemini 4 CarbonGoogle
    Sample task
    $0.04
    Cost per task (GPQA)
    –
    Cost per task (τ² airline)
    –
    Time per task (GPQA)
    –
    Time per task (τ² airline)
    –
  • Claude Opus 5.5Anthropic
    Sample task
    $0.08
    Cost per task (GPQA)
    $0.054
    Cost per task (τ² airline)
    $0.35best
    Time per task (GPQA)
    27 s
    Time per task (τ² airline)
    98 sbest
  • Claude Fable 5.2Anthropic
    Sample task
    $0.16
    Cost per task (GPQA)
    $0.087
    Cost per task (τ² airline)
    –
    Time per task (GPQA)
    26 sbest
    Time per task (τ² airline)
    –
  • GPT-6.2 AstraOpenAI
    Sample task
    $0.16
    Cost per task (GPQA)
    –
    Cost per task (τ² airline)
    –
    Time per task (GPQA)
    –
    Time per task (τ² airline)
    –
  • GPT-6 AstraOpenAI
    Sample task
    $0.20
    Cost per task (GPQA)
    $0.11
    Cost per task (τ² airline)
    –
    Time per task (GPQA)
    48 s
    Time per task (τ² airline)
    –
  • Claude Fable 5.1Anthropic
    Sample task
    $0.20
    Cost per task (GPQA)
    $0.14
    Cost per task (τ² airline)
    $0.81
    Time per task (GPQA)
    40 s
    Time per task (τ² airline)
    182 s

Sorted by Sample task. A dash: that source has no result for the model among the rows we keep, or we could not match its name to the model.

  • Sample task: Our calculation, Source: OpenRouter, as of Nov 20, 06:00 UTC
  • Cost per task (GPQA), Time per task (GPQA): Source: OpenRouter (openrouter.ai), licensed under CC BY 4.0, as of Nov 20, 2026 · newest model scored came out Nov 9, 2026Source
  • Cost per task (τ² airline), Time per task (τ² airline): Source: OpenRouter (openrouter.ai), licensed under CC BY 4.0, as of Nov 20, 2026 · newest model scored came out Oct 23, 2026Source
Vendor-reported

What the vendors say

Scores quoted word for word from a vendor's own launch post or model card, as the Drops collected them. Claims, not measurements: no independent lab or Signal Sentry measured them, and they never appear in the tables above.

Coding

Agents and tool use

Reasoning and knowledge

Other benchmarks

What changed

Model changelog

New models, settled price changes and new benchmark results from our snapshots, with the Drops' model stories. Newest first, last 14 days.

  • BenchmarkGPT-6.2 Astra (OpenAI) enters LMArena · WebDev at #3 with 1,781.as listed: GPT-6.2 Astra (Max)
  • BenchmarkGPT-6.2 Astra (OpenAI) enters LMArena · Agent at #6 with 0.152.as listed: GPT-6.2 Astra (Max)
  • BenchmarkGPT-6.2 Sol (OpenAI) enters LMArena · WebDev at #8 with 1,757.as listed: GPT-6.2 Sol (Max)
  • BenchmarkGPT-6.2 Sol (OpenAI) enters LMArena · Agent at #11 with 0.113.as listed: GPT-6.2 Sol (Max)
  • BenchmarkGemini 4 Hydrogen (Google) enters Epoch AI · GPQA diamond at #2 with 96.8%.as listed: gemini-4-hydrogen_xhigh
  • BenchmarkGemini 4 Hydrogen (Google) enters Epoch AI · FrontierMath-Tiers-1-3-v2-Private at #12 with 88.1%.as listed: gemini-4-hydrogen_xhigh
  • BenchmarkGemini 4 Hydrogen (Google) enters Epoch AI · FrontierMath-Tier-4-v2-Private at #13 with 87.8%.as listed: gemini-4-hydrogen_xhigh
  • BenchmarkGemini 4 Hydrogen (Google) enters Epoch AI · SimpleQA Verified at #9 with 70.9%.as listed: gemini-4-hydrogen_xhigh
  • BenchmarkGemini 4 Hydrogen (Google) enters Epoch AI · OTIS Mock AIME 2024-2025 at #1 with 100.0%.as listed: gemini-4-hydrogen_xhigh
  • BenchmarkGemini 4 Hydrogen (Google) enters Epoch AI · Chess Puzzles at #12 with 60.0%.as listed: gemini-4-hydrogen_xhigh
  • BenchmarkGemini 4 Hydrogen (Google) enters Epoch AI · Mystery Game Puzzles at #8 with 65.0%.as listed: gemini-4-hydrogen_xhigh
  • BenchmarkGemini 4 Hydrogen (Google) enters Epoch AI · Furniture Assembly at #13 with 57.5%.as listed: gemini-4-hydrogen_xhigh
  • BenchmarkGemini 4 Neon (Google) enters LMArena · Text (style control) at #3 with 1,518.as listed: Gemini 4 Neon (High)
  • BenchmarkGemini 4 Neon (Google) enters LMArena · WebDev at #6 with 1,764.as listed: Gemini 4 Neon (High)
  • BenchmarkGemini 4 Neon (Google) enters LMArena · Agent at #3 with 0.184.as listed: Gemini 4 Neon (High)

