Skip to content
ITG NowFollow

Home › AI model benchmarks

Which AI model is best for what?

13 AI models on 8 tests, ranked from published scores. Every number names its source.

Quick answer

No single AI model wins every test. Leaders right now: all-round, Claude Opus 5.5; chat and writing, Gemini 4 Argon; build websites and apps, Claude Opus 5.5; coding, Claude Fable 5.1; reasoning, Gemini 4 Argon; science questions, GPT-6 Astra; hardest exam, Claude Opus 5.5; speed, Gemini 3.5 Flash-Lite.

Who is best at what

Each dot is one model's rank on one test.

Rank of each model on each test
Model most top-3 finishes firstAll-roundChat & writingBuild websites & appsCodingReasoningScience questionsmaker's ownHardest exammaker's ownSpeed
Claude Opus 5.5rank 1 of 12rank 2 of 12rank 1 of 13rank 2 of 10rank 3 of 10rank 2 of 6rank 1 of 6rank 6 of 11
Claude Fable 5.1rank 2 of 12rank 3 of 12rank 3 of 13rank 1 of 10rank 4 of 10rank 2 of 6rank 8 of 11
GPT-6 Astrarank 2 of 12rank 4 of 12rank 2 of 13rank 3 of 10rank 2 of 10rank 1 of 6rank 9 of 11
Gemini 4 Argonrank 2 of 12rank 1 of 12rank 4 of 13rank 4 of 10rank 1 of 10
Claude Haiku 5.5rank 6 of 12rank 8 of 13rank 7 of 10rank 6 of 10rank 5 of 6rank 3 of 6rank 2 of 11
DeepSeek-V4.1-Flashrank 7 of 12rank 5 of 12rank 6 of 13rank 6 of 10rank 5 of 10rank 3 of 6rank 4 of 6rank 3 of 11
Gemini 3.5 Flash-Literank 10 of 12rank 6 of 12rank 10 of 13rank 1 of 11
Gemma 4 31Brank 11 of 12rank 7 of 12rank 11 of 13rank 10 of 10rank 10 of 10rank 6 of 6rank 6 of 6rank 11 of 11
GPT-6 Lunarank 8 of 12rank 9 of 12rank 9 of 13rank 7 of 10rank 8 of 10rank 5 of 11
Grok 4.3rank 10 of 12rank 12 of 13
Grok 4.7rank 5 of 12rank 8 of 12rank 5 of 13rank 7 of 10rank 7 of 10rank 7 of 11
Mistral Medium 3.5rank 12 of 12rank 12 of 12rank 13 of 13rank 4 of 11
Qwen3.8-27Brank 9 of 12rank 11 of 12rank 7 of 13rank 5 of 10rank 9 of 10rank 4 of 6rank 5 of 6rank 10 of 11

winner bigger and brighter = higher rank no published scoreHover or tap a dot for the number. Tap a column name for the full ranking.

Winners by job

Tap a job for the full ranking.

Scoreboard

Every model on every test with 3 or more models. Greener = better rank.

Scores by model and test
ModelArtificial Analysis Intelligence Indexindex pointsLMArena TextArena scoreLMArena WebDevArena scoreLiveBench 2.0 Codingscore out of 100LiveBench 2.0 Reasoningscore out of 100GPQA Diamond%Humanity's Last Exam (no tools)%Output speedtokens per second
Claude Opus 5.558 (best)15081807 (best)96.392.29164.4 (best)97
Claude Fable 5.1531502174197.7 (best)91.7–60.971
GPT-6 Astra531477177793.792.796.0 (best)–57
Gemini 4 Argon531524 (best)167889.795.1 (best)–––
Claude Haiku 5.543–158584.584.68545.9235
DeepSeek-V4.1-Flash391475161885.486.790.936.8176
Gemini 3.5 Flash-Lite2214551441––––449 (best)
Gemma 4 31B151452136749.068.384.319.535
GPT-6 Luna381443157884.581.8––120
Grok 4.3–14421358–––––
Grok 4.7461444163984.582.7––93
Mistral Medium 3.51414261264––––160
Qwen3.8-27B341438159386.880.089.230.841
How to read the scoreboard
  • Stronger green = better rank on that test. Bold underlined = top score. – = no published score.
  • Rows are sorted by top-3 finishes. Tap a test name for its full ranking and source, or a model for everything we know about it.
  • Each column is its own test with its own scale, so do not compare numbers across columns.
  • Hover a cell to see who measured it (independent or the maker's own number), the date and the setting used.

