Cookies on AIGROW

We use necessary cookies to run the site. With your permission, we also use cookies for analytics, session recordings and marketing attribution.

Cookie Policy

AI leaderboard · Reasoning

AI reasoning benchmark: 28 models ranked on measured results

Three tests Epoch AI runs itself: the best move in new chess positions, the best move in a game whose name is kept secret so models are less likely to have prepared for it, and how many goals a model reaches in a little-known board game after practising it.

Run Claude Fable 5.2 up to 60% below list price. $25 a month brings $110 of API credit and cancels any time; $25 once buys $100. Results from tests Epoch AI runs, data of 12 November 2026, under CC BY 4.0. Prices from each maker’s pricing page, read on 12 November 2026.

Top score
86.5%Claude Fable 5.2 from Anthropic
Models ranked
28
Level with the leader
2
On sale with AIGROW credit
9
Data of
12 November 2026

Level: its interval overlaps the leader’s.

On the three reasoning tests Epoch AI runs itself, Claude Fable 5.2 from Anthropic leads the 28 models with results on all three, averaging 86.5% (data of 12 November 2026). 2 other models are statistically level with it. AIGROW API credit sells it at $3.20 / $16.00 instead of $8.00 / $40.00 per million input / output tokens.

Run models from this ranking below list price

AIGROW API credit runs these models through one key, billed per token at the rates below. $25 a month brings $110 of API credit and cancels any time; $25 once buys $100.

  1. No. 1Claude Fable 5.2AnthropicList price $8.00 / $40.00$3.20 / $16.00with AIGROW API creditRun Claude Fable 5.2 up to 60% below list priceCertified genuine by AIGROW on 11 November 2026
  2. No. 3GPT-6 AstraOpenAIList price $10.00 / $50.00$2.50 / $12.50with AIGROW API creditRun GPT-6 Astra up to 75% below list priceCertified genuine by AIGROW on 11 November 2026
  3. No. 4Claude Opus 5.5AnthropicList price $4.00 / $20.00$1.60 / $8.00with AIGROW API creditRun Claude Opus 5.5 up to 60% below list priceCertified genuine by AIGROW on 11 November 2026

US dollars per million input / output tokens. Certified genuine by AIGROW means the model passed AIGROW’s own check against the same model at its maker on 11 November 2026. No model maker reviewed or endorsed it.

AI reasoning benchmark

Epoch uses puzzles and games to stand in for planning and learning on the job; they do not test writing, coding or knowledge.

