Part of Best LLM for AI Agents

Best LLM for Autonomous Agents

Claude Fable 5.2 ranks first for Autonomous Agents as of November 15, 2026.[1][2][3][4][5] Leads Terminal-Bench 4.1 with 65.76% accuracy and holds #1 rankings across LMArena's agentic arena (30.4), explicit task-outcome signal (22), and Design Arena's full-stack agent category (Elo 1400), demonstrating consistent top-tier performance in multi-step tool use and terminal task execution. Claude Opus 5.5 is the next-ranked alternative. It scores 65.76% ± 3.11 (95% interval) on Terminal-Bench 4.1 (#1 of 46) at max reasoning effort, using Claude Code, measuring a model-and-agent configuration on reproducible terminal tasks.

About this recommendation

Updated
Evidence through
Sources
17
Revision
v140

Decision audit

Why this result

Inspect the inputs and the computed order behind the recommendation.

Models screened

24

live candidates

Evaluation feeds

6

task-weighted

Winner coverage

54%

intended feed weight

Largest provider share

3 of 8

Anthropic

Established source breadth. 5 citation families and 0 practitioner families provide attributed capability reports; 0 cautionary threads retained. The largest citation family contributes 39%.

Anthropic: Claude Fable 5.2 leads the measured task comparison with 2 independent evaluation families. Community volume, release age and usage popularity add no ranking points.

Real-world track record is still developing. The recommendation can lead on independent measurements while practitioner reports accumulate.

Sources evaluated

These are the weights used in scoring. Missing measurements are omitted from each model’s average. A feed with no result for any model released in the last 60 days is stale and does not score until it catches up.

winner: Claude Fable 5.2
Evaluation feedWeightWinner resultField measured
Terminal-Bench (4.1)Observed 2026-11-15 · evaluated 2026-11-13 · Claude Code; max effort · 95% interval ±2.85
47%
#120/24
SWE-rebenchstale
0%
no current result0/24
LiveBench Agentic Codingfailing; last successful check 2026-10-18
20%
not measured14/24
LMArena Agent Task OutcomeObserved 2026-11-11 · evaluation date not supplied · max effort
20%
#124/24
LMArena AgentObserved 2026-11-11 · evaluation date not supplied · max effort
13%
#124/24
OSWorld Verifiedstale
0%
no current result0/24

Provider concentration

Each exact model is scored separately; provider identity is not a ranking input.

Anthropic38%
  • Anthropic3 models
  • Google2 models
  • OpenAI2 models
  • Moonshot AI1 model

Decision table

Every published model is shown in computed order. Practitioner sources are distinct community threads, not the citations repeated in the prose below.

  1. Anthropic · coverage 54%

    #1 LMArena Agent · #1 LMArena Agent Task Outcome

  2. Anthropic · coverage 67%

    #1 Terminal-Bench · #2 LiveBench Agentic Coding

  3. Google · coverage 54%

    #3 LMArena Agent · #3 Terminal-Bench

  4. Anthropic · coverage 67%

    #3 LMArena Agent Task Outcome · #4 LMArena Agent

  5. OpenAI · coverage 67%

    #6 LMArena Agent · #6 LMArena Agent Task Outcome

  6. Google · coverage 54%

    #7 LMArena Agent Task Outcome · #9 LMArena Agent

  7. OpenAI · coverage 67%

    #5 Terminal-Bench · #7 LMArena Agent

  8. Moonshot AI · coverage 54%

    #7 Terminal-Bench · #14 LMArena Agent

Task score averages the available, task-weighted measurements. Benchmark positions compare models matched to our catalog. Models measured by at least two independent evaluator families are considered first. Missing tests, practitioner report counts, popularity and release age add no points. Coverage describes evidence breadth, not a probability of correctness. Scores within two points are shown as close under this policy; this is not a statistical confidence interval. When the leading scores are that close, the model measured by more evaluator families ranks first.

  1. 01

    Claude Fable 5.2Anthropic · $8 in / $40 out per M tokens · 1M context

    Leads Terminal-Bench 4.1 with 65.76% accuracy and holds #1 rankings across LMArena's agentic arena (30.4), explicit task-outcome signal (22), and Design Arena's full-stack agent category (Elo 1400), demonstrating consistent top-tier performance in multi-step tool use and terminal task execution.