Value picks

Price, memory, open models and the newest release, straight from the file.

Every test, ranked

Open a test for every score, what it measures and where the numbers come from.

All-roundArtificial Analysis Intelligence IndexClaude Opus 5.558 index points

One combined score from many independent tests of how smart a model is.

Higher is better · 12 models · independent

  1. 1Claude Opus 5.558 index points
  2. 2GPT-6 Astra53 index points
  3. 2Claude Fable 5.153 index points
  4. 2Gemini 4 Argon53 index points
  5. 5Grok 4.746 index points
  6. 6Claude Haiku 5.543 index points
  7. 7DeepSeek-V4.1-Flash39 index points
  8. 8GPT-6 Luna38 index points
  9. 9Qwen3.8-27B34 index points
  10. 10Gemini 3.5 Flash-Lite22 index points
  11. 11Gemma 4 31B15 index points
  12. 12Mistral Medium 3.514 index points

Source: Artificial Analysis, as of 2026-10-17 · independent test

Each model's highest-scoring setting is used. The page shows no index version or date; as_of is the day it was read.

Chat & writingLMArena TextGemini 4 Argon1524 Arena score

When real people compare two AI answers side by side without knowing which AI wrote them, which one do they pick more often?

Higher is better · 12 models · independent

  1. 1Gemini 4 Argon1524 Arena score
  2. 2Claude Opus 5.51508 Arena score
  3. 3Claude Fable 5.11502 Arena score
  4. 4GPT-6 Astra1477 Arena score
  5. 5DeepSeek-V4.1-Flash1475 Arena score
  6. 6Gemini 3.5 Flash-Lite1455 Arena score
  7. 7Gemma 4 31B1452 Arena score
  8. 8Grok 4.71444 Arena score
  9. 9GPT-6 Luna1443 Arena score
  10. 10Grok 4.31442 Arena score
  11. 11Qwen3.8-27B1438 Arena score
  12. 12Mistral Medium 3.51426 Arena score

Source: LMArena (arena.ai), as of 2026-10-15 · independent test

Default view with style control. Each score has a plus or minus range, so models within about 10 points are close to tied. Each model's best listed setting is used.

Build websites & appsLMArena WebDevClaude Opus 5.51807 Arena score

When people ask two AIs to build the same website or app and compare the results, which one do they pick?

Higher is better · 13 models · independent

  1. 1Claude Opus 5.51807 Arena score
  2. 2GPT-6 Astra1777 Arena score
  3. 3Claude Fable 5.11741 Arena score
  4. 4Gemini 4 Argon1678 Arena score
  5. 5Grok 4.71639 Arena score
  6. 6DeepSeek-V4.1-Flash1618 Arena score
  7. 7Qwen3.8-27B1593 Arena score
  8. 8Claude Haiku 5.51585 Arena score
  9. 9GPT-6 Luna1578 Arena score
  10. 10Gemini 3.5 Flash-Lite1441 Arena score
  11. 11Gemma 4 31B1367 Arena score
  12. 12Grok 4.31358 Arena score
  13. 13Mistral Medium 3.51264 Arena score

Source: LMArena (arena.ai), as of 2026-10-17 · independent test

Default view with style control. Scores within each model's plus or minus range are close to tied. Each model's best listed setting is used.

CodingLiveBench 2.0 CodingClaude Fable 5.197.7 score out of 100

Fresh coding tasks with tests that check the answer, changed often so models cannot memorize them.

Higher is better · 10 models · independent

  1. 1Claude Fable 5.197.7 score out of 100
  2. 2Claude Opus 5.596.3 score out of 100
  3. 3GPT-6 Astra93.7 score out of 100
  4. 4Gemini 4 Argon89.7 score out of 100
  5. 5Qwen3.8-27B86.8 score out of 100
  6. 6DeepSeek-V4.1-Flash85.4 score out of 100
  7. 7GPT-6 Luna84.5 score out of 100
  8. 7Claude Haiku 5.584.5 score out of 100
  9. 7Grok 4.784.5 score out of 100
  10. 10Gemma 4 31B49.0 score out of 100

Source: LiveBench 2.0, as of 2026-10-14 · independent test

LiveBench 2.0 release of 2 Sep 2026, shown on the site on 14 Oct 2026. Models run at their highest supported effort.