  • NewHy4 flash (Tencent) is on OpenRouter: $0.035 in · $0.1405 out per 1M, 1.05M context.In Drop #091 →
  • NewQwen4 Max (Qwen) is on OpenRouter: $1.5 in · $4.5 out per 1M, 1M context.In Drop #089 →
  • BenchmarkGemini 4 Hydrogen (Google) enters LMArena · Agent at #5 with 0.149.as listed: Gemini 4 Hydrogen (High)
  • BenchmarkQwen4 Max (Qwen) enters LMArena · WebDev at #6 with 1,750.as listed: Qwen4-Max (xHigh)
  • BenchmarkQwen4 Max (Qwen) enters LMArena · Agent at #14 with 0.0746.as listed: Qwen4-Max (xHigh)
  • PriceGLM 5.3 (Z.ai) price cut on OpenRouter: $0.084 in · $3.6 out → $0.0699 in · $1.4 out per 1M.
  • PriceQwen3.5 397B A17B (Qwen) price cut on OpenRouter: $0.6 in · $3.6 out → $0.45 in · $3 out per 1M.
6 earlier days

  • GoneLFM2.5-2.6B (LiquidAI) is no longer in OpenRouter's priced list.
  • GoneLing 3.0 Flash Fin (inclusionAI) is no longer in OpenRouter's priced list.
  • GoneLing 3.0 Flash Sante (inclusionAI) is no longer in OpenRouter's priced list.
  • BenchmarkGemini 4 Hydrogen (Google) enters LMArena · Text (style control) at #15 with 1,490.as listed: Gemini 4 Hydrogen (High)
  • BenchmarkGemini 4 Hydrogen (Google) enters LMArena · WebDev at #10 with 1,704.as listed: Gemini 4 Hydrogen (High)
  • BenchmarkInkling 2 Small (Thinking Machines) enters OpenRouter · GPQA Diamond: cost per task at #20 with $0.031.
  • Benchmarkclaude-fable-5-2-fallback_max (Anthropic) enters Epoch AI · GPQA diamond at #3 with 96.0%.
  • Benchmarkclaude-fable-5-fallback_max (Anthropic) enters Epoch AI · GPQA diamond at #20 with 93.9%.
  • Benchmarkclaude-opus-5-5-fallback_max (Anthropic) enters Epoch AI · GPQA diamond at #16 with 94.1%.
  • PriceGLM 5.3 (Z.ai) price rise on OpenRouter: $0.0699 in · $1.4 out → $0.084 in · $3.6 out per 1M.
  • PriceKimi K2.5 (MoonshotAI) price cut on OpenRouter: $0.532 in · $3.325 out → $0.5 in · $2.5 out per 1M.
  • PriceMiMo-V2.5-Pro (Xiaomi) price cut on OpenRouter: $0.48 in · $1.8 out → $0.435 in · $0.87 out per 1M.
  • PriceRipple 65B Preview (Ricursive) price cut on OpenRouter: $0.2 in · $0.6 out → $0.1 in · $0.3 out per 1M.
  • DropEdge0 open-sources a 145B model it says runs on a 48GB laptop#084
  • Dropllama.cpp reaches v1.0.0 and ties any API break to a major-version bump#084

  • GoneERNIE-5.5-Flash-Preview (Baidu) is no longer in OpenRouter's priced list.
  • GoneSeed-2.0-Code (ByteDance Seed) is no longer in OpenRouter's priced list.
  • New #1Claude Fable 5.2 (Anthropic) now leads LMArena · WebDev with 1,836.as listed: Claude Fable 5.2 (xHigh)
  • New #1Claude Sonnet 5.5 (Anthropic) now leads Epoch AI · MirrorCode with 97.8%.as listed: claude-sonnet-5-5_max
  • New #1Gemini 4 Carbon (Google) now leads Epoch AI · GPQA diamond with 96.5%.as listed: gemini-4-carbon_high
  • BenchmarkClaude Fable 5.2 (Anthropic) enters OpenRouter · GPQA Diamond at #5 with 95.1%.
  • BenchmarkClaude Fable 5.2 (Anthropic) enters OpenRouter · GPQA Diamond: time per task at #2 with 26 s.
  • BenchmarkClaude Sonnet 5 (Anthropic) enters Epoch AI · MirrorCode at #13 with 63.3%.as listed: claude-sonnet-5_max
  • BenchmarkDeepSeek V4.1 Flash (DeepSeek) enters Epoch AI · MirrorCode at #11 with 64.4%.as listed: deepseek-v4.1-flash_max
  • BenchmarkGPT-6.1 Terra (OpenAI) enters Epoch AI · MirrorCode at #7 with 76.7%.as listed: gpt-6.1-terra_max
  • BenchmarkGemini 4 Carbon (Google) enters Epoch AI · FrontierMath-Tiers-1-3-v2-Private at #6 with 90.5%.as listed: gemini-4-carbon_high
  • BenchmarkGemini 4 Carbon (Google) enters Epoch AI · FrontierMath-Tier-4-v2-Private at #7 with 95.1%.as listed: gemini-4-carbon_high
  • BenchmarkGemini 4 Carbon (Google) enters Epoch AI · SimpleQA Verified at #2 with 77.7%.as listed: gemini-4-carbon_high
  • BenchmarkGemini 4 Carbon (Google) enters Epoch AI · OTIS Mock AIME 2024-2025 at #1 with 100.0%.as listed: gemini-4-carbon_high
  • BenchmarkGemini 4 Carbon (Google) enters Epoch AI · Chess Puzzles at #2 with 78.0%.as listed: gemini-4-carbon_high
  • BenchmarkGemini 4 Carbon (Google) enters Epoch AI · Mystery Game Puzzles at #5 with 73.0%.as listed: gemini-4-carbon_high
  • BenchmarkGemini 4 Carbon (Google) enters Epoch AI · Furniture Assembly at #6 with 77.5%.as listed: gemini-4-carbon_high
  • BenchmarkGrok 4.7 (xAI) enters Epoch AI · MirrorCode at #17 with 54.4%.as listed: grok-4.7_xhigh
  • BenchmarkMuse Spark 1.3 (Meta) enters Epoch AI · MirrorCode at #17 with 54.4%.as listed: muse-spark-1.3_max
  • PriceDeepSeek V4 Flash 0423 (DeepSeek) price cut on OpenRouter: $0.091 in · $0.182 out → $0.09 in · $0.18 out per 1M.
  • DropDeepMind's AlphaGenome 2 predicts DNA variant effects cell type by cell type#080
  • DropTether's 0.8B BotPsy targets on-device agents and tool calls#080

