Anthropic · closed model
#1 of 17 models we track on the Artificial Analysis Intelligence Index, an all-round score (58 index points). Best result: #1 of 18 on LMArena WebDev (build websites & apps, 1807 Arena score). Anthropic has 2 more models on this page.
See all its scores and sourcesGoogle · closed model
Tied #2 of 17 models we track on the Artificial Analysis Intelligence Index, an all-round score (53 index points). Best result: #1 of 15 on LMArena Text (chat & writing, 1524 Arena score). Google has 2 more models on this page.
See all its scores and sourcesOpenAI · closed model
Tied #2 of 17 models we track on the Artificial Analysis Intelligence Index, an all-round score (53 index points). Best result: #1 of 11 on ARC-AGI-2 (new puzzles, 95.0%). OpenAI has 1 more model on this page.
See all its scores and sourcesxAI · closed model
Tied #5 of 17 models we track on the Artificial Analysis Intelligence Index, an all-round score (46 index points). xAI has 1 more model on this page.
See all its scores and sourcesXiaomi · open model, free to download
Tied #5 of 17 models we track on the Artificial Analysis Intelligence Index, an all-round score (46 index points). Best open model on that index.
See all its scores and sourcesZ.ai · open model, free to download
#7 of 17 models we track on the Artificial Analysis Intelligence Index, an all-round score (45 index points). Best result: #4 of 6 on Terminal-Bench 4.0 (agent tasks, 41.8 % of tasks done).
See all its scores and sourcesMoonshot AI · open model, free to download
#8 of 17 models we track on the Artificial Analysis Intelligence Index, an all-round score (44 index points). Best result: #4 of 15 on LMArena Text (chat & writing, 1487 Arena score).
See all its scores and sourcesDeepSeek · open model, free to download
Tied #10 of 17 models we track on the Artificial Analysis Intelligence Index, an all-round score (39 index points). Best result: #1 of 4 on OpenRouter usage (this week) (most used (this week), 36.8 trillion tokens this week).
See all its scores and sourcesMeta · open model, free to download
Tied #10 of 17 models we track on the Artificial Analysis Intelligence Index, an all-round score (39 index points). Best result: #6 of 16 on Output speed (speed, 131 tokens per second).
See all its scores and sourcesAlibaba (Qwen) · open model, free to download
#13 of 17 models we track on the Artificial Analysis Intelligence Index, an all-round score (34 index points). Best result: #4 of 6 on GPQA Diamond (science questions, 89.2%).
See all its scores and sourcesReflection AI · open model, free to download
#14 of 17 models we track on the Artificial Analysis Intelligence Index, an all-round score (30 index points). Best result: #3 of 16 on Output speed (speed, 228 tokens per second).
See all its scores and sourcesMistral AI · open model, free to download
#17 of 17 models we track on the Artificial Analysis Intelligence Index, an all-round score (14 index points). Best result: #5 of 16 on Output speed (speed, 160 tokens per second).
See all its scores and sourcesHome › AI model benchmarks
18 AI models on 11 tests, ranked from published scores. Every number names its source.
No single model wins every test: Claude Opus 5.5 leads 4 of 11, Gemini 4 Argon and GPT-6 Astra lead 2 each.
Open = you can download the model. Closed = you can only use it through the maker's app or paid API.
Open-weight models handled the most text-model tokens on OpenRouter this week. Closed models still lead the independent boards we track (10 of 11 measures).
| Job | Best open | Best closed | Gap |
|---|---|---|---|
| All-roundArtificial Analysis Intelligence Index · index points | OpenMiMo-V2.6-ProXiaomi · MIT: free for almost any use46 | ClosedClaude Opus 5.5Anthropic58 | 12 points behind |
| Chat & writingLMArena Text · Arena score | OpenKimi K3Moonshot AI · Kimi K3 license: free to use with conditions1487 | ClosedGemini 4 ArgonGoogle1524 | 37 points behind |
| Build websites & appsLMArena WebDev · Arena score | OpenKimi K3Moonshot AI · Kimi K3 license: free to use with conditions1655 | ClosedClaude Opus 5.5Anthropic1807 | 152 points behind |
| CodingLiveBench 2.0 Coding · score out of 100 | OpenQwen3.8-27BAlibaba (Qwen) · Apache 2.0: free for almost any use86.8 | ClosedClaude Fable 5.1Anthropic97.7 | 10.9 points behind |
| ReasoningLiveBench 2.0 Reasoning · score out of 100 | OpenKimi K3Moonshot AI · Kimi K3 license: free to use with conditions90.7 | ClosedGemini 4 ArgonGoogle95.1 | 4.4 points behind |
| Science questionsGPQA Diamond · % | OpenDeepSeek-V4.1-Flash maker's own numberDeepSeek · MIT: free for almost any use90.9% | ClosedGPT-6 Astra maker's own numberOpenAI96.0% | 5.1 points behind |
| Hardest examHumanity's Last Exam (no tools) · % | OpenDeepSeek-V4.1-Flash maker's own numberDeepSeek · MIT: free for almost any use36.8% | ClosedClaude Opus 5.5 maker's own numberAnthropic64.4% | 27.6 points behind |
| New puzzlesARC-AGI-2 · % | OpenDeepSeek-V4.1-FlashDeepSeek · MIT: free for almost any use72.9% | ClosedGPT-6 AstraOpenAI95.0% | 22.1 points behind |
| Agent tasksTerminal-Bench 4.0 · % of tasks done | OpenGLM-5.3Z.ai · GLM-5.3 license: free to use with conditions41.8 | ClosedClaude Opus 5.5Anthropic64.8 | 23 points behind |
| Most used (this week)OpenRouter usage (this week) · trillion tokens this week | OpenDeepSeek-V4.1-FlashDeepSeek · MIT: free for almost any use36.8 | ClosedGPT-6 LunaOpenAI6.5 | Open ahead by 30.3 trillion tokens this week |
| SpeedOutput speed · tokens per second | OpenBeam 1.0Reflection AI · Apache 2.0: free for almost any use228 | ClosedGemini 3.5 Flash-LiteGoogle449 | 221 tokens per second slower |
Run it yourself. 9 of the 18 models we track have open weights: you can download them and run them on your own computer or server, so your data stays with you (the biggest ones need a powerful machine). 5 of them use MIT or Apache 2.0, free for almost any use.