ReasoningLiveBench 2.0 ReasoningGemini 4 Argon95.1 score out of 100

Fresh logic and puzzle questions with one right answer, changed often so models cannot memorize them.

Higher is better · 10 models · independent

  1. 1Gemini 4 Argon95.1 score out of 100
  2. 2GPT-6 Astra92.7 score out of 100
  3. 3Claude Opus 5.592.2 score out of 100
  4. 4Claude Fable 5.191.7 score out of 100
  5. 5DeepSeek-V4.1-Flash86.7 score out of 100
  6. 6Claude Haiku 5.584.6 score out of 100
  7. 7Grok 4.782.7 score out of 100
  8. 8GPT-6 Luna81.8 score out of 100
  9. 9Qwen3.8-27B80.0 score out of 100
  10. 10Gemma 4 31B68.3 score out of 100

Source: LiveBench 2.0, as of 2026-10-14 · independent test

LiveBench 2.0 release of 2 Sep 2026, shown on the site on 14 Oct 2026. Models run at their highest supported effort.

Science questionsGPQA DiamondGPT-6 Astra96.0%

Very hard science questions in biology, chemistry and physics, written by PhD experts.

Higher is better · 6 models · maker's own number

  1. 1GPT-6 Astra96.0%
  2. 2Claude Opus 5.591%
  3. 3DeepSeek-V4.1-Flash90.9%
  4. 4Qwen3.8-27B89.2%
  5. 5Claude Haiku 5.585%
  6. 6Gemma 4 31B84.3%

Source: Makers' own model cards, as of 2026-10-17 · numbers can include makers' own claims

Maker's own numbers, run with different settings, so compare with care. Close to saturated at the top and questioned for possible training leaks. Anthropic's figures are whole numbers from its cost guide.

Hardest examHumanity's Last Exam (no tools)Claude Opus 5.564.4%

2,500 expert questions written to be too hard for AI, answered without web search or code.

Higher is better · 6 models · maker's own number

  1. 1Claude Opus 5.564.4%
  2. 2Claude Fable 5.160.9%
  3. 3Claude Haiku 5.545.9%
  4. 4DeepSeek-V4.1-Flash36.8%
  5. 5Qwen3.8-27B30.8%
  6. 6Gemma 4 31B19.5%

Source: Makers' own model cards, as of 2026-10-17 · numbers can include makers' own claims

The official leaderboard lists none of these models yet, so these are the makers' own numbers, graded by different judges. Some answers in the test are disputed.

SpeedOutput speedGemini 3.5 Flash-Lite449 tokens per second

How many tokens (pieces of words) per second the model writes once it starts answering. Higher is faster.

Higher is better · 11 models · independent

  1. 1Gemini 3.5 Flash-Lite449 tokens per second
  2. 2Claude Haiku 5.5235 tokens per second
  3. 3DeepSeek-V4.1-Flash176 tokens per second
  4. 4Mistral Medium 3.5160 tokens per second
  5. 5GPT-6 Luna120 tokens per second
  6. 6Claude Opus 5.597 tokens per second
  7. 7Grok 4.793 tokens per second
  8. 8Claude Fable 5.171 tokens per second
  9. 9GPT-6 Astra57 tokens per second
  10. 10Qwen3.8-27B41 tokens per second
  11. 11Gemma 4 31B35 tokens per second

Source: Artificial Analysis, as of 2026-10-17 · independent test

Median speed for the same setting used in the Intelligence Index. Gemini 4 Argon has no speed listed yet.

Every model, in detail

Price, memory, release date and every score with its source.

GPT-6 Astra OpenAI

Near the top for coding and hard reasoning; slower to start on its deepest settings.