    Best when: Use for terminal-based autonomous workflows where reproducible task execution matters, as it scores 65.76% on Terminal-Bench 4.1 (#1 of 46) with Claude Code.
    Source 1
    “Within the 46 catalog models matched to this evaluation: Scores 65.76% ± 2.85 (95% interval) on Terminal-Bench 4.1 (#1 of 46) at max reasoning effort, using Claude Code, measuring a model-and-agent configuration on reproducible terminal tasks.”
    Terminal-BenchOpen original ↗

    More tips

    • Deploy for full-stack agent development where human preference for final output quality is critical, ranking #1 of 44 on Design Arena (Elo 1400).
      Source 2
      “Within the 44 catalog models matched to this evaluation: Ranks #1 of 44 on Design Arena's full-stack agent category (Elo 1400), based on blind human preference between agent-built results.”
      Design Arena full-stack agentsOpen original ↗
    • Select when explicit task completion rates are the priority, achieving #1 of 52 on LMArena Agent's task-outcome signal (score 22).
      Source 4
      “Within the 52 catalog models matched to this evaluation: Ranks #1 of 52 on LMArena Agent's explicit task-outcome signal (score 22), measuring tasks users marked as completed.”
      LMArena Agent task outcomeOpen original ↗

    No practitioner reports are linked yet, so no cautions are listed. Absence of reports is not evidence of reliability.

  2. 02

    Claude Opus 5.5Anthropic · $4 in / $20 out per M tokens · 1M context

    Claude Opus 5.5 scores 65.76% ± 3.11 (95% interval) on Terminal-Bench 4.1 (#1 of 46) at max reasoning effort, using Claude Code, measuring a model-and-agent configuration on reproducible terminal tasks.

    Best when: Its strongest result for this task is on Terminal-Bench, which measures completing difficult repository and terminal tasks through an agent harness.
    Source 5
    “Within the 46 catalog models matched to this evaluation: Scores 65.76% ± 3.11 (95% interval) on Terminal-Bench 4.1 (#1 of 46) at max reasoning effort, using Claude Code, measuring a model-and-agent configuration on reproducible terminal tasks.”
    Terminal-BenchOpen original ↗
  3. 03

    Gemini 4 CarbonGoogle · $2 in / $10 out per M tokens · 1.0M context

    Gemini 4 Carbon scores 64.55% ± 3.23 (95% interval) on Terminal-Bench 4.1 (#3 of 46) at high reasoning effort, using Antigravity CLI, measuring a model-and-agent configuration on reproducible terminal tasks.

    Best when: Its strongest result for this task is on Terminal-Bench, which measures completing difficult repository and terminal tasks through an agent harness.
    Source 11
    “Within the 46 catalog models matched to this evaluation: Scores 64.55% ± 3.23 (95% interval) on Terminal-Bench 4.1 (#3 of 46) at high reasoning effort, using Antigravity CLI, measuring a model-and-agent configuration on reproducible terminal tasks.”
    Terminal-BenchOpen original ↗
  4. 04

    Claude Sonnet 5.5Anthropic · $2 in / $10 out per M tokens · 1M context

    Designated as sole model for autonomous agent functions in production deployments, ranking #3 on LMArena Agent's explicit task-outcome signal (19.5) and #4 on Terminal-Bench 4.1 (63.94%), though showing lower LiveBench Agentic Coding performance (#21 of 59, 56.31%).

    Best when: Use as a unified agent model across all roles (observer, tutor-editor, crop, generators) where organizations have standardized on a single model for consistency.
    Source 12
    “Human decision (2026-09-29): the app uses ONLY Claude Sonnet 5.5 (claude-sonnet-5-5) for every role (observer, tutor-editor, crop, generators), replacing Opus 5.5 and Sonnet 5. Update AGENTS.md, docs/VISION.md and docs/PLAN.md accordingly (role names unchanged). Do not touch dated history in docs/runbooks/operations.md.”

    More tips

    • Select when explicit task completion signal matters, ranking #3 of 52 on LMArena Agent's task-outcome metric (score 19.5).
      Source 13
      “Within the 52 catalog models matched to this evaluation: Ranks #3 of 52 on LMArena Agent's explicit task-outcome signal (score 19.5), measuring tasks users marked as completed.”
      LMArena Agent task outcomeOpen original ↗

    Watch out for

    • Expect lower performance on executable coding tasks, scoring 56.31% on LiveBench Agentic Coding (#21 of 59) for JavaScript, TypeScript, and Python.
      Source 14
      “Within the 59 catalog models matched to this evaluation: Scores 56.31% on LiveBench Agentic Coding (#21 of 59), covering executable JavaScript, TypeScript, and Python tasks.”
      LiveBench Agentic CodingOpen original ↗
  5. 05

    GPT-6 AstraOpenAI · $10 in / $50 out per M tokens · 1.1M context

    GPT-6 Astra scores 60.3% ± 2.82 (95% interval) on Terminal-Bench 4.1 (#6 of 46) at max reasoning effort, using Codex, measuring a model-and-agent configuration on reproducible terminal tasks.

