Claude Benchmarks

As of Nov 6, 2026, Claude Opus 5.5 is first on the public Terminal-Bench 4.0 leaderboard (64.85% inside Claude Code) and first on WebDev Arena (1810), but third on Arena's text board behind Gemini 4 Carbon and Gemini 4 Argon, second on the Artificial Analysis Intelligence Index (59 against Gemini 4 Carbon's 60), and third on ARC-AGI-2 behind GPT-6 Astra and GPT-6.1 Sol (ARC Prize). This page lists only measured, dated results with their sources: Anthropic's own launch tables first, labeled as Anthropic's claims, then the independent boards, then what none of them show. Scores move weekly, so every number carries the day it was read.

Learn more: AI Leaderboards on ai.vi.ki and OpenAI Benchmarks on chatgpt.vi.ki

Quick facts

Sources: Terminal-Bench 4.0, WebDev Arena, Text Arena, Artificial Analysis, ARC Prize, SWE-bench, all read Nov 6, 2026.

What Anthropic reports

Anthropic's launch post for Claude Opus 5.5 (Sep 22, 2026) compares it with Claude Fable 5.1, Claude Opus 5, GPT-6 Astra and GPT-5.6 Sol. These are Anthropic's own runs, at max effort unless noted, with production safeguards switched on (Introducing Claude Opus 5.5).

Source: Introducing Claude Opus 5.5, Anthropic, Sep 22, 2026.

On its own table, Anthropic shows GPT-6 Astra ahead on two rows: AutomationBench (41.4% against 40.0%) and Terminal-Bench-Science (64.6% against 58.7%). Anthropic gives a standard error of about 2.6 points on Terminal-Bench 4.0, so gaps of a point or two on that test are within noise (Introducing Claude Opus 5.5).

Claude Sonnet 5.5's post (Sep 28, 2026) reports 70.6% on Terminal-Bench 4.0, above Opus 5.5, and 55.5% on CursorBench 4.0; on FrontierCode it scores 46.2% at max effort, below GPT-6 Sol's 49.3% (Introducing Claude Sonnet 5.5). Claude Haiku 5.5's post (Oct 7, 2026) gives it 39.2% on Terminal-Bench 4.0 and 1620 on GDPval-AA, against 1437 for GPT-6 Luna (Introducing Claude Haiku 5.5).

Anthropic's posts disagree with each other in places. The Sonnet 5.5 post footnotes Opus 5.5's 66.4% as an "Xhigh" result, while the Opus 5.5 post says max effort unless noted. The Fable 5.1 post (Sep 15, 2026) gives Fable 5.1 1853 on GDPval-AA v2, and the Opus 5.5 post a week later gives it 1735 on the newer v2.1, so those two numbers are not comparable (Introducing Claude Mythos and Fable 5.1; Introducing Claude Opus 5.5).

Terminal-Bench 4.0

Terminal-Bench, run by Stanford and the Laude Institute, scores an agent and a model together on real tasks in a terminal, 330 trials per entry, with a 95% confidence interval. Every Claude row runs inside Claude Code and every OpenAI row inside Codex, so the harness is part of each score.

As read on Nov 6, 2026: Claude Code + Opus 5.5 (max) 64.85% ±3.11 at a total run cost of $4,718.65; Claude Code + Sonnet 5.5 (max) 61.82%; Codex + GPT-6.1 Sol (max) and Codex + GPT-6 Astra (max) 58.18% each; Claude Code + Fable 5.1 (xhigh) 57.88%; Grok Build + Grok 4.7 (xhigh) 37.58%; the best Google entry, Gemini 3.8 Flash in mini-SWE-agent, 19.09%. No Gemini 4 model had an entry (Terminal-Bench 4.0 leaderboard).

The cost column matters as much as the score. Codex + GPT-6.1 Sol's run cost $634.24, about a seventh of Opus 5.5's, for 6.7 fewer points. The public board also sits below Anthropic's own figures: 64.85% against Anthropic's 66.4% for Opus 5.5, and 61.82% against 70.6% for Sonnet 5.5.

