RankingTerminal-Bench · 4.0
Terminal-Bench · 4.0
Terminal-Bench · 4.0 is published here as its own result table, version 4.0. Scores use percent, and higher values rank better. The display harness is Claude Fable 5.1 and Claude Mythos 5.1 System Card. Data updated 22 Nov 2026.
- Bucket
- Coding
- Unit
- percent
- Direction
- Higher is better
- Version
- 4.0
- Display harness
- Claude Fable 5.1 and Claude Mythos 5.1 System Card
- Board
- https://www.anthropic.com/claude-fable-5-1-mythos-5-1-system-card
Compare published benchmark results with category weights →
Models
1–18 of 18 entries
| Model | Score | Harness | Evidence | Source-recorded date |
|---|---|---|---|---|
| Claude Fable 5.1Anthropic | 55.8%Reported settings & sourceClaude Code --bare max effort, 15 trials/task over 66 tasks; Anthropic internal reruns. Uses more precise section values rather than rounded headline table. Distinct from2.1; timeouts/resources changed. Fable is the safeguarded deployed configuration; selected tasks may use disclosed Opus fallback, except evaluations explicitly counting safety blocks as failures. Claude Fable 5.1 and Claude Mythos 5.1 System Card · 8.6 · reviewed 2026-11-22 | Claude Fable 5.1 and Claude Mythos 5.1 System Card | lab self-report | 2026-11-22 |
| Claude Opus 5Anthropic | 52.3%Reported settings & sourceClaude Code --bare max effort, 15 trials/task over 66 tasks; Anthropic internal reruns. Uses more precise section values rather than rounded headline table. Distinct from2.1; timeouts/resources changed. Claude Fable 5.1 and Claude Mythos 5.1 System Card · 8.6 · reviewed 2026-11-22 | Claude Fable 5.1 and Claude Mythos 5.1 System Card | lab self-report | 2026-11-22 |
| Claude Fable 5Anthropic | 42%Reported settings & sourceClaude Code --bare max effort, 15 trials/task over 66 tasks; Anthropic internal reruns. Uses more precise section values rather than rounded headline table. Distinct from2.1; timeouts/resources changed. Fable is the safeguarded deployed configuration; selected tasks may use disclosed Opus fallback, except evaluations explicitly counting safety blocks as failures. Claude Fable 5.1 and Claude Mythos 5.1 System Card · 8.6 · reviewed 2026-11-22 | Claude Fable 5.1 and Claude Mythos 5.1 System Card | lab self-report | 2026-11-22 |
| Claude Fable 5.1Anthropic | 57.9%Reported settings & sourceComparator maximum available thinking/reasoning when reported, otherwise best available result; exact setting not identified by this grid. Argon self-computed; other models from the official public leaderboard. Highest scoring thinking level reported by Terminal-Bench authors. Argon agent identity is not specified in this methodology. Provider-published launch claim, not independent measurement by Google for every row. Argon self-computed; other models from the official public leaderboard. Highest scoring thinking level reported by Terminal-Bench authors. Argon agent identity is not specified in this methodology. Methodology https://deepmind.google/models/evals-methodology/gemini-4-argon, SHA256 ff1df6bdeddc4c0f48840c09a8a7813b10ae7ad9e053892da68dabf64f09bff7. Preserve source rounding. Methodology says capabilities as of September 2026 and results as of October 2026; review date is source retrieval, not evaluation date. Comparison limit: Provider-published claim. Independent source results are admitted separately with exact checkpoint, effort and compatible protocol; this grid does not establish a matched comparison. Google Gemini 4 Argon launch and evaluation methodology · Launch grid row 8: Terminal-Bench 4.0 / Claude Fable 5.1; methodology Additional Details · reviewed 2026-09-30 | Google Gemini 4 Argon launch and evaluation methodology | lab self-report | 2026-09-30 |
| Claude Haiku 5.5Anthropic | 39.2%Reported settings & sourceClaude Code --bare; max thinking effort; 66 tasks, ten trials per task (660 trials); no internet egress; safeguards enabled, no fallback. Blocked requests ended the trial (1.8% of trials, all failed). Provider-published lab_self_report; retained as launch evidence, not an independent board admission. Comparison limit: Provider publication; exact benchmark-specific configuration and independent evaluation provenance require separate review before matched-board admission. Claude Haiku 5.5 System Card · section 8.4; pp.115–116 · reviewed 2026-10-07 | Claude Haiku 5.5 System Card | lab self-report | 2026-10-07 |
