Skip to content

AI Model & Agent Leaderboards

Which model or agent should you use for the job? We collect results from public evaluations and official pricing pages, grouped by use case. Every figure links to where it was published and shows the date it applies to. We never estimate or fill in missing numbers.

Last updated

Coding

Which models write and fix real code best?

Code Arena (WebDev): top 8

People compare two anonymous models on the same web-development task and vote for the better result. Higher is better; overlapping ± ranges mean the models are statistically tied.

Scroll the table sideways to see every column.

Code Arena (WebDev): top 8. Higher is better. Source: LMArena, as of Nov 13, 2026.
RankModelLabScoreVotes
1Claude Fable 5.2 (xHigh)Anthropic1,826 ±191,445
2Claude Opus 5.5 (Max)Anthropic1,816 ±124,399
3Claude Sonnet 5.5 (xHigh)Anthropic1,772 ±106,120
4Gemini 4 Carbon (High)Marked preliminary by LMArena.Google1,768 ±152,105
5GPT-6 Astra (Max)OpenAI1,763 ±88,685
6Qwen4-Max (xHigh)Marked preliminary by LMArena.Alibaba1,750 ±151,814
7Claude Fable 5.1 (Max)Anthropic1,743 ±89,990
8GPT-6.1 Sol (Max)OpenAI1,742 ±105,804
Source: LMArena — Code Arena (WebDev) leaderboard, as of .

SWE-Bench Pro (public): top 6Old: no new results since Jul 2026

Share of real, long-horizon issues from copyleft open-source repositories that a model's patch resolves without breaking existing tests. Higher is better.

Scroll the table sideways to see every column.

SWE-Bench Pro (public): top 6. Old: no new results since Jul 2026. Higher is better. Source: Scale AI, as of Jul 9, 2026.
RankModelLabResolvedAdded
1Muse Spark 1.1*Meta61.5% ±3.12026-07-09
2GPT-5.4 (xHigh)*OpenAI59.1% ±3.62026-04-08
3Muse Spark*Meta55.0% ±3.62026-04-08
4Claude Opus 4.6 (thinking)*Anthropic51.9% ±3.62026-04-08
5Gemini 3.1 Pro (thinking)*Google46.1% ±3.62026-04-08
6Claude Opus 4.5 (2025-11-01)Anthropic45.9% ±3.62025-12-11
Source: Scale AI — SWE-Bench Pro (public set), as of .* Run with the mini-swe-agent harness (Scale AI's footnote).The newest entry is from July 2026, so the latest model generation is not on this board yet.

SWE-bench Verified, same harness: top 7Old: no new results since Feb 2026

Share of 500 human-validated GitHub issues resolved when every model runs in the same mini-SWE-agent environment. Higher is better.

Scroll the table sideways to see every column.

SWE-bench Verified, same harness: top 7. Old: no new results since Feb 2026. Higher is better. Source: SWE-bench team, as of Feb 19, 2026.
ModelResolvedReported run cost (500 tasks)Run date
Claude 4.5 Opus (high)76.8%$376.952026-02-17
Gemini 3 Flash (high)75.8%$177.982026-02-17
MiniMax M2.5 (high)75.8%$36.642026-02-17
Claude 4.6 Opus75.6%$275.762026-02-17
GLM 5 (high)72.8%$267.192026-02-17
GPT 5.2 (high)72.8%$236.782026-02-17
GPT 5.2 Codex72.8%$224.712026-02-19
Source: SWE-bench team — SWE-bench Verified (mini-SWE-agent runs), as of .No new Verified results have been published since February 2026, so current frontier models are missing. Treat it as a historical baseline.

Agents

Which agent + model pairs finish long, multi-step terminal tasks?

Terminal-Bench 4.1: best run per model, top 12

Share of hard, real terminal tasks an agent harness + model pair completes end to end. Higher is better. Cost is the total the submitter reported for all 330 trials.

Scroll the table sideways to see every column.