19 models' OpenRouter prices moved back and forth between runs, so their changes are left out.

31 older entries are not shown (the list is capped).

From our models snapshots (Source: OpenRouter; Source: Epoch AI (CC BY); LMArena (CC BY 4.0)) and the Drops.

Price board

Newest models: price and context

The 15 newest of the 195 models in our latest snapshot, newest on OpenRouter first. Prices are US dollars per 1M tokens.

Newest models with their price per 1M input and output tokens and their context window
ModelMakerAddedIn $/1MOut $/1MContext
Pixtral Large 2502MistralNov 20, 2026$2$6128K
Gemini 4 NeonGoogleNov 18, 2026$3$151.05M
GPT-6.2 Astra ProOpenAINov 17, 2026$8$401.05M
GPT-6.2 AstraOpenAINov 17, 2026$8$401.05M
GPT-6.2 Sol ProOpenAINov 17, 2026$2$101.05M
GPT-6.2 SolOpenAINov 17, 2026$2$101.05M
Hy4 flashTencentNov 16, 2026$0.035$0.14051.05M
Qwen4 MaxQwenNov 16, 2026$1.425$4.2751M
Gemini 4 HydrogenGoogleNov 12, 2026$1.5$7.51.05M
ERNIE-Light-DT-1.0BaiduNov 10, 2026$1.13$3.941.02M
Gemini 4 CarbonGoogleNov 10, 2026$2$101.05M
Nex-N2.6Nex AGINov 10, 2026$0.12$0.4262K
Inkling 2 SmallThinking MachinesNov 9, 2026$0.3$11.05M
Claude Fable 5.2AnthropicNov 9, 2026$8$401M
Muse EmberMetaNov 9, 2026$0.1$0.51.05M
See 85 more models
More of the newest models with their price per 1M input and output tokens and their context window
ModelMakerAddedIn $/1MOut $/1MContext
LFM3-1.7BLiquidAINov 6, 2026$0.08$0.2262K
Gemini Omni 1.2 FlashGoogleNov 5, 2026$1.5$9131K
Mi:dm K 3.0 ProKTNov 5, 2026$0.2$0.8262K
Gemini Omni 1.1 FlashGoogleNov 4, 2026$1.5$9131K
Ripple 65B PreviewRicursiveNov 4, 2026$0.1$0.31.05M
Hy4TencentNov 4, 2026$0.417$1.25051.05M
Motif 3Motif TechnologiesNov 2, 2026$0.3$1.2262K
GLM 5.3 FlashX LegalZ.aiNov 2, 2026$0.37$1.251.05M
Particle-1Particle AIOct 28, 2026$0.07$0.51.05M
GPT-6.1 Terra ProOpenAIOct 26, 2026$1$51.05M
GPT-6.1 TerraOpenAIOct 26, 2026$1$51.05M
Kimi K3.1 ProMoonshotAIOct 23, 2026$0.253$8.81.05M
MiniMax M3.1 FlashMiniMaxOct 19, 2026$0.15$0.61.05M
DeepSeek V4.1 Flash 1017DeepSeekOct 17, 2026$0.15$0.61.05M
GlimpseElorianOct 17, 2026$0.6$3262K
Beam 1.0 501B A23BReflectionOct 16, 2026$0.25$11.05M
Ternary Bonsai 2 9BPrismMLOct 16, 2026$0.08$0.13262K
Moondream 3 PreviewMoondreamOct 15, 2026$0.3$2.532K
Moondream 3.1MoondreamOct 15, 2026$0.3$2.532K
Muse Spark OpenMetaOct 15, 2026$1.25$4.251.05M
Nemotron 3.5 Super VLNVIDIAOct 15, 2026$0.15$0.51.05M
Pareto 26.10n/aOct 15, 2026$0.8$3.21.05M
Pareto 26.9n/aOct 13, 2026$2.5$7.5262K
LFM2.5-24B-A2BLiquidAIOct 13, 2026$0.05$0.7128K
Gemini 4 ArgonGoogleOct 13, 2026$2$101.05M
SummarizeScaleDownOct 13, 2026$0.04$01M
ExtractScaleDownOct 13, 2026$0.04$01M
Muse WonderMetaOct 12, 2026$4$201.05M
Step 5 PreviewStepFunOct 8, 2026$1$2.71M
Claude Haiku 5.5AnthropicOct 7, 2026$0.1$0.51M
Nano Banana 2.1GoogleOct 6, 2026$1.5$7.566K
Mistral Large 4MistralOct 6, 2026$0.7$2.11.05M
Ling 3.1 FlashinclusionAIOct 2, 2026$0.045$0.1351M
Pareto 26.10 Previewn/aOct 1, 2026$0.8$3.21.05M
GPT-6.1 Sol ProOpenAISep 29, 2026$2$101.05M
GPT-6.1 SolOpenAISep 29, 2026$2$101.05M
Claude Sonnet 5.5AnthropicSep 28, 2026$2$101M