| Claude Opus 5Anthropic | 51.8%Reported settings & sourceOfficialpublicboard highest scoring thinking level; nativeagents may differ Gemini3.8Flash Model Card · Model card page5 Results table · reviewed 2026-09-06 | Gemini3.8Flash Model Card | lab self-report | 2026-09-06 |
| Claude Opus 5.5Anthropic | 66.36%Reported settings & sourceClaude Code --bare; xhigh thinking effort; safeguards enabled with server-side fallback (2.5% of requests, 10% of trials); five trials per task (330 trials) on 66 tasks. Provider-published lab self-report from the Claude Opus 5.5 system card (SHA256 7311c9c6bbb16d012f1c12c7418b05949fcf7ae3e30d2c40f22050074b2a7378). Not an independent board. Table 8.1.A and the launch grid round this xhigh result to 66.4%. Section 8.5 states 66.36%. Comparison limit: Provider-published lab self-report. Shown as a labeled claim and not admitted as a matched board comparison. Claude Opus 5.5 System Card · section 8.5 · reviewed 2026-09-22 | Claude Opus 5.5 System Card | lab self-report | 2026-09-22 |
| Claude Opus 5.5Anthropic | 64.8%Reported settings & sourceClaude Code --bare; max thinking effort; safeguards enabled with the default server-side fallback; five trials per task on 66 tasks. Section 8.5 says this is within noise of xhigh. Provider-published lab self-report from the Claude Opus 5.5 system card (SHA256 7311c9c6bbb16d012f1c12c7418b05949fcf7ae3e30d2c40f22050074b2a7378). Not an independent board. Comparison limit: Provider-published lab self-report. Shown as a labeled claim and not admitted as a matched board comparison. Claude Opus 5.5 System Card · section 8.5 · reviewed 2026-09-22 | Claude Opus 5.5 System Card | lab self-report | 2026-09-22 |
| Claude Opus 5.5Anthropic | 66.4%Reported settings & sourceComparator maximum available thinking/reasoning when reported, otherwise best available result; exact setting not identified by this grid. Argon self-computed; other models from the official public leaderboard. Highest scoring thinking level reported by Terminal-Bench authors. Argon agent identity is not specified in this methodology. Provider-published launch claim, not independent measurement by Google for every row. Argon self-computed; other models from the official public leaderboard. Highest scoring thinking level reported by Terminal-Bench authors. Argon agent identity is not specified in this methodology. Methodology https://deepmind.google/models/evals-methodology/gemini-4-argon, SHA256 ff1df6bdeddc4c0f48840c09a8a7813b10ae7ad9e053892da68dabf64f09bff7. Preserve source rounding. Methodology says capabilities as of September 2026 and results as of October 2026; review date is source retrieval, not evaluation date. Comparison limit: Provider-published claim. Independent source results are admitted separately with exact checkpoint, effort and compatible protocol; this grid does not establish a matched comparison. Google Gemini 4 Argon launch and evaluation methodology · Launch grid row 8: Terminal-Bench 4.0 / Claude Opus 5.5; methodology Additional Details · reviewed 2026-09-30 | Google Gemini 4 Argon launch and evaluation methodology | lab self-report | 2026-09-30 |
| Claude Sonnet 5Anthropic | 12.4%Reported settings & sourceOfficialpublicboard highest scoring thinking level; nativeagents may differ Gemini3.8Flash Model Card · Model card page5 Results table · reviewed 2026-09-06 | Gemini3.8Flash Model Card | lab self-report | 2026-09-06 |
| Gemini 3.7 FlashGoogle | 11.2%Reported settings & sourceOfficialpublicboard highest scoring thinking level; nativeagents may differ Gemini3.8Flash Model Card · Model card page5 Results table · reviewed 2026-09-06 | Gemini3.8Flash Model Card | lab self-report | 2026-09-06 |
| Gemini 3.8 FlashGoogle | 19.1%Reported settings & sourceOfficialpublicboard highest scoring thinking level; nativeagents may differ Gemini3.8Flash Model Card · Model card page5 Results table · reviewed 2026-09-06 | Gemini3.8Flash Model Card | lab self-report | 2026-09-06 |