Terminal-Bench 4.1: best run per model, top 12. Higher is better. Source: Terminal-Bench (Stanford / Laude Institute), as of Nov 12, 2026.
RankAgent + modelEffortResolvedTotal cost (330 trials)Cost per trial (total ÷ 330)Date
1Claude Code + Claude Opus 5.5max65.8% ±3.1$4,719$14.302026-09-22
1Claude Code + Claude Fable 5.2max65.8% ±2.9$16,493$49.982026-11-09
3Antigravity CLI + Gemini 4 Carbonhigh64.6% ±3.2$1,358$4.122026-10-28
4Claude Code + Claude Sonnet 5.5max63.9% ±3.0$7,334$22.222026-09-28
5Codex + GPT-6.1 Solmax60.9% ±3.0$634$1.922026-09-29
6Codex + GPT-6 Astramax60.3% ±2.8$3,267$9.902026-09-03
8Claude Code + Kimi K3.1 Promax59.7% ±4.2$1,087$3.292026-10-24
12Antigravity CLI + Gemini 4 Hydrogenhigh56.4% ±3.0$784$2.382026-11-12
13Antigravity CLI + Gemini 4 Argonhigh55.5% ±2.9$1,109$3.362026-09-30
17Codex + GPT-6.1 Terramax53.3% ±3.1$874$2.652026-10-26
29Grok Build + Grok 4.7xhigh38.8% ±3.6$3,683$11.162026-09-21
30Claude Code + DeepSeek V4.1 Flash 1017max38.5% ±3.8$185$0.562026-10-17
Source: Terminal-Bench (Stanford / Laude Institute) — Terminal-Bench 4.1 leaderboard, as of .Ranks are the board's own; other efforts of the same model are left out, which is why some ranks are skipped.Cost per trial is Metarix arithmetic: the reported total divided by 330 trials.

Hallucination

Which models invent the fewest facts?

Vectara hallucination leaderboard: summarisation

How often a model adds facts that are not in a document it was asked to summarise, judged by Vectara's HHEM model. Lower hallucination rate is better.

Scroll the table sideways to see every column.

Vectara hallucination leaderboard: summarisation. Lower is better. Source: Vectara, as of Oct 22, 2026.
Position (of 111)ModelHallucination rateFactual consistencyAnswer rate
1antgroup/finix_s1_32b1.8%98.2%99.5%
2openai/gpt-5.4-nano-2026-03-173.1%96.9%100.0%
3google/gemini-2.5-flash-lite3.3%96.7%99.5%
6deepseek-ai/DeepSeek-V4.1-Flash4.2%95.8%99.9%
30openai/gpt-6-sol6.5%93.5%100.0%
45deepseek-ai/DeepSeek-V4-Pro8.6%91.4%97.2%
46openai/gpt-6-astra8.7%91.3%100.0%
60anthropic/claude-opus-5-59.9%90.1%99.7%
71moonshotai/kimi-k2.610.8%89.2%99.7%
85google/gemini-4-argon12.1%87.9%99.1%
103xai-org/grok-4-1-fast-reasoning19.2%80.8%99.7%
Source: Vectara — Hallucination Leaderboard (HHEM), as of .Shows the top three overall plus frontier models from major labs; model IDs are as Vectara lists them.Several of the newest models (for example Claude Fable 5.2, GPT-6.1 Sol and Gemini 4 Carbon) are not on this board yet.

AA-Omniscience: knowledge and hallucination

6,000 factual questions across 42 topics. Hallucination rate (lower is better) is the share of questions a model didn't get right where it gave a wrong answer instead of declining. Net Score runs from −100 to 100 and rewards correct answers while penalising wrong ones.

Scroll the table sideways to see every column.