    Best when: Its strongest result for this task is on Terminal-Bench, which measures completing difficult repository and terminal tasks through an agent harness.
    Source 15
    “Within the 46 catalog models matched to this evaluation: Scores 60.3% ± 2.82 (95% interval) on Terminal-Bench 4.1 (#6 of 46) at max reasoning effort, using Codex, measuring a model-and-agent configuration on reproducible terminal tasks.”
    Terminal-BenchOpen original ↗
  6. 06

    Gemini 4 ArgonGoogle · $2 in / $10 out per M tokens · 1.0M context

    Gemini 4 Argon ranks #7 of 52 on LMArena Agent's explicit task-outcome signal (score 13.2), measuring tasks users marked as completed.

    Best when: Its strongest result for this task is on LMArena Agent task outcome, which measures finishing the task the user actually set.
    Source 16
    “Within the 52 catalog models matched to this evaluation: Ranks #7 of 52 on LMArena Agent's explicit task-outcome signal (score 13.2), measuring tasks users marked as completed.”
    LMArena Agent task outcomeOpen original ↗

    No practitioner reports are linked yet, so no cautions are listed. Absence of reports is not evidence of reliability.

  7. 07

    GPT-6.1 SolOpenAI · $2 in / $10 out per M tokens · 1.1M context

    GPT-6.1 Sol scores 60.91% ± 2.97 (95% interval) on Terminal-Bench 4.1 (#5 of 46) at max reasoning effort, using Codex, measuring a model-and-agent configuration on reproducible terminal tasks.

    Best when: Its strongest result for this task is on Terminal-Bench, which measures completing difficult repository and terminal tasks through an agent harness.
    Source 17
    “Within the 46 catalog models matched to this evaluation: Scores 60.91% ± 2.97 (95% interval) on Terminal-Bench 4.1 (#5 of 46) at max reasoning effort, using Codex, measuring a model-and-agent configuration on reproducible terminal tasks.”
    Terminal-BenchOpen original ↗

    No practitioner reports are linked yet, so no cautions are listed. Absence of reports is not evidence of reliability.

  8. 08

    Kimi K3.1 ProMoonshot AI · $0.24 in / $8.40 out per M tokens · 1.0M context · open-weight

    Strong open-weight performer ranking #3 on Design Arena's full-stack (Elo 1321) and web-app (Elo 1312) agent categories.

    Best when: Use for full-stack agent development where open-weight deployment is required, ranking #3 of 44 on Design Arena (Elo 1321) by blind human preference.
    Source 6
    “Within the 44 catalog models matched to this evaluation: Ranks #3 of 44 on Design Arena's full-stack agent category (Elo 1321), based on blind human preference between agent-built results.”
    Design Arena full-stack agentsOpen original ↗

    More tips

    • Deploy for web-app agent tasks with open-weight constraints, ranking #3 of 47 on Design Arena (Elo 1312).
      Source 7
      “Within the 47 catalog models matched to this evaluation: Ranks #3 of 47 on Design Arena's web-app agent category (Elo 1312), based on blind human preference between agent-built results.”
      Design Arena web-app agentsOpen original ↗

    Watch out for

    • Expect poor explicit task completion rates, ranking #23 of 52 on LMArena Agent's task-outcome signal (score 2.7).
      Source 9
      “Within the 52 catalog models matched to this evaluation: Ranks #23 of 52 on LMArena Agent's explicit task-outcome signal (score 2.7), measuring tasks users marked as completed.”
      LMArena Agent task outcomeOpen original ↗

Frequently asked

What is the best model for Autonomous Agents?
Claude Fable 5.2 ranks first as of November 15, 2026. Leads Terminal-Bench 4.1 with 65.76% accuracy and holds #1 rankings across LMArena's agentic arena (30.4), explicit task-outcome signal (22), and Design Arena's full-stack agent category (Elo 1400), demonstrating consistent top-tier performance in multi-step tool use and terminal task execution.[1][2][3][4]
What is the best alternative to Claude Fable 5.2?
Claude Opus 5.5 ranks second. Claude Opus 5.5 scores 65.76% ± 3.11 (95% interval) on Terminal-Bench 4.1 (#1 of 46) at max reasoning effort, using Claude Code, measuring a model-and-agent configuration on reproducible terminal tasks.[5]
Which open-weight model ranks highest for Autonomous Agents?
Kimi K3.1 Pro is the highest-ranked open-weight model, at #8. Strong open-weight performer ranking #3 on Design Arena's full-stack (Elo 1321) and web-app (Elo 1312) agent categories.[6][7][8][9][10]

Sources

  1. 1

    “Within the 46 catalog models matched to this evaluation: Scores 65.76% ± 2.85 (95% interval) on Terminal-Bench 4.1 (#1 of 46) at max reasoning effort, using Claude Code, measuring a model-and-agent configuration on reproducible terminal tasks.”

    Terminal-Bench · Benchmark · Nov 15, 2026
  2. 2

    “Within the 44 catalog models matched to this evaluation: Ranks #1 of 44 on Design Arena's full-stack agent category (Elo 1400), based on blind human preference between agent-built results.”