| Gemini 4 ArgonGoogle | 57.4%Reported settings & sourceGemini API highest thinking settings; exact named effort not specified by Google. Argon self-computed; other models from the official public leaderboard. Highest scoring thinking level reported by Terminal-Bench authors. Argon agent identity is not specified in this methodology. Provider-published launch claim, not independent measurement by Google for every row. Argon self-computed; other models from the official public leaderboard. Highest scoring thinking level reported by Terminal-Bench authors. Argon agent identity is not specified in this methodology. Methodology https://deepmind.google/models/evals-methodology/gemini-4-argon, SHA256 ff1df6bdeddc4c0f48840c09a8a7813b10ae7ad9e053892da68dabf64f09bff7. Preserve source rounding. Methodology says capabilities as of September 2026 and results as of October 2026; review date is source retrieval, not evaluation date. Comparison limit: Provider-published claim. Independent source results are admitted separately with exact checkpoint, effort and compatible protocol; this grid does not establish a matched comparison. Google Gemini 4 Argon launch and evaluation methodology · Launch grid row 8: Terminal-Bench 4.0 / Gemini 4 Argon; methodology Additional Details · reviewed 2026-09-30 | Google Gemini 4 Argon launch and evaluation methodology | lab self-report | 2026-09-30 |
| GPT-5.6 SolOpenAI | 37.3%Reported settings & sourceClaude Code --bare max effort, 15 trials/task over66tasks for Claude; GPT Codex CLI max from public board. Uses more precise section values rather than rounded headline table. Distinct from2.1; timeouts/resources changed. Comparison limit: System card section 8.6 cites the public Codex CLI result for Sol, while the Claude rows are internal Claude Code --bare reruns. Different agent harnesses and runs cannot form a matched comparison. Claude Fable 5.1 and Claude Mythos 5.1 System Card · 8.6 · reviewed 2026-11-22 | Claude Fable 5.1 and Claude Mythos 5.1 System Card | lab self-report | 2026-11-22 |
| GPT-5.6 SolOpenAI | 37.3%Reported settings & sourceOfficialpublicboard highest scoring thinking level; nativeagents may differ Gemini3.8Flash Model Card · Model card page5 Results table · reviewed 2026-09-06 | Gemini3.8Flash Model Card | lab self-report | 2026-09-06 |
| GPT-5.6 TerraOpenAI | 23.6%Reported settings & sourceOfficialpublicboard highest scoring thinking level; nativeagents may differ Gemini3.8Flash Model Card · Model card page5 Results table · reviewed 2026-09-06 | Gemini3.8Flash Model Card | lab self-report | 2026-09-06 |
| GPT-6 AstraOpenAI | 58.2%Reported settings & sourceComparator maximum available thinking/reasoning when reported, otherwise best available result; exact setting not identified by this grid. Argon self-computed; other models from the official public leaderboard. Highest scoring thinking level reported by Terminal-Bench authors. Argon agent identity is not specified in this methodology. Provider-published launch claim, not independent measurement by Google for every row. Argon self-computed; other models from the official public leaderboard. Highest scoring thinking level reported by Terminal-Bench authors. Argon agent identity is not specified in this methodology. Methodology https://deepmind.google/models/evals-methodology/gemini-4-argon, SHA256 ff1df6bdeddc4c0f48840c09a8a7813b10ae7ad9e053892da68dabf64f09bff7. Preserve source rounding. Methodology says capabilities as of September 2026 and results as of October 2026; review date is source retrieval, not evaluation date. Comparison limit: Provider-published claim. Independent source results are admitted separately with exact checkpoint, effort and compatible protocol; this grid does not establish a matched comparison. Google Gemini 4 Argon launch and evaluation methodology · Launch grid row 8: Terminal-Bench 4.0 / GPT-6 Astra; methodology Additional Details · reviewed 2026-09-30 | Google Gemini 4 Argon launch and evaluation methodology | lab self-report | 2026-09-30 |
| Mistral Large 4Mistral | 28.3%Reported settings & sourcePublic Mistral Large4 Preview. Effort, sampling and evaluation budgets are not stated. Chart footnote attributes private Artificial Analysis evaluation before public harness launch; provider-republished claim, not a new independent result. Provider-published lab_self_report; retained as launch evidence, not an independent board admission. Narrative states28.3%; chart rounds to28. Exact printed chart label; asset https://mistral.ai/_astro/code-benchmarks---terminal-bench-4%201-v3_Z2iGo8O.webp?dpl=6ac6a8353c2a68969e300fd2; asset SHA256 c3201abcd5c42890988e41c33719804fe09ff7ffe90bd0c71553e964a1f4d340. Comparison limit: Provider publication; exact benchmark-specific configuration and independent evaluation provenance require separate review before matched-board admission. Introducing Mistral Large 4 — public preview · Capabilities deep-dive; Terminal-Bench; chart 1 (mistral-chart-1.webp) · reviewed 2026-10-07 | Introducing Mistral Large 4 — public preview | lab self-report | 2026-10-07 |