Company
OpenAI
Released
2026-09-03
Open weights
No
Context
1,050,000 tokens
Price per million tokens
$10 in / $50 out · price source

All its scores

  • LMArena Text: 1477 Arena score (Max) independentas of 2026-10-15 · source
  • LMArena WebDev: 1777 Arena score (Max) independentas of 2026-10-17 · source
  • Artificial Analysis Intelligence Index: 53 index points (max) independentas of 2026-10-17 · source
  • Output speed: 57 tokens per second (max) independentas of 2026-10-17 · source
  • LiveBench 2.0 Reasoning: 92.7 score out of 100 (highest supported effort) independentas of 2026-10-14 · source
  • LiveBench 2.0 Coding: 93.7 score out of 100 (highest supported effort) independentas of 2026-10-14 · source
  • GPQA Diamond: 96.0% maker's own numberas of 2026-09-03 · source

Official page for GPT-6 Astra

GPT-6 Luna OpenAI

Very cheap and quick for simple everyday jobs.

Company
OpenAI
Released
2026-09-22
Open weights
No
Context
1,050,000 tokens
Price per million tokens
$0.1 in / $0.5 out · price source

All its scores

  • LMArena Text: 1443 Arena score (Max) independentas of 2026-10-15 · source
  • LMArena WebDev: 1578 Arena score (Max) independentas of 2026-10-17 · source
  • Artificial Analysis Intelligence Index: 38 index points (max) independentas of 2026-10-17 · source
  • Output speed: 120 tokens per second (max) independentas of 2026-10-17 · source
  • LiveBench 2.0 Reasoning: 81.8 score out of 100 (highest supported effort) independentas of 2026-10-14 · source
  • LiveBench 2.0 Coding: 84.5 score out of 100 (highest supported effort) independentas of 2026-10-14 · source

Official page for GPT-6 Luna

Claude Opus 5.5 Anthropic

Top pick for building apps and websites, and the highest overall intelligence score here.

Company
Anthropic
Released
2026-09-22
Open weights
No
Price per million tokens
$4 in / $20 out · price source

All its scores

  • LMArena Text: 1508 Arena score (High) independentas of 2026-10-15 · source
  • LMArena WebDev: 1807 Arena score (Max) independentas of 2026-10-17 · source
  • Artificial Analysis Intelligence Index: 58 index points (max with fallback) independentas of 2026-10-17 · source
  • Output speed: 97 tokens per second (max with fallback) independentas of 2026-10-17 · source
  • LiveBench 2.0 Reasoning: 92.2 score out of 100 (highest supported effort) independentas of 2026-10-14 · source
  • LiveBench 2.0 Coding: 96.3 score out of 100 (highest supported effort) independentas of 2026-10-14 · source
  • GPQA Diamond: 91% maker's own numberas of 2026-10-17 · source
  • Humanity's Last Exam (no tools): 64.4% maker's own numberas of 2026-09-22 · source

Official page for Claude Opus 5.5

Claude Fable 5.1 Anthropic

Strong all-rounder for writing and coding, but one of the priciest.

Company
Anthropic
Released
2026-09-01
Open weights
No
Context
1,000,000 tokens
Price per million tokens
$10 in / $50 out · price source

All its scores

  • LMArena Text: 1502 Arena score (Max) independentas of 2026-10-15 · source
  • LMArena WebDev: 1741 Arena score (Max) independentas of 2026-10-17 · source
  • Artificial Analysis Intelligence Index: 53 index points (max with fallback) independentas of 2026-10-17 · source
  • Output speed: 71 tokens per second (max with fallback) independentas of 2026-10-17 · source
  • LiveBench 2.0 Reasoning: 91.7 score out of 100 (highest supported effort) independentas of 2026-10-14 · source
  • LiveBench 2.0 Coding: 97.7 score out of 100 (highest supported effort) independentas of 2026-10-14 · source
  • Humanity's Last Exam (no tools): 60.9% maker's own numberas of 2026-09-22 · source

Official page for Claude Fable 5.1

Claude Haiku 5.5 Anthropic

Fast and very cheap, with surprisingly strong reasoning for its price.