AA-Omniscience: knowledge and hallucination. Lower is better. Source: Artificial Analysis, as of Nov 16, 2026.
Model (setting)Hallucination rateAccuracyNet Score
Gemini 4 Hydrogen (Xhigh)8.5%50.5%46.3
Gemini 4 Carbon (High)11.4%58.3%53.5
Muse Wonder Full (Max)17.6%41.0%30.6
Grok 4.7 (Xhigh)20.0%42.7%31.2
Qwen4 Max28.4%43.7%27.7
GPT-6.1 Terra (Max)34.5%47.6%29.6
Kimi K3.1 Pro37.3%48.9%29.9
Claude Sonnet 5.5 (Max)43.3%51.5%30.5
GPT-6 Astra (Max)49.8%61.6%42.4
Claude Opus 5.5 (Max)50.3%65.5%48.1
GPT-6.1 Sol (Max)51.9%61.8%42.0
Claude Fable 5.2 (Max)59.3%79.1%66.8
Source: Artificial Analysis — AA-Omniscience, as of .A low hallucination rate and a high Net Score are different strengths: Claude Fable 5.2 knows the most, while the Gemini 4 models most often admit when they don't know.Figures rounded to one decimal from the values Artificial Analysis publishes.

Cost per task

What does a task actually cost, and what are the list prices?

Intelligence vs. cost per task

Artificial Analysis Intelligence Index (higher is better) next to the average cost in USD of one task in that index. A cheap model that needs many retries can still cost more per finished job.

Scroll the table sideways to see every column.

Intelligence vs. cost per task. Higher is better. Source: Artificial Analysis, as of Nov 16, 2026.
Model (setting)Intelligence IndexCost per task
Claude Fable 5.2 (Max)64.6$26.54
Gemini 4 Carbon (High)60.0$5.48
Claude Opus 5.5 (Max)59.1$10.04
GPT-6 Astra (Max)56.9$6.24
GPT-6.1 Sol (Max)54.6$1.27
Claude Sonnet 5.5 (Max)54.5$10.02
Gemini 4 Hydrogen (Xhigh)53.9$3.16
GPT-6.1 Terra (Max)51.3$0.96
Kimi K3.1 Pro46.5$3.68
Qwen4 Max42.9$3.27
Claude Haiku 5.5 (Max)36.5$0.41
Grok 4.7 (Xhigh)35.5$7.77
DeepSeek V4.1 Flash 1017 (Max)32.6$0.60
Source: Artificial Analysis — LLM leaderboard (Intelligence Index, cost per task, release dates), as of .

Official API list prices

Standard, pay-as-you-go prices per million tokens from each lab's own pricing page. Batch, caching and long-context rates differ; follow the source link for those.

Scroll the table sideways to see every column.

Official API list prices. Source: OpenAI, as of Nov 16, 2026; Anthropic, as of Nov 16, 2026; Google, as of Nov 16, 2026; xAI, as of Nov 16, 2026.
ModelLabInput / 1M tokensOutput / 1M tokensSource
Claude Fable 5.2Anthropic$8.00$40.00Anthropic
Claude Opus 5.5Anthropic$4.00$20.00Anthropic
Claude Sonnet 5.5Anthropic$2.00$10.00Anthropic
Claude Haiku 5.5Prompts up to 100,000 tokens; longer prompts cost more.Anthropic$0.10$0.50Anthropic
GPT-6 AstraOpenAI$10.00$50.00OpenAI
GPT-6.1 SolOpenAI$2.00$10.00OpenAI
GPT-6.1 TerraOpenAI$1.00$5.00OpenAI
GPT-6 LunaOpenAI$0.10$0.50OpenAI
Gemini 4 CarbonRises to $4 input / $20 output on 1 December 2026.Google$2.00$10.00Google
Gemini 4 HydrogenGoogle$1.50$7.50Google
Grok 4.7xAI$2.00$6.00xAI
Sources: OpenAI — OpenAI API pricing, as of ; Anthropic — Claude API pricing, as of ; Google — Gemini Developer API pricing, as of ; xAI — xAI models and pricing, as of .

Release tracker

New models (release dates as listed by Artificial Analysis) and new agent results (submission dates on Terminal-Bench 4.1), newest first.