    Design Arena full-stack agents · Benchmark · Nov 15, 2026
  3. 3

    “Within the 52 catalog models matched to this evaluation: Ranks #1 of 52 on LMArena's agentic arena (score 30.4), measuring tool-use and multi-step task performance from human preference.”

    LMArena agentic arena · Benchmark · Nov 11, 2026
  4. 4

    “Within the 52 catalog models matched to this evaluation: Ranks #1 of 52 on LMArena Agent's explicit task-outcome signal (score 22), measuring tasks users marked as completed.”

    LMArena Agent task outcome · Benchmark · Nov 11, 2026
  5. 5

    “Within the 46 catalog models matched to this evaluation: Scores 65.76% ± 3.11 (95% interval) on Terminal-Bench 4.1 (#1 of 46) at max reasoning effort, using Claude Code, measuring a model-and-agent configuration on reproducible terminal tasks.”

    Terminal-Bench · Benchmark · Nov 15, 2026
  6. 6

    “Within the 44 catalog models matched to this evaluation: Ranks #3 of 44 on Design Arena's full-stack agent category (Elo 1321), based on blind human preference between agent-built results.”

    Design Arena full-stack agents · Benchmark · Nov 15, 2026
  7. 7

    “Within the 47 catalog models matched to this evaluation: Ranks #3 of 47 on Design Arena's web-app agent category (Elo 1312), based on blind human preference between agent-built results.”

    Design Arena web-app agents · Benchmark · Nov 15, 2026
  8. 8

    “Within the 46 catalog models matched to this evaluation: Scores 59.7% ± 4.18 (95% interval) on Terminal-Bench 4.1 (#7 of 46) at max reasoning effort, using Claude Code, measuring a model-and-agent configuration on reproducible terminal tasks.”

    Terminal-Bench · Benchmark · Nov 15, 2026
  9. 9

    “Within the 52 catalog models matched to this evaluation: Ranks #23 of 52 on LMArena Agent's explicit task-outcome signal (score 2.7), measuring tasks users marked as completed.”

    LMArena Agent task outcome · Benchmark · Nov 11, 2026
  10. 10

    “Within the 52 catalog models matched to this evaluation: Ranks #14 of 52 on LMArena's agentic arena (score 4.9), measuring tool-use and multi-step task performance from human preference.”

    LMArena agentic arena · Benchmark · Nov 11, 2026
  11. 11

    “Within the 46 catalog models matched to this evaluation: Scores 64.55% ± 3.23 (95% interval) on Terminal-Bench 4.1 (#3 of 46) at high reasoning effort, using Antigravity CLI, measuring a model-and-agent configuration on reproducible terminal tasks.”

    Terminal-Bench · Benchmark · Nov 15, 2026
  12. 12

    “Human decision (2026-09-29): the app uses ONLY Claude Sonnet 5.5 (claude-sonnet-5-5) for every role (observer, tutor-editor, crop, generators), replacing Opus 5.5 and Sonnet 5. Update AGENTS.md, docs/VISION.md and docs/PLAN.md accordingly (role names unchanged). Do not touch dated history in docs/runbooks/operations.md.”

    MatillaM · GitHub · Sep 29, 2026
  13. 13

    “Within the 52 catalog models matched to this evaluation: Ranks #3 of 52 on LMArena Agent's explicit task-outcome signal (score 19.5), measuring tasks users marked as completed.”

    LMArena Agent task outcome · Benchmark · Nov 11, 2026
  14. 14

    “Within the 59 catalog models matched to this evaluation: Scores 56.31% on LiveBench Agentic Coding (#21 of 59), covering executable JavaScript, TypeScript, and Python tasks.”

    LiveBench Agentic Coding · Benchmark · Oct 18, 2026
  15. 15

    “Within the 46 catalog models matched to this evaluation: Scores 60.3% ± 2.82 (95% interval) on Terminal-Bench 4.1 (#6 of 46) at max reasoning effort, using Codex, measuring a model-and-agent configuration on reproducible terminal tasks.”

    Terminal-Bench · Benchmark · Nov 15, 2026
  16. 16

    “Within the 52 catalog models matched to this evaluation: Ranks #7 of 52 on LMArena Agent's explicit task-outcome signal (score 13.2), measuring tasks users marked as completed.”

    LMArena Agent task outcome · Benchmark · Nov 11, 2026
  17. 17

    “Within the 46 catalog models matched to this evaluation: Scores 60.91% ± 2.97 (95% interval) on Terminal-Bench 4.1 (#5 of 46) at max reasoning effort, using Codex, measuring a model-and-agent configuration on reproducible terminal tasks.”

    Terminal-Bench · Benchmark · Nov 15, 2026

Rankings synthesized from community evidence and open benchmarks. See methodology. Not driven by vendor marketing.