Task ranking
Best LLM for AI Agents
Claude Opus 5.5 ranks first for AI Agents as of October 20, 2026.[1][2][3][4][5][6] Leads Terminal-Bench 4.0 with 64.85% and ranks #1 on LMArena's agentic arena (20.4) and explicit task-outcome signal (21.4), demonstrating strong performance on reproducible terminal tasks and human-preferred multi-step tool use. Claude Sonnet 5.5 is the next-ranked alternative. Ranks #2 on LMArena Agent's explicit task-outcome signal (20.7) and #5 on the agentic arena (12.8), with a practitioner selecting it as the sole model for all agent roles in a production application.
- Best open-weight pick: DeepSeek V4.1 Flash (#6)
- Compare Claude Opus 5.5 vs Claude Sonnet 5.5 →
About this recommendation
- Updated
- Evidence through
- Sources
- 23
- Revision
- v115
Decision audit
Why this result
Inspect the inputs and the computed order behind the recommendation.
Models screened
24
live candidates
Evaluation feeds
9
task-weighted
Winner coverage
50%
intended feed weight
Largest provider share
3 of 8
Anthropic
Anthropic: Claude Opus 5.5 leads the measured task comparison with 3 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.
| Evaluation feed | Weight | Winner result | Field measured |
|---|---|---|---|
| Berkeley Function Callingstale | 0% | no current result | 0/24 |
| Terminal-Bench (4.0)Observed 2026-10-20 · evaluated 2026-10-06 · Claude Code; max effort · 95% interval ±3.11 | 32% | #1 | 14/24 |
| SWE-rebenchstale | 0% | no current result | 0/24 |
| LMArena AgentObserved 2026-10-19 · evaluation date not supplied · max effort | 21% | #1 | 24/24 |
| LiveBench Agentic Coding (2026-06-25)Observed 2026-10-18 · evaluation date not supplied · claude-opus-5-5-max-effortfailing; last successful check 2026-10-18 | 18% | #2 | 24/24 |
| OSWorld Verifiedstale | 0% | no current result | 0/24 |
| tau2-benchstale | 0% | no current result | 0/24 |
| LMArena Agent Task OutcomeObserved 2026-10-19 · evaluation date not supplied · max effort | 14% | #1 | 24/24 |
| LMArena Agent Tool HallucinationObserved 2026-10-19 · evaluation date not supplied · max effort | 14% | #10 | 24/24 |
Provider concentration
Each exact model is scored separately; provider identity is not a ranking input.
- Anthropic3 models
- OpenAI2 models
- deepseek1 model
- Moonshot AI1 model
- Z.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.
Anthropic · coverage 50%
#1 LMArena Agent · #1 LMArena Agent Task Outcome
Anthropic · coverage 50%
#2 LMArena Agent Task Outcome · #2 Terminal-Bench
Anthropic · coverage 50%
#1 LMArena Agent Tool Hallucination · #3 LiveBench Agentic Coding
OpenAI · coverage 50%
#1 LMArena Agent Tool Hallucination · #2 LMArena Agent
OpenAI · coverage 50%
#1 LMArena Agent Tool Hallucination · #3 Terminal-Bench
deepseek · coverage 