Perceptron Mk1.5PerceptronSep 25, 2026$0.15$1.537K
Ember-1FireworksSep 24, 2026$3$151.05M
GLM 5.3 PrimeZ.aiSep 23, 2026$1.68$5.281M
Qwen3.8 Max PrimeQwenSep 23, 2026$4$121M
Aion 3.5 MiniAionLabsSep 23, 2026$0.7$1.4262K
Aion 3.5AionLabsSep 23, 2026$3$6262K
Solar Mini 4UpstageSep 23, 2026$0.05$0.2524K
Command A+CohereSep 22, 2026$0.3$1.5192K
GPT-6 Luna ProOpenAISep 22, 2026$0.1$0.51.05M
GPT-6 LunaOpenAISep 22, 2026$0.1$0.51.05M
GPT-6 Sol ProOpenAISep 22, 2026$2$101.05M
GPT-6 SolOpenAISep 22, 2026$2$101.05M
Claude Opus 5.5AnthropicSep 22, 2026$4$201M
MiMo-V2.6-Pro-UltraSpeedXiaomiSep 21, 2026$4.35$8.71.05M
MiMo-V2.6-FlashXiaomiSep 21, 2026$0.14$0.281.05M
MiMo-V2.6-ProXiaomiSep 21, 2026$0.435$0.871.05M
Grok 4.7SpaceXSISep 21, 2026$2$6500K
Qwen3.8 Omni FlashQwenSep 21, 2026$0.15$0.471M
Ternary Bonsai 2 27BPrismMLSep 18, 2026$0.075$0.5262K
GLM 5.3 FlashXZ.aiSep 18, 2026$0.37$1.251.05M
Schematron V2 TurboInference.netSep 12, 2026$0.03$0.15128K
Schematron V2 SmallInference.netSep 12, 2026$0.05$0.23128K
Fugu Ultra v2SakanaSep 11, 2026$5$301M
Fugu MaxSakanaSep 11, 2026$2$61M
Ling 3.0 Flash VLinclusionAISep 10, 2026$0.06$0.18131K
DeepSeek V4.1 FlashDeepSeekSep 10, 2026$0.3$1.21.05M
Nex-N2.5-MiniNex AGISep 8, 2026$0.025$0.1262K
Nex-N2.5-ProNex AGISep 8, 2026$0.075$0.25262K
GPT-6 AstraOpenAISep 4, 2026$10$501.05M
GPT-6 Astra ProOpenAISep 4, 2026$10$501.05M
Qwen3.8 Max (0902)QwenSep 3, 2026$2$61M
Muse Spark 1.3 ContributorMetaSep 2, 2026$0.1$0.21.05M
Muse Spark 1.3MetaSep 2, 2026$1.25$4.251.05M
Gemini 3.8 FlashGoogleSep 2, 2026$0.825$4.1251.05M
Claude Fable 5.1AnthropicSep 1, 2026$10$501M
Granite 4.2 8BIBMAug 31, 2026$0.06$0.15131K
Hy4 previewTencentAug 28, 2026$0.834$2.5011.05M
Qwen3.8 FlashQwenAug 26, 2026$0.15$0.471M
GLM 5.3 FlashZ.aiAug 26, 2026$0.0399$0.07841.05M
Muse Spark 1.2 ContributorMetaAug 21, 2026$0.1$0.21.05M
DeepSeek V4 Flash Vision ExpDeepSeekAug 21, 2026$0.44$1.321.05M
Hy-MT2-1.8BTencentAug 20, 2026$0.044$0.1778K
Hy-MT2-30B-A3BTencentAug 20, 2026$0.074$0.2958K
Hy-MT2-7BTencentAug 19, 2026$0.074$0.2958K
GLM 5.3Z.aiAug 18, 2026$0.0699$1.41.05M
Qwen3.8 27BQwenAug 14, 2026$0.06$2.21.02M
Gemini 3.7 FlashGoogleAug 13, 2026$0.75$3.751.05M
Seed 2.1 TurboByteDance SeedAug 12, 2026$0.5$2.5262K

Showing the 100 newest of the 195 models in the snapshot.

Source: OpenRouter, as of Nov 20, 06:00 UTC.

Independent results

Benchmark boards

Measured by Epoch AI, LMArena, OpenRouter's benchmark runs, τ²-bench and Vectara, not by the vendors. Each board names its source and its date. Artificial Analysis numbers are not shown on this site's boards: its terms allow single quoted numbers, not tables.

Show all 21 boards, top 10 of each as the source lists them

Epoch AI · GPQA diamond

Epoch AI · GPQA diamond: top results
RankModel (as the source lists it)Accuracy
1gpt-6.2-astra_max96.8%
2gemini-4-hydrogen_xhigh96.8%
3gemini-4-carbon_high96.5%
4gemini-4-argon_high96.4%
5claude-fable-5-2-fallback_max96.0%
6gpt-6-astra_max95.8%
7claude-sonnet-5-5_max95.6%
8gemini-3.8-flash_high95.4%
8gpt-6.1-sol_max95.4%
10gemini-3.7-flash_high94.8%

Top 10 of the 30 we keep.