SWE-bench Verified

SWE-bench Verified asks a model to fix 500 real GitHub issues, judged by each repository's tests. It is no longer a current measure: the newest submission on swebench.com on Nov 6, 2026 is dated Feb 26, 2026, and no Claude 5.x, GPT-6 or Gemini 4 model appears. Its top rows are two agents on Claude Opus 4.5 at 79.2% (Dec 2025). With one simple harness for every model, Claude Opus 4.5 (high) led at 76.8%, ahead of Gemini 3 Flash at 75.8% and GPT 5.2 (high) at 72.8% (SWE-bench leaderboards).

Anthropic stopped quoting SWE-bench Verified with its 5.x models, switching to Terminal-Bench 4.0, FrontierCode and CursorBench (Introducing Claude Opus 5.5). Its earlier self-reported figures ran higher than the public rows because of its own harness and prompt: Opus 4.6 at 81.42% in Anthropic's post against 75.6% on the public board (Introducing Claude Opus 4.6).

Arena (formerly LMArena)

Arena (lmarena.ai now redirects there) ranks models by people's blind votes between two answers. On the text board (updated Oct 28, 2026, about 9 million votes), Gemini 4 Carbon (High) leads with 1528, Gemini 4 Argon (High) has 1520 and Claude Opus 5.5 (High) 1509; Opus 5.5's rank spread runs from 2 to 8, so the top three are not cleanly separated. Anthropic holds six of the top ten entries, counting variants; OpenAI's best is GPT-5.6 Sol (xHigh) at 22nd (Text Arena).

On WebDev Arena, where people vote on web apps the models build, Opus 5.5 (Max) is first at 1810 and the only model whose rank spread is 1 to 1, ahead of GPT-6 Astra (1773), Sonnet 5.5 (1770) and Gemini 4 Carbon (1768). On the vision board (Oct 19, 2026), Claude Fable 5 is first at 1309 but the spreads are wide (Vision Arena). Claude has no entry on the text-to-image board because it does not generate images; OpenAI holds the top five places there (Text-to-Image Arena).

Artificial Analysis

Artificial Analysis runs nine tests of its own (knowledge work, coding, science and general knowledge) and combines them into one index. On Nov 6, 2026: Gemini 4 Carbon (high) 60; Claude Opus 5.5 (max) 59; GPT-6 Astra (max) 57; GPT-6.1 Sol (max) 55; Claude Sonnet 5.5 (max) 55; Gemini 4 Argon (high) 50; Claude Fable 5.1 (xhigh) 50; Grok 4.7 (xhigh) 35.

Cost per task separates them more than the score does: Opus 5.5 at max effort cost $10.04 a task against $1.27 for GPT-6.1 Sol at max effort, four index points lower. Fable 5.1, Anthropic's top general model, scores below both Opus 5.5 and Sonnet 5.5 on this index.

ARC-AGI, Humanity's Last Exam and GPQA

On ARC Prize's leaderboard (puzzles easy for people, hard for AI), ARC-AGI-2's top results are GPT-6 Astra (Max) 95.0%, GPT-6.1 Sol (Max) 94.2%, Claude Opus 5.5 (High) 93.3% at $0.408 a task, and Gemini 4 Argon (High) 91.3%. On the interactive ARC-AGI-3, OpenAI and Google post high results and Opus 5.5 has none listed.

Humanity's Last Exam (2,500 expert questions) gives different orders on different boards. Artificial Analysis's runs put Opus 5.5 (max) first at 61.4%, ahead of Gemini 4 Argon (57.1%) and GPT-6 Astra (54.7%). Scale AI's board, whose newest date is Sep 17, 2026, puts GPT-6 Astra first at 54.80 and does not list Opus 5.5 (Scale AI).

On GPQA Diamond, graduate-level science questions where top models now score in the 90s, GPT-6 Astra leads at 96.1%; the best Claude rows are Fable 5.1 and Opus 5 at 93.7%, and Opus 5.5 had no score on Nov 6, 2026.

What benchmarks do not show

7:50 A benchmark measures one task at one setting. Claude's scores change a great deal with effort: Opus 5.5 scores 59 at max and 52 at high on the Artificial Analysis index, and the Claude apps do not run every model at max effort. Harnesses differ too: a Claude result inside Claude Code is not the same model test as a GPT result inside Codex.

