Reasoning
198-question Diamond subset; independent AA evaluation
- 01
- 02
- 03
- 03
- 03
- 22
Claude Opus 4.6
Across the task benchmarks we track, Claude Opus 4.6’s strongest relative results are in reasoning. On GPQA Diamond, it scores 89.6%, ranking #22 among 30 models with reported results in our dataset.
Updated: Data as of:
Strongest results
| Benchmark | Measures | Score |
|---|---|---|
| SWE-bench ↗Verified / mini-SWE-agent 2.0.0 / Bash Only | Coding agents | 75.6%#3 of 7 |
| Humanity's Last Exam ↗Stable HLE — text-only subset | Expert reasoning | 39.9%#32 of 68 |
| GPQA ↗Diamond / Artificial Analysis | Reasoning | 89.6%#22 of 30 |
| MCP-Atlas ↗April 2026 / public 500 | Tool use | 79%#5 of 5 |
Model details
| Model | Claude Opus 4.6 |
|---|---|
| Agent Index v1.2 | Insufficient coverage (0/7) · Preview |
| Creator | Anthropic (opens in a new tab) |
| Availability | Proprietary |
| Context size | 1,000,000 tokens (opens in a new tab) |
| Max output | 128,000 tokens (opens in a new tab) |
| Knowledge cutoff | Not reported |
| License | Not reported |
Pricing and performance
Per provider and reported configuration. Prices are USD per 1M tokens.
| Provider / configuration | Contexttokens | InputUSD / 1M tokens | OutputUSD / 1M tokens | LatencyTTFT · seconds | Speedtokens / sec |
|---|---|---|---|---|---|
| Claude Opus 4.6 (non-reasoning, high) | 1,000,000 | $5.00 (opens in a new tab) | $25.00 (opens in a new tab) | 1.35 (opens in a new tab) | 58.564 (opens in a new tab) |
| Claude Opus 4.6 (non-reasoning, high) | 1,000,000 | $5.00 (opens in a new tab) | $25.00 (opens in a new tab) | 1.778 (opens in a new tab) | 39.935 (opens in a new tab) |
| Claude Opus 4.6 (non-reasoning, high) | 1,000,000 | $5.00 (opens in a new tab) | $25.00 (opens in a new tab) | 2.086 (opens in a new tab) | 40.802 (opens in a new tab) |
| Claude Opus 4.6 (non-reasoning, high) | 1,000,000 | $5.00 (opens in a new tab) | $25.00 (opens in a new tab) | 1.898 (opens in a new tab) | 42.424 (opens in a new tab) |
Benchmark performance
Claude Opus 4.6 in color, alongside leading models. Latest stored reports, best effort per report date.
GPQA Diamond ↗
Reported benchmarks
All published results in our dataset, with versions, evaluation settings, and sources.
| Benchmark | Score | Evaluation setup | Source |
|---|---|---|---|
| AppWorldTest-ChallengeTask Goal Completion · Higher is better | 72.4% | agent: Button Agent · metric: TGC · selection: All 23 leaderboard submissions; latest report date and best score at that date shown per model. Normal-split and SGC values retained in the source data/settings, not mixed into the ranking. · source model label: Claude Opus 4.6 · split: Test-Challenge · test challenge sgc: 64.8 · test normal sgc: 89.3 · test normal tgc: 91.1 | AppWorld research team ↗ (opens in a new tab)benchmark publisher · 2026-03-23 |
| Gaia2CLI · OpenClaw 2026.4.1Pass@1 · Higher is better | 57% | api host: anthropic · capabilities: search, execution, adaptability, ambiguity, time · edition: Gaia2-CLI · harness: OpenClaw 2026.4.1 · reasoning: high · reasoning context: Not specified | Meta Agents Research Environments ↗ (opens in a new tab)benchmark publisher · 2026-04-13 |
| GPQADiamond / Artificial AnalysisAccuracy · Higher is better | 89.596% | Claude Opus 4.6 (Max); 198-question Diamond subset; independent AA evaluation | Artificial Analysis (opens in a new tab)independent |
| GPQADiamond / Artificial AnalysisAccuracy · Higher is better | 84.04% | Claude Opus 4.6 (Non-reasoning, High); 198-question Diamond subset; independent AA evaluation | Artificial Analysis (opens in a new tab)independent |
| Humanity's Last ExamStable HLE — full setAccuracy · Higher is better | 53.1% | Opus 4.6 comparison in Mythos Preview launch table. · subset: full · Tools enabled | Anthropic ↗ (opens in a new tab)provider reported |
| Humanity's Last ExamStable HLE — full setAccuracy · Higher is better | 40% | Opus 4.6 comparison in Mythos Preview launch table. · subset: full · No tools | Anthropic ↗ (opens in a new tab)provider reported |
| Humanity's Last ExamStable HLE — text-only subsetAccuracy · Higher is better | 39.9% | question count: 2158 · reported precision decimals: 1 · Claude Opus 4.6 (max); original HLE; 2,158 text-only questions; no tools; pass@1; chart labels rounded to one decimal. · source model href: /models/claude-opus-4-6-adaptive · subset: text · No tools | Artificial Analysis (opens in a new tab)independent |
| MCP-AtlasApril 2026 / public 500Task success · Higher is better | 79% | max; public 500; 100 tool calls | Scale AI ↗ (opens in a new tab)benchmark publisher |
| MyPCBench184 tasks · 100 stepsPerfect-task rate · Higher is better | 58.2% | interface: computer + bash · judge: gemini-3.1-flash-lite-preview · max steps: 100 · rubric score percent: 85.1 · selection: Canonical matched-protocol rows only; four 200-step self-reported rows retained in source snapshot, excluded from this track. · tasks: 184 | MyPCBench / Carnegie Mellon University ↗ (opens in a new tab)benchmark publisher |
| SWE-benchVerified / mini-SWE-agent 2.0.0 / Bash OnlyResolved · Higher is better | 75.6% | attempts: 1 · harness: mini-SWE-agent · harness version: 2.0.0 · Official Bash Only leaderboard; mini-SWE-agent 2.0.0; 500 tasks; one attempt; effort not specified · tasks: 500 | SWE-bench ↗ (opens in a new tab)benchmark publisher · 2026-02-17 |
| π-BenchOverall · PROCProactivity (PROC) · Higher is better | 65.5% | completeness score: 67.6 · completeness uncertainty: 1.5 · metric: Avg PROC · personas: 5 · proc uncertainty: 1.4 · selection: Overall rows; not per-persona averages. COMP is a separate metric and does not determine this ranking. · tasks: 100 | π-Bench research team ↗ (opens in a new tab)benchmark publisher |