Source: Epoch AI (CC BY), as of Nov 20, 2026. Source

Epoch AI · FrontierMath-Tiers-1-3-v2-Private

Epoch AI · FrontierMath-Tiers-1-3-v2-Private: top results
RankModel (as the source lists it)Accuracy
1claude-fable-5-2_max95.1%
2gpt-6-astra_max93.7%
2gpt-6.1-sol_max93.7%
4gpt-6.2-astra_max92.6%
5claude-opus-5-5_max91.2%
6gpt-6.1-terra_max90.9%
7gemini-4-carbon_high90.5%
8claude-fable-5-1_max90.2%
9gpt-6-sol_max89.8%
10gpt-5.6-sol_max89.1%

Top 10 of the 30 we keep.

Source: Epoch AI (CC BY), as of Nov 20, 2026. Source

Epoch AI · FrontierMath-Tier-4-v2-Private

Epoch AI · FrontierMath-Tier-4-v2-Private: top results
RankModel (as the source lists it)Accuracy
1claude-fable-5-2_max100.0%
1gpt-6.1-sol_max100.0%
1gpt-6.2-astra_max100.0%
4gpt-6-astra_high97.6%
4gpt-6-astra_max97.6%
4gpt-6-astra_xhigh97.6%
7gpt-6-astra_medium97.6%
8gemini-4-carbon_high95.1%
9claude-opus-5-5_max95.0%
10gpt-6.1-terra_max92.7%

Top 10 of the 30 we keep.

Source: Epoch AI (CC BY), as of Nov 20, 2026. Source

Epoch AI · SimpleQA Verified

Epoch AI · SimpleQA Verified: top results
RankModel (as the source lists it)Accuracy
1claude-fable-5-2_max94.0%
2gpt-6.2-astra_max78.5%
3gemini-4-carbon_high77.7%
4gemini-4-argon_high76.0%
5gpt-6-astra_max75.6%
6gpt-6.1-sol_max73.9%
7gemini-3.1-pro-preview_high73.5%
8claude-opus-5-5_max72.2%
9gemini-4-hydrogen_xhigh70.9%
10claude-fable-5-1_max70.8%

Top 10 of the 30 we keep.

Source: Epoch AI (CC BY), as of Nov 20, 2026. Source

Epoch AI · OTIS Mock AIME 2024-2025

Epoch AI · OTIS Mock AIME 2024-2025: top results
RankModel (as the source lists it)Accuracy
1claude-fable-5-1_max100.0%
1claude-fable-5_high100.0%
1claude-opus-5-5_max100.0%
1claude-sonnet-5-5_max100.0%
1gemini-4-argon_high100.0%
1gemini-4-carbon_high100.0%
1gemini-4-hydrogen_xhigh100.0%
1gpt-5.5-pre-release_xhigh100.0%
1gpt-5.5-pro-pre-release_xhigh100.0%
1gpt-5.6-sol_max100.0%

Top 10 of the 30 we keep.

Source: Epoch AI (CC BY), as of Nov 20, 2026. Source

Epoch AI · MirrorCode

Epoch AI · MirrorCode: top results
RankModel (as the source lists it)Accuracy
1claude-sonnet-5-5_max97.8%
2claude-opus-5-5_max95.6%
3claude-fable-5-1_max93.3%
4claude-opus-5_max87.8%
5claude-opus-5-5_high78.9%
6claude-haiku-5-5_max78.9%
7gpt-6.1-terra_max76.7%
8claude-fable-5-1_high73.3%
9claude-fable-5_max70.0%
10gpt-6.2-astra_max67.8%

Top 10 of the 30 we keep.

Source: Epoch AI (CC BY), as of Nov 20, 2026. Source

Epoch AI · Chess Puzzles

Epoch AI · Chess Puzzles: top results
RankModel (as the source lists it)Accuracy
1claude-fable-5-2_max79.0%
2gemini-4-carbon_high78.0%
2gpt-6.2-astra_max78.0%
4gpt-6-astra_promax76.0%
5gpt-6-astra_max74.0%
6gpt-5.6-sol_promax68.0%
6gpt-6.1-sol_max68.0%
8claude-opus-5-5_max67.0%
9gemini-3.7-flash_high66.0%
10gpt-5.5_xhigh63.0%

Top 10 of the 30 we keep.

Source: Epoch AI (CC BY), as of Nov 20, 2026. Source

Epoch AI · Mystery Game Puzzles

Epoch AI · Mystery Game Puzzles: top results
RankModel (as the source lists it)Accuracy
1gpt-6.2-astra_max95.0%
2claude-fable-5-2_max90.0%
3gpt-6-astra_max84.0%
4gpt-6.1-sol_max80.0%
5gpt-6.1-terra_max74.0%
6gemini-4-carbon_high73.0%
7claude-opus-5-5_max71.0%
8claude-sonnet-5-5_max65.0%
8gemini-4-hydrogen_xhigh65.0%
10gemini-4-argon_high64.0%

Top 10 of the 30 we keep.

Source: Epoch AI (CC BY), as of Nov 20, 2026. Source

Epoch AI · Furniture Assembly

Epoch AI · Furniture Assembly: top results
RankModel (as the source lists it)Accuracy
1claude-fable-5-2_max90.0%
2gpt-6.2-astra_max89.2%
3claude-opus-5-5_max83.3%
4gpt-6-astra_max80.0%
4gpt-6.1-sol_max80.0%
6gpt-6.1-terra_max79.2%
7gemini-4-carbon_high77.5%
8claude-sonnet-5-5_max75.0%
9gemini-4-argon_high72.5%
10claude-fable-5-1_max70.0%

Top 10 of the 30 we keep.