Company
Anthropic
Released
2026-10-07
Open weights
No
Context
1,000,000 tokens
Price per million tokens
$0.1 in / $0.5 out · price source

All its scores

  • LMArena WebDev: 1585 Arena score (High) independentas of 2026-10-17 · source
  • Artificial Analysis Intelligence Index: 43 index points (max) independentas of 2026-10-17 · source
  • Output speed: 235 tokens per second (max) independentas of 2026-10-17 · source
  • LiveBench 2.0 Reasoning: 84.6 score out of 100 (highest supported effort) independentas of 2026-10-14 · source
  • LiveBench 2.0 Coding: 84.5 score out of 100 (highest supported effort) independentas of 2026-10-14 · source
  • GPQA Diamond: 85% maker's own numberas of 2026-10-17 · source
  • Humanity's Last Exam (no tools): 45.9% maker's own numberas of 2026-10-07 · source

Official page for Claude Haiku 5.5

Gemini 4 Argon Google

Most preferred by people for everyday chat and writing, and the best fresh-puzzle reasoning score.

Company
Google
Released
2026-10-13
Open weights
No
Context
1,048,576 tokens
Price per million tokens
$2 in / $10 out · price source

All its scores

  • LMArena Text: 1524 Arena score (High) independentas of 2026-10-15 · source
  • LMArena WebDev: 1678 Arena score (High) independentas of 2026-10-17 · source
  • Artificial Analysis Intelligence Index: 53 index points (high) independentas of 2026-10-17 · source
  • LiveBench 2.0 Reasoning: 95.1 score out of 100 (highest supported effort) independentas of 2026-10-14 · source
  • LiveBench 2.0 Coding: 89.7 score out of 100 (highest supported effort) independentas of 2026-10-14 · source

Official page for Gemini 4 Argon

Gemini 3.5 Flash-Lite Google

The fastest model here, for quick and simple tasks.

Company
Google
Released
2026-07-21
Open weights
No
Context
1,048,576 tokens
Price per million tokens
$0.3 in / $2.5 out · price source

All its scores

  • LMArena Text: 1455 Arena score independentas of 2026-10-15 · source
  • LMArena WebDev: 1441 Arena score independentas of 2026-10-17 · source
  • Artificial Analysis Intelligence Index: 22 index points independentas of 2026-10-17 · source
  • Output speed: 449 tokens per second independentas of 2026-10-17 · source

Official page for Gemini 3.5 Flash-Lite

Grok 4.7 xAI

Solid mid-range model at a moderate price.

Company
xAI
Released
2026-09-21
Open weights
No
Context
500,000 tokens
Price per million tokens
$2 in / $6 out · price source

All its scores

  • LMArena Text: 1444 Arena score (xHigh) independentas of 2026-10-15 · source
  • LMArena WebDev: 1639 Arena score (xHigh) independentas of 2026-10-17 · source
  • Artificial Analysis Intelligence Index: 46 index points (xhigh) independentas of 2026-10-17 · source
  • Output speed: 93 tokens per second (xhigh) independentas of 2026-10-17 · source
  • LiveBench 2.0 Reasoning: 82.7 score out of 100 (highest supported effort) independentas of 2026-10-14 · source
  • LiveBench 2.0 Coding: 84.5 score out of 100 (highest supported effort) independentas of 2026-10-14 · source

Official page for Grok 4.7

Grok 4.3 xAI

An older, cheaper xAI model; it scores below the newer ones.

Company
xAI
Open weights
No
Context
1,000,000 tokens
Price per million tokens
$1.25 in / $2.5 out · price source

All its scores

  • LMArena Text: 1442 Arena score independentas of 2026-10-15 · source
  • LMArena WebDev: 1358 Arena score independentas of 2026-10-17 · source

Official page for Grok 4.3

DeepSeek-V4.1-Flash DeepSeek

The strongest open model here: free to download, cheap to use and quick.