Results of 28 AI models on Chess Puzzles, Mystery Game Puzzles, EBR-bench, with their 95% intervals and list prices
#ModelAverageChess PuzzlesMystery Game PuzzlesEBR-benchPer 1M tokens, in / out
1Claude Fable 5.2AnthropicRun Claude Fable 5.2 up to 60% below list priceCertified genuine by AIGROW on 11 November 202686.5%95% interval 83.0% to 90.0%79.0%max, run 6 Nov 202690.0%max, run 6 Nov 202690.5%max, run 6 Nov 2026List price $8.00 / $40.00$3.20 / $16.00with AIGROW API credit
2Gemini 4 CarbonGoogle DeepMind79.5%95% interval 73.6% to 85.5%Its interval overlaps 1 model above it78.0%high, run 2 Nov 202673.0%high, run 2 Nov 202687.6%high, run 5 Nov 2026$2.00 / $10.00
3GPT-6 AstraOpenAIRun GPT-6 Astra up to 75% below list priceCertified genuine by AIGROW on 11 November 202678.7%95% interval 72.1% to 85.3%Its interval overlaps 2 models above it76.0%promax, run 15 Oct 202684.0%max, run 30 Aug 202676.2%max, run 30 Aug 2026List price $10.00 / $50.00$2.50 / $12.50with AIGROW API credit
4Claude Opus 5.5AnthropicRun Claude Opus 5.5 up to 60% below list priceCertified genuine by AIGROW on 11 November 202669.8%95% interval 63.5% to 76.1%Its interval overlaps 2 models above it67.0%max, run 13 Oct 202671.0%max, run 29 Sept 202671.4%max, run 22 Sept 2026List price $4.00 / $20.00$1.60 / $8.00with AIGROW API credit
5GPT-6.1 SolOpenAIRun GPT-6.1 Sol up to 75% below list priceCertified genuine by AIGROW on 11 November 202667.4%95% interval 61.9% to 72.9%Its interval overlaps 2 models above it68.0%max, run 13 Oct 202680.0%max, run 29 Sept 202654.3%max, run 29 Sept 2026List price $2.00 / $10.00$0.50 / $2.50with AIGROW API credit
6Gemini 4 ArgonGoogle DeepMindRun Gemini 4 Argon up to 50% below list priceCertified genuine by AIGROW on 11 November 202662.6%95% interval 57.1% to 68.1%Its interval overlaps 2 models above it59.0%high, run 13 Oct 202664.0%high, run 13 Oct 202664.8%high, run 13 Oct 2026List price $2.00 / $10.00$1.00 / $5.00with AIGROW API credit
7GPT-5.6 SolOpenAI56.9%95% interval 51.7% to 62.1%Its interval overlaps 2 models above it68.0%promax, run 14 Oct 202658.0%max, run 28 Jul 202644.8%max, run 4 Sept 2026$4.00 / $20.00
8GPT-6 SolOpenAI55.1%95% interval 49.6% to 60.6%Its interval overlaps 2 models above it56.0%max, run 13 Oct 202656.0%max, run 22 Sept 202653.3%max, run 22 Sept 2026$2.00 / $10.00
9Claude Fable 5.1AnthropicRun Claude Fable 5.1 up to 60% below list priceCertified genuine by AIGROW on 11 November 202654.4%95% interval 48.3% to 60.4%Its interval overlaps 3 models above it48.0%max, run 13 Oct 202658.0%max, run 1 Sept 202657.1%max, run 8 Sept 2026List price $10.00 / $50.00$4.00 / $20.00with AIGROW API credit
10GPT-5.5OpenAI51.1%95% interval 46.4% to 55.8%Its interval overlaps 3 models above it63.0%xhigh, run 13 Oct 202656.0%xhigh, run 24 Jul 202634.3%xhigh, run 27 Jul 2026$5.00 / $30.00
11Claude Opus 5Anthropic49.2%95% interval 44.1% to 54.4%Its interval overlaps 4 models above it43.0%max, run 13 Oct 202659.0%max, run 25 Jul 202645.7%max, run 5 Sept 2026$5.00 / $25.00
12Claude Fable 5Anthropic46.2%95% interval 41.4% to 51.0%Its interval overlaps 4 models above it47.0%max, run 13 Oct 202652.0%max, run 31 Jul 202639.5%max, run 17 Jul 2026No list price
13Claude Haiku 5.5Anthropic42.9%95% interval 38.3% to 47.5%Its interval overlaps 3 models above it55.0%max, run 13 Oct 202630.0%max, run 7 Oct 202643.8%max, run 9 Oct 2026$0.10 / $0.50
14Claude Opus 4.8Anthropic38.9%95% interval 34.2% to 43.5%Its interval overlaps 2 models above it52.0%max, run 13 Oct 202636.0%max, run 25 Jul 202628.6%max, run 7 Aug 2026No list price
15GPT-5.4OpenAI36.8%No published margin48.0%xhigh, run 13 Oct 202637.0%xhigh, run 24 Jul 202625.4%xhigh, run 25 Jun 2026Not updated in over 120 daysNo list price
16Gemini 3.1 Pro (preview)Google DeepMind33.1%No published margin51.0%high, run 13 Oct 202634.0%high, run 27 Jul 202614.3%Run 25 Jun 2026Not updated in over 120 days$2.00 / $12.00
17Grok 4.6xAI32.5%95% interval 28.0% to 37.0%Its interval overlaps 1 model above it33.0%xhigh, run 13 Oct 202634.0%xhigh, run 14 Aug 202630.5%xhigh, run 18 Aug 2026No list price
18GPT-5.2OpenAI31.3%No published margin48.0%xhigh, run 14 Oct 202623.0%high, run 6 Aug 202623.0%xhigh, run 26 Jun 2026Not updated in over 120 daysNo list price
19Claude Opus 4.7Anthropic29.3%No published margin41.0%max, run 13 Oct 202628.0%max, run 25 Jul 202619.1%max, run 30 Jun 2026Not updated in over 120 daysNo list price
20Gemini 3.5 FlashGoogle DeepMind27.3%No published margin45.0%high, run 13 Oct 202632.0%high, run 27 Jul 20264.8%high, run 25 Jun 2026Not updated in over 120 daysNo list price
21GPT-5OpenAI23.9%No published margin36.0%high, run 13 Oct 202623.0%high, run 5 Aug 202612.7%high, run 29 Jun 2026Not updated in over 120 daysNo list price
22Qwen3.7-MaxAlibaba22.8%No published margin27.0%Run 13 Oct 202632.0%Run 28 Jul 20269.5%Run 25 Jun 2026Not updated in over 120 daysNo list price
23Mistral Large 4Mistral AI22.4%95% interval 17.9% to 26.8%29.0%high, run 2 Nov 202620.0%high, run 2 Nov 202618.1%high, run 2 Nov 2026$0.70 / $2.10
24Claude Opus 4.6Anthropic19.6%No published margin21.0%max, run 13 Oct 202625.0%max, run 25 Jul 202612.7%max, run 29 Jun 2026Not updated in over 120 daysNo list price
25GLM-5.2Z.ai (Zhipu AI)16.8%No published margin22.0%max, run 14 Oct 202619.0%low, run 27 Aug 20269.5%max, run 29 Jun 2026Not updated in over 120 daysNo list price