Source: Epoch AI (CC BY), as of Nov 20, 2026. Source

Epoch AI · Epoch Capabilities Index

Epoch AI · Epoch Capabilities Index: top results
RankModel (as the source lists it)Index
1GPT-6.2 Astra174.17
2Claude Fable 5.2173.5
3Gemini 4 Carbon169.78
4GPT-6 Astra169.2
5Claude Opus 5.5168.7
6GPT-6.1 Sol168.63
7GPT-6.1 Terra167.72
8Claude Fable 5.1167.09
9Claude Sonnet 5.5166.59
10Gemini 4 Argon166.57

Top 10 of the 30 we keep.

Source: Epoch AI (CC BY), as of Nov 20, 2026. Source

LMArena · Text (style control)

LMArena · Text (style control): top results
RankModel (as the source lists it)Arena score
1Gemini 4 Carbon (High)1,526
2Gemini 4 Argon (High)1,521
3Gemini 4 Neon (High)1,518
4Muse Wonder (Max)1,511
5Claude Opus 5.5 (High)1,507
6Claude Fable 5 (High)1,504
7Claude Opus 4.6 (High)1,503
8Claude Fable 5.1 (Max)1,502
9Claude Opus 4.7 (High)1,501
10Muse Spark 1.3 (Max)1,498

Top 10 of the 30 we keep.

LMArena (CC BY 4.0), as of Nov 18, 2026. Source

LMArena · WebDev

LMArena · WebDev: top results
RankModel (as the source lists it)Arena score
1Claude Fable 5.2 (xHigh)1,824
2Claude Opus 5.5 (Max)1,818
3GPT-6.2 Astra (Max)1,781
4Claude Sonnet 5.5 (xHigh)1,773
5Gemini 4 Carbon (High)1,769
6Gemini 4 Neon (High)1,764
7GPT-6 Astra (Max)1,763
8GPT-6.2 Sol (Max)1,757
9Qwen4-Max (xHigh)1,750
10GPT-6.1 Sol (Max)1,744

Top 10 of the 30 we keep.

LMArena (CC BY 4.0), as of Nov 18, 2026. Source

LMArena · Vision

Top 10 of the 30 we keep.

LMArena (CC BY 4.0), as of Nov 9, 2026. Source

LMArena · Agent

Top 10 of the 30 we keep.

LMArena (CC BY 4.0), as of Nov 18, 2026. Source

OpenRouter · GPQA Diamond

OpenRouter · GPQA Diamond: top results
RankModel (as the source lists it)Accuracy
1Gemini 4 Argonvia Google, its own provider96.5%
2Gemini 3.8 Flashvia Google AI Studio, its own provider95.3%
3Gemini 3.1 Pro Previewvia Google, its own provider95.2%
4GPT-6 Astra Provia OpenAI, its own provider95.1%
5Claude Fable 5.2via Anthropic, its own provider95.1%
6Ling 3.1 Flashvia DeepInfra, the most accurate provider94.8%
7GPT-6 Astravia OpenAI, its own provider94.7%
8Gemini 3.7 Flashvia Google, its own provider94.5%
9GPT-6.1 Solvia OpenAI, its own provider94.4%
10GPT-5.6 Sol Provia OpenAI, its own provider93.9%

Top 10 of the 60 we keep.

Source: OpenRouter (openrouter.ai), licensed under CC BY 4.0, as of Nov 20, 2026. Source

OpenRouter · GPQA Diamond: cost per task

Lower is better on this board.

OpenRouter · GPQA Diamond: cost per task: top results
RankModel (as the source lists it)Cost per task
1GPT-6 Lunavia OpenAI, its own provider$0.003
2Claude Haiku 5.5via Anthropic, its own provider$0.004
3MiniMax M3.1 Flashvia Minimax, its own provider$0.0068
4DeepSeek V4 Flash 0423via Crusoe, the most accurate provider$0.008
5GPT-6 Luna Provia OpenAI, its own provider$0.0082
6Ling 3.1 Flashvia DeepInfra, the most accurate provider$0.0093
7GLM 5.3 Flashvia Z.AI, its own provider$0.01
8GPT-6.1 Terravia OpenAI, its own provider$0.011
9DeepSeek V4.1 Flash 1017via Tenstorrent, the most accurate provider$0.012
10Nemotron 3.5 Super VLvia Io Net, the most accurate provider$0.013

Top 10 of the 60 we keep.

Source: OpenRouter (openrouter.ai), licensed under CC BY 4.0, as of Nov 20, 2026. Source

OpenRouter · GPQA Diamond: time per task

Lower is better on this board.

OpenRouter · GPQA Diamond: time per task: top results
RankModel (as the source lists it)Time per task
1Claude Sonnet 5.5via Anthropic, its own provider20 s
2Claude Fable 5.2via Anthropic, its own provider26 s
3Claude Opus 5.5via Anthropic, its own provider27 s
4GPT-6.1 Terravia OpenAI, its own provider28 s
5GPT-6 Solvia OpenAI, its own provider34 s
6Claude Haiku 5.5via Anthropic, its own provider37 s
7Claude Fable 5.1via Anthropic, its own provider40 s
8Claude Fable 5via Anthropic, its own provider47 s
9GPT-6.1 Solvia OpenAI, its own provider48 s
10GPT-6 Astravia OpenAI, its own provider48 s

Top 10 of the 60 we keep.