Price per million tokens (closed is cheaper here). Cheapest open model with a listed price: DeepSeek-V4.1-Flash: $0.3 in / $1.2 out. Cheapest closed: GPT-6 Luna and Claude Haiku 5.5: $0.1 in / $0.5 out. Only 1 of the 9 open models list a price; the rest you run yourself or through a host you pick.
Best open-weight vs best closed model among the models we track, per test. Updated 18 Oct 2026. Each row's source and date: Sources.
Every model on every test with 3 or more models. Greener = better rank.
Swipe sideways for all 11 tests
Price, memory and the newest release.
Open a test for every score, what it measures and where the numbers come from.
One combined score from many independent tests of how smart a model is.
Higher is better · 17 models · independent
Source: Artificial Analysis, as of 17 Oct 2026 · 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.
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 · 15 models · independent · bars start at 1421, not 0
Source: LMArena (arena.ai), as of 15 Oct 2026 · 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.
When people ask two AIs to build the same website or app and compare the results, which one do they pick?
Higher is better · 18 models · independent · bars start at 1237, not 0
Source: LMArena (arena.ai), as of 17 Oct 2026 · 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.
Fresh coding tasks with tests that check the answer, changed often so models cannot memorize them.
Higher is better · 15 models · independent
Source: LiveBench 2.0, as of 14 Oct 2026 · 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.
Fresh logic and puzzle questions with one right answer, changed often so models cannot memorize them.
Higher is better · 15 models · independent
Source: LiveBench 2.0, as of 14 Oct 2026 · 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.
Very hard science questions in biology, chemistry and physics, written by PhD experts.
Higher is better · 6 models · maker's own number
Source: Makers' own model cards, as of 17 Oct 2026 · 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.
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
Source: Makers' own model cards, as of 17 Oct 2026 · 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.
New picture puzzles a model has never seen, so it has to reason instead of remember. Harder for AI than for people.
Higher is better · 11 models · independent
Source: ARC Prize (arcprize.org), as of 18 Oct 2026 · independent test
Each model's best listed setting; the board also shows cost per task, kept in each score's variant.
Can an AI agent finish real jobs in a computer's command line, like setting up software or fixing a broken project?
Higher is better · 6 models · independent
Source: Terminal-Bench (tbench.ai), as of 18 Oct 2026 · independent test
Scores depend on the agent program used (Claude Code, Codex, Grok Build), listed in each score's variant. Each has a plus or minus range of about 3 points.
How much developers actually use each model through OpenRouter, a popular service that connects apps to many AI models.
Higher is better · 4 models · independent
Source: OpenRouter rankings, as of 16 Oct 2026 · independent test
Real usage, not quality. Only models in OpenRouter's weekly top 10 are listed. Free and cheap models get more traffic.
How many tokens (pieces of words) per second the model writes once it starts answering. Higher is faster.
Higher is better · 16 models · independent
Source: Artificial Analysis, as of 17 Oct 2026 · independent test
Median speed for the same setting used in the Intelligence Index. Gemini 4 Argon has no speed listed yet.
Price, memory, release date and every score with its source.
Near the top for coding and hard reasoning; slower to start on its deepest settings.
All its scores
Very cheap and quick for simple everyday jobs.
All its scores
Top pick for building apps and websites, and the highest overall intelligence score here.
All its scores
Strong all-rounder for writing and coding, but one of the priciest.
All its scores
Fast and very cheap, with surprisingly strong reasoning for its price.
All its scores
Most preferred by people for everyday chat and writing, and the best fresh-puzzle reasoning score.
All its scores
The fastest model here, for quick and simple tasks.
All its scores
Solid mid-range model at a moderate price.
All its scores
An older, cheaper xAI model; it scores below the newer ones.
All its scores
The strongest open model here: free to download, cheap to use and quick.
All its scores
A capable open model small enough to run on a good home computer.
All its scores
Google's open model you can run yourself; best for simple tasks.
All its scores
The open model people rate highest for chat and writing on LMArena right now; very large, so most people use it through a hosted service.
All its scores
Xiaomi's open model: the highest overall intelligence score among open models here, under the simple MIT license.
All its scores
Strong open all-rounder with one of the best open coding scores on LiveBench.
All its scores
Meta's new open model: quick to answer, mid-pack on hard tests.
All its scores
Reflection AI's open model for coding agents: among the fastest models here, with a strong LiveBench coding score, Apache 2.0.
All its scores
Open model that answers quickly, but scores lower on hard tests.
All its scores
Updated 18 Oct 2026, 12:00 IST. Scores and prices come only from these pages and each model's own source links above:
Open vs closed, row by row:
We count visits with Google Analytics to see what helps. OK with that?
ITG Now