  1. Qwen4 Max · Alibaba · Model

    Listed with an Intelligence Index of 42.9. Artificial Analysis

  2. Gemini 4 Hydrogen · Google · Model

    Lowest AA-Omniscience hallucination rate in our table (8.5%). Artificial Analysis

  3. Antigravity CLI + Gemini 4 Hydrogen · Google · Agent result

    New Terminal-Bench 4.1 entry: 56.4% resolved. Terminal-Bench (Stanford / Laude Institute)

  4. Claude Fable 5.2 · Anthropic · Model

    Highest Intelligence Index on the board (64.6 at Max). Artificial Analysis

  5. Claude Code + Claude Fable 5.2 · Anthropic · Agent result

    Ties for first on Terminal-Bench 4.1 at 65.8%. Terminal-Bench (Stanford / Laude Institute)

  6. Muse Ember · Meta · Model

    New release listed. Artificial Analysis

  7. Tencent Hy4 · Tencent · Model

    New release listed. Artificial Analysis

  8. Mistral Large 4 · Mistral · Model

    Open-weights release. Artificial Analysis

  9. Gemini 4 Carbon · Google · Model

    Intelligence Index 60.0 at High. Artificial Analysis

  10. Antigravity CLI + Gemini 4 Carbon · Google · Agent result

    New Terminal-Bench 4.1 entry: 64.6% resolved. Terminal-Bench (Stanford / Laude Institute)

  11. GPT-6.1 Terra · OpenAI · Model

    Intelligence Index 51.3 at about $0.96 per task. Artificial Analysis

  12. Kimi K3.1 Pro · Kimi · Model

    Open-weights; Intelligence Index 46.5. Artificial Analysis

  13. DeepSeek V4.1 Flash 1017 · DeepSeek · Model

    Open-weights; $0.60 per task on the Intelligence Index. Artificial Analysis

  14. Muse Wonder · Meta · Model

    New release listed. Artificial Analysis

  15. Claude Haiku 5.5 · Anthropic · Model

    Lowest cost per task in our table ($0.41). Artificial Analysis

  16. Gemini 4 Argon · Google · Model

    New release listed. Artificial Analysis

  17. GPT-6.1 Sol · OpenAI · Model

    Intelligence Index 54.6 at $1.27 per task. Artificial Analysis

  18. Claude Sonnet 5.5 · Anthropic · Model

    New release listed. Artificial Analysis

  19. Claude Opus 5.5 · Anthropic · Model

    Intelligence Index 59.1 at Max. Artificial Analysis

Sources & method

Figures are copied as published. We round only where noted and never estimate, fill in or combine scores across benchmarks. Ranks are the publisher's own. Tables show a selection of rows, so read the full board at the source before you decide.

Benchmarks measure different things and change over time. A model missing from a table usually means the publisher has not tested it yet, not that it did badly.

Spotted an outdated or wrong number? Tell the Metarix Team via the contact page.

Every source used on this page, with the date its figures apply to and the date we read it.
SourcePublisherFigures as ofChecked
Code Arena (WebDev) leaderboardElo-style score from blind human votes on front-end and agentic web-dev tasks.LMArena
Search Arena leaderboardElo-style score from blind human votes on answers that use live web search.LMArena
SWE-Bench Pro (public set)asOf is the newest entry on the board.Scale AI
SWE-bench Verified (mini-SWE-agent runs)Figures from the board's published data file (data/leaderboards.json); asOf is the newest Verified entry.SWE-bench team
Terminal-Bench 4.1 leaderboardasOf is the newest submission date on the board. Each entry is an agent + model pair run for 330 trials.Terminal-Bench (Stanford / Laude Institute)
Hallucination Leaderboard (HHEM)Vectara
AA-OmniscienceThe page shows no stamp; asOf is the day it was read.Artificial Analysis
LLM leaderboard (Intelligence Index, cost per task, release dates)The page shows no stamp; asOf is the day it was read.Artificial Analysis
Claude API pricingAnthropic
OpenAI API pricingOpenAI
Gemini Developer API pricingGoogle
xAI models and pricingxAI