Source: OpenRouter (openrouter.ai), licensed under CC BY 4.0, as of Nov 20, 2026. Source

OpenRouter · τ²-bench airline

OpenRouter · τ²-bench airline: top results
RankModel (as the source lists it)Accuracy
1DeepSeek V4 Flash 0423via OpenInference, the most accurate provider81.7%
2DeepSeek V4 Flash 0731via BaseTen, the most accurate provider79.8%
3Qwen3.5 397B A17Bvia Alibaba, its own provider79.4%
4Ling 3.1 Flashvia DeepInfra, the most accurate provider79.3%
5Claude Fable 5via Anthropic, its own provider79.3%
6Qwen3.5-35B-A3Bvia Alibaba, its own provider79.2%
7Claude Fable 5.1via Anthropic, its own provider79.1%
8MiMo-V2.6-Provia Xiaomi, its own provider78.9%
9Gemini 4 Argonvia Google, its own provider78.7%
10DeepSeek V4 Pro 0813via DeepInfra, the most accurate provider78.6%

Top 10 of the 59 we keep.

Source: OpenRouter (openrouter.ai), licensed under CC BY 4.0, as of Nov 20, 2026. Source

OpenRouter · τ²-bench airline: cost per task

Lower is better on this board.

OpenRouter · τ²-bench airline: cost per task: top results
RankModel (as the source lists it)Cost per task
1Ling 3.0 Flashvia Novita, the most accurate provider$0.0021
2DeepSeek V4 Flash 0731via BaseTen, the most accurate provider$0.0076
3MiMo-V2.5via Xiaomi, its own provider$0.0082
4GLM 5.3 Flashvia Z.AI, its own provider$0.0092
5MiMo-V2.6-Flashvia Xiaomi, its own provider$0.012
6Qwen3.8 Flashvia Alibaba, its own provider$0.015
7GLM 5.3via digitalocean, the most accurate provider$0.016
8DeepSeek V4.1 Flash 1017via AtlasCloud, the most accurate provider$0.017
9Beam 1.0 501B A23Bvia Io Net, the most accurate provider$0.025
10MiniMax M3via Minimax, its own provider$0.03

Top 10 of the 60 we keep.

Source: OpenRouter (openrouter.ai), licensed under CC BY 4.0, as of Nov 20, 2026. Source

OpenRouter · τ²-bench airline: time per task

Lower is better on this board.

OpenRouter · τ²-bench airline: time per task: top results
RankModel (as the source lists it)Time per task
1DeepSeek V4 Flash 0731via BaseTen, the most accurate provider31 s
2GLM 5.3via digitalocean, the most accurate provider42 s
3Muse Spark Openvia Modal, the most accurate provider66 s
4DeepSeek V4 Pro 0423via BaseTen, the most accurate provider70 s
5Claude Sonnet 5.5via Anthropic, its own provider73 s
6Ling 3.0 Flashvia Novita, the most accurate provider76 s
7Claude Opus 4.7via Anthropic, its own provider86 s
8DeepSeek V4.1 Flash 1017via AtlasCloud, the most accurate provider93 s
9Claude Sonnet 5via Anthropic, its own provider95 s
10Nova Micro 1.0via Anthropic, the most accurate provider96 s

Top 10 of the 60 we keep.

Source: OpenRouter (openrouter.ai), licensed under CC BY 4.0, as of Nov 20, 2026. Source

Vectara · Hallucination rate in summaries

Lower is better on this board.

Vectara · Hallucination rate in summaries: top results
RankModel (as the source lists it)Percent
1antgroup/finix_s1_32b1.8%
2openai/gpt-5.4-nano-2026-03-173.1%
3google/gemini-2.5-flash-lite3.3%
4microsoft/Phi-43.7%
5meta-llama/Llama-3.3-70B-Instruct-Turbo4.1%
6deepseek-ai/DeepSeek-V4.1-Flash4.2%
7snowflake/snowflake-arctic-instruct4.3%
8google/gemma-3-12b-it4.4%
9mistralai/mistral-large-24114.5%
10qwen/qwen3-8b4.8%

Top 10 of the 73 we keep.

Vectara Hallucination Leaderboard (HHEM-2.3), Apache-2.0, as of Oct 22, 2026. Source

From the Drops

Models

Google ships Gemini 4 Neon at $3 in, $15 out per 1M tokens, pitched on cost per task

Google DeepMind says Neon has a 1M-token context, is rolling out in the Gemini API, Antigravity and the Gemini app, and costs 32% less per task than Carbon at its default level. Artificial Analysis scores it 61.4.

Why it matters: A mid-priced model with frontier-level vendor claims is a reason to rerun your own evals; Google's benchmark and cost figures are its own, so check them on your tasks first.

From Drop #095 · Thu, Nov 19, 2026 · Morning
Models

Microsoft AI lists seven models, including Cyber 2 Flash, and offers licensed weights

Microsoft AI says its new Cyber 2 Flash model outscores GPT-6 Astra on CyberGym-E2E for under 2% of the per-task cost, and its Licensed Weights let customers tune a MAI model and run it in their own Azure environment.

Why it matters: These are Microsoft's own comparisons, not independent tests. If you run security or coding agents around the clock, price a MAI pilot against your current model on your own tasks.