Company
DeepSeek
Released
2026-09-10
Open weights
Yes, you can download it
Context
1,000,000 tokens
Price per million tokens
$0.3 in / $1.2 out · price source

All its scores

  • LMArena Text: 1475 Arena score (Max) independentas of 2026-10-15 · source
  • LMArena WebDev: 1618 Arena score (Max) independentas of 2026-10-17 · source
  • Artificial Analysis Intelligence Index: 39 index points (max) independentas of 2026-10-17 · source
  • Output speed: 176 tokens per second (max) independentas of 2026-10-17 · source
  • LiveBench 2.0 Reasoning: 86.7 score out of 100 (highest supported effort) independentas of 2026-10-14 · source
  • LiveBench 2.0 Coding: 85.4 score out of 100 (highest supported effort) independentas of 2026-10-14 · source
  • GPQA Diamond: 90.9% maker's own numberas of 2026-10-17 · source
  • Humanity's Last Exam (no tools): 36.8% maker's own numberas of 2026-10-17 · source

Official page for DeepSeek-V4.1-Flash

Qwen3.8-27B Alibaba (Qwen)

A capable open model small enough to run on a good home computer.

Company
Alibaba (Qwen)
Released
2026-08-14
Open weights
Yes, you can download it
Context
262,144 tokens

All its scores

  • LMArena Text: 1438 Arena score independentas of 2026-10-15 · source
  • LMArena WebDev: 1593 Arena score independentas of 2026-10-17 · source
  • Artificial Analysis Intelligence Index: 34 index points (xhigh) independentas of 2026-10-17 · source
  • Output speed: 41 tokens per second (xhigh) independentas of 2026-10-17 · source
  • LiveBench 2.0 Reasoning: 80.0 score out of 100 (highest supported effort) independentas of 2026-10-14 · source
  • LiveBench 2.0 Coding: 86.8 score out of 100 (highest supported effort) independentas of 2026-10-14 · source
  • GPQA Diamond: 89.2% maker's own numberas of 2026-10-17 · source
  • Humanity's Last Exam (no tools): 30.8% maker's own numberas of 2026-10-17 · source

Official page for Qwen3.8-27B

Gemma 4 31B Google

Google's open model you can run yourself; best for simple tasks.

Company
Google
Released
2026-03-31
Open weights
Yes, you can download it
Context
262,144 tokens

All its scores

  • LMArena Text: 1452 Arena score independentas of 2026-10-15 · source
  • LMArena WebDev: 1367 Arena score independentas of 2026-10-17 · source
  • Artificial Analysis Intelligence Index: 15 index points independentas of 2026-10-17 · source
  • Output speed: 35 tokens per second independentas of 2026-10-17 · source
  • LiveBench 2.0 Reasoning: 68.3 score out of 100 (highest supported effort) independentas of 2026-10-14 · source
  • LiveBench 2.0 Coding: 49.0 score out of 100 (highest supported effort) independentas of 2026-10-14 · source
  • GPQA Diamond: 84.3% maker's own numberas of 2026-10-17 · source
  • Humanity's Last Exam (no tools): 19.5% maker's own numberas of 2026-10-17 · source

Official page for Gemma 4 31B

Mistral Medium 3.5 Mistral AI

Open model that answers quickly, but scores lower on hard tests.

Company
Mistral AI
Released
2026-04-28
Open weights
Yes, you can download it
Context
262,144 tokens

All its scores

  • LMArena Text: 1426 Arena score independentas of 2026-10-15 · source
  • LMArena WebDev: 1264 Arena score independentas of 2026-10-17 · source
  • Artificial Analysis Intelligence Index: 14 index points independentas of 2026-10-17 · source
  • Output speed: 160 tokens per second independentas of 2026-10-17 · source

Official page for Mistral Medium 3.5

How to read this
  • What a benchmark is: a fixed test, like an exam, that every model takes so you can compare them on the same questions.
  • Maker's own numbers can be biased: a company picks the tests it does well on. Scores tagged "independent" were run by an outside group.
  • Ranks change fast: new models come out every few weeks, so check the "as of" date on each test.
  • One number never tells the whole story: price, speed, privacy and how a model feels to use matter too. Try two or three on your own tasks.
What we left out on purpose
  • Vision: no trusted vision score covers these models yet, so vision is left out.
  • Left out on purpose: SWE-bench and Aider (none of these models on the official boards yet), a maker's number for a rival's model, and averages we would have to compute ourselves.
Sources (6 pages)

Updated 2026-10-17 18:00 IST. Scores and prices come only from these pages and each model's own source links above:

📲 Want the news when a new model tops these charts? Follow ITG Now on Instagram, or read today's AI news.