Price per task is not price per token. Claude 4.7 and later models use a tokenizer that produces about 30% more tokens for the same text, and models differ in how much they think, so equal list prices can mean unequal bills. Benchmarks also say nothing about usage limits, features such as image generation, or how a model feels to write with; for those, see the comparison pages such as Claude vs ChatGPT and Claude vs Gemini.

What reviewers found

6:11 In an Aug 27, 2026 video, Tim Ruscica of Tech With Tim warned that the labs are "marking their own homework": OpenAI's launch chart for GPT-5.6 Sol led with benchmarks where it beat Claude, and Anthropic's for Claude Opus 5 led with others where Claude won. On one independent benchmark that also grades whether the code is good, he said Fable 5, Opus 5 and GPT-5.6 Sol tied at 34.7%, on only about 100 tasks. Those were the models of August, before Opus 5.5 and GPT-6.

0:14 In a Sep 24, 2026 video, Károly Zsolnai-Fehér of Two Minute Papers asked Claude Opus 5.5 to reproduce physics simulations from research papers: it ran a honey-coiling simulation in real time, which he said GPT-6 Astra could not, and built muscle-driven walking characters, another task Astra failed, with results "not perfect" but close to the paper's. He added that such runs can get expensive and, reading Anthropic's system card, that hallucinations are still not solved.

Frequently asked questions

Is Claude the best AI model in 2026? No single board says so. On Nov 6, 2026, Claude Opus 5.5 leads Terminal-Bench 4.0 and WebDev Arena, while Gemini 4 Carbon leads Arena's text board and the Artificial Analysis index, and GPT-6 Astra leads ARC-AGI-2 and GPQA Diamond.

What is Claude's SWE-bench score? The public SWE-bench Verified board has no Claude 5.x result; its best Claude rows are Claude Opus 4.5 agents at 79.2% from Dec 2025. Anthropic now reports Terminal-Bench 4.0 and FrontierCode instead.

Which Claude model is best at coding? On the public Terminal-Bench 4.0 board, Claude Opus 5.5 in Claude Code scores 64.85% and Sonnet 5.5 61.82%, while Anthropic's own post reports Sonnet 5.5 higher (70.6%). Opus 5.5 also tops WebDev Arena.

Why are Anthropic's numbers higher than the leaderboards? Anthropic runs its own harness, effort settings and trials, and the public boards run theirs. Sonnet 5.5's own Terminal-Bench 4.0 claim is about nine points above its public leaderboard row, so compare models on the same independent board.

Where does Claude rank on LMArena? Claude Opus 5.5 (High) is third on Arena's text board at 1509 (updated Oct 28, 2026) and first on WebDev Arena at 1810. lmarena.ai now redirects to arena.ai.

Sources

• Codex vs Claude - an Honest Comparison, Tech With Tim, Aug 27, 2026 (video)
• Claude Opus 5.5 AI: An Incredible Leap Forward, Two Minute Papers, Sep 24, 2026 (video)
• Introducing Claude Opus 5.5, Anthropic, Sep 22, 2026
• Introducing Claude Sonnet 5.5, Anthropic, Sep 28, 2026
• Introducing Claude Haiku 5.5, Anthropic, Oct 7, 2026
• Introducing Claude Mythos and Fable 5.1, Anthropic, Sep 15, 2026
• Introducing Claude Opus 4.6, Anthropic, Feb 5, 2026
• Pricing, Claude Docs, read Nov 6, 2026
• Terminal-Bench 4.0 leaderboard, tbench.ai, read Nov 6, 2026
• SWE-bench leaderboards, SWE-bench, read Nov 6, 2026
• Text Arena, Arena, updated Oct 28, 2026, read Nov 6, 2026
• WebDev Arena, Arena, read Nov 6, 2026
• Vision Arena, Arena, updated Oct 19, 2026, read Nov 6, 2026
• Text-to-Image Arena, Arena, updated Nov 2, 2026, read Nov 6, 2026
• LLM Leaderboard, Artificial Analysis, read Nov 6, 2026
• Humanity's Last Exam Benchmark, Artificial Analysis, read Nov 6, 2026
• GPQA Diamond, Artificial Analysis, read Nov 6, 2026
• Humanity's Last Exam leaderboard, Scale AI, read Nov 6, 2026
• ARC Prize Leaderboard, ARC Prize Foundation, read Nov 6, 2026