From Drop #094 · Wed, Nov 18, 2026 · Evening
Models

OpenAI ships GPT-6.2 Astra and Sol; Astra cuts price to $8 in, $40 out per 1M

OpenAI released GPT-6.2 Astra and GPT-6.2 Sol in its API. Simon Willison notes Astra costs 80% of 6 Astra's price and cached input drops to $0.40 per 1M from $1; Sol stays at $2 in, $10 out.

Why it matters: Long agent sessions lean on cached input, so rerun your cost model on cached share. Willison reports mixed early reviews, so test it on your own tasks before switching.

From Drop #093 · Wed, Nov 18, 2026 · Morning
Models

Microsoft's MAI Code 2 and Code 2 Flash land in GitHub Copilot, pitched on price

Microsoft says MAI Code 2 gives 98% of GPT 6.1 Sol's performance at 40% of the cost across four coding benchmarks, and Code 2 Flash targets routine edits. It says neither is tuned yet for very long agentic work.

Why it matters: Microsoft's comparison prices MAI Code 2 at a 50% promotional rate on its own harness, so check costs at list price on your repos before routing everyday tasks to it.

From Drop #093 · Wed, Nov 18, 2026 · Morning
Models

K2-Horizon releases six fully open models, from 0.9B to a 375B-A23B MoE

The K2 Team's technical report on arXiv says it releases checkpoints, pre-training data, logs and training code, and that the models lead or compete with open models of their size on agentic and coding benchmarks.

Why it matters: A fully open stack lets teams audit the training data behind an agent model. Check the benchmark claims on your own tasks before swapping models.

From Drop #092 · Tue, Nov 17, 2026 · Evening
Models

Tencent's Hy4 Flash lands on OpenRouter at $0.035 in, $0.1405 out per 1M tokens

OpenRouter lists Tencent's Hy4 Flash, added Nov. 16, with a 1048576-token context window. No independent benchmark scores for it were in our data this run.

Why it matters: A very low-priced, long-context option for bulk work. Test it on your own tasks before routing traffic; token price alone says little about cost per finished task.

From Drop #091 · Tue, Nov 17, 2026 · Morning
Models

Alibaba's Qwen4-Max lands on OpenRouter at $1.5 in, $4.5 out per 1M tokens

OpenRouter lists Qwen4-Max at $1.5 per 1M input tokens and $4.5 per 1M output. Qwen posted on X that it is a new MoE model with native vision and a 1M-token context; an HN thread drew 163 points.

Why it matters: A new low-priced large model worth a trial on your own workload; so far the performance claims come from X posts, not a published model card, so measure before switching.

From Drop #089 · Mon, Nov 16, 2026 · Morning
Models

Epoch Capabilities Index: Claude Fable 5.2 leads at 173.44, Gemini 4 Carbon next

On Epoch AI's capabilities index, Claude Fable 5.2 scores 173.44, ahead of Gemini 4 Carbon at 169.76, GPT-6 Astra at 169, Claude Opus 5.5 at 168.54 and GPT-6.1 Sol at 168.48.

Why it matters: The top five sit close together, so price and fit on your own tasks may matter more than rank. Test two or three before standardizing.

From Drop #088 · Sun, Nov 15, 2026 · Evening
Models

LMArena agent board: Claude Fable 5.2 leads at 0.304, ahead of Opus 5.5 at 0.195

On LMArena's agent leaderboard as of 2026-11-11, Claude Fable 5.2 (Max) scores 0.304, Claude Opus 5.5 (Max) 0.195, Gemini 4 Carbon (High) 0.17 and Claude Sonnet 5.5 (Max) 0.16. Source: LMArena (CC BY 4.0).

Why it matters: Agent and chat rankings differ: Gemini 4 Carbon tops LMArena's text board but is third here. Pick agent models on agent results and your own tasks.

From Drop #087 · Sun, Nov 15, 2026 · Morning
Models

Unsloth updates Qwen3.8-27B quants, says 5-10% better accuracy at the same size

Unsloth says its Dynamic V3.1 GGUFs for Qwen3.8-27B, a 27B model with vision and a 256K context window, raise average accuracy 5-10% at comparable sizes on tests including Terminal Bench and DeepSWE.

Why it matters: Unsloth says the model runs locally on 17GB setups, which makes it a test case for keeping coding agents on your own hardware. Rerun your own evals before swapping quants.

From Drop #086 · Sat, Nov 14, 2026 · Evening
Models

Perplexity releases Decider v1.3 decision models under an Apache license

Perplexity says Decider v1.3 and Decider Flash v1.3 return probabilities for yes/no, multiple-choice and score questions in one pass, and add locate for pulling out text spans. Both are on Hugging Face, OpenRouter and its API.

Why it matters: Jobs like routing, judging and screening tool output for prompt injection may not need a generative model. Price one of these against the LLM calls you use today.

From Drop #085 · Sat, Nov 14, 2026 · Morning
Models

llama.cpp reaches v1.0.0 and ties any API break to a major-version bump

The llama.cpp team at Hugging Face released the llama.cpp and ggml libraries as v1.0.0 and says API or ABI breaks will now come with major-version bumps. It supports more than 150 architecture variants, the team says.

Why it matters: Apps built on local inference can pin a major version and expect stable interfaces within it. Check how your builds pull llama.cpp before moving across majors.

From Drop #084 · Fri, Nov 13, 2026 · Evening

Showing the newest 12 of 154. Older ones stay on their Drop's page in the archive.

Not tracked yet

Speed, badges, trending and notes

Speed (tokens per second), open-weights vs commercial badges, Hugging Face trending and a "worth a look if…" note per release are not tracked yet.