Not ranked: models missing a test

These models have results on some of the tests but not all, so an average would not compare with the ones above. Their results are here, best average first, without a number.

82 models with results on some of the tests
ModelChess PuzzlesMystery Game PuzzlesEBR-bench
GPT-6.1 TerraOpenAIRun GPT-6.1 Terra up to 75% below list priceCertified genuine by AIGROW on 11 November 2026Not run74.0%max, run 26 Oct 202678.1%max, run 26 Oct 2026
Claude Sonnet 5.5AnthropicRun Claude Sonnet 5.5 up to 60% below list priceCertified genuine by AIGROW on 11 November 202659.0%max, run 13 Oct 202665.0%max, run 29 Sept 2026Not run
GPT-5.5 ProOpenAI62.0%xhigh, run 14 Oct 2026Not runNot run
GPT-5.2 ProOpenAI60.0%xhigh, run 14 Oct 2026Not runNot run
GPT-5.4 ProOpenAI59.0%xhigh, run 14 Oct 2026Not runNot run
Gemini 3.8 FlashGoogle DeepMind58.0%high, run 13 Oct 202647.0%high, run 2 Sept 2026Not run
Gemini 3.7 FlashGoogle DeepMind66.0%high, run 13 Oct 202637.0%high, run 14 Aug 2026Not run
GPT-5 ProOpenAI48.0%high, run 14 Oct 2026Not runNot run
DeepSeek V4.1 Flash 1017DeepSeek46.0%max, run 19 Oct 202649.0%max, run 19 Oct 2026Not run
DeepSeek V4.1 FlashDeepSeek48.0%max, run 13 Oct 202643.0%max, run 8 Oct 2026Not run

How to read this ranking

What the scores measure, what they leave out, and where each number comes from.

What it measures
Three tests Epoch AI runs itself: the best move in new chess positions, the best move in a game whose name is kept secret so models are less likely to have prepared for it, and how many goals a model reaches in a little-known board game after practising it.
What it leaves out
Epoch uses puzzles and games to stand in for planning and learning on the job; they do not test writing, coding or knowledge.
How to read the margin
Each result carries a 95% interval, 1.96 standard errors either side as Epoch AI publishes them; an average combines the errors of its tests. Where two intervals overlap, the tests cannot say which model is better. Rows are still numbered by score, and a row says how many models above it overlap.
Settings
Under each result is the reasoning setting Epoch AI ran the model at, as it names it, and the day of the run. Where a model ran at several settings, the best result counts.
Complete results only
A model is numbered only when it has results on all three tests, and its place is the plain average of the three, with no weights. Models with fewer results are listed below the ranking without a number.
Not updated in over 120 days
A result run more than 120 days before the newest data on these pages carries this mark, so an older result can be weighed as one.
Prices
The list price per million tokens from the maker’s own pricing page, read on 12 November 2026. It is a price per token, not per answer: a higher reasoning setting spends more tokens on the same question. Where AIGROW API credit sells a model, the list price is struck through and the price through the credit sits under it.

Source: Epoch AI, Chess Puzzles, Mystery Game Puzzles, EBR-bench (data of 12 November 2026), licensed under CC BY 4.0. We change no score: we add the prices and the links to our model pages.

What each of them costs to run

The price list works three everyday jobs on every model, and our own benchmark runs them on real business tasks to see which pass.

See the pricesSee the benchmark

Data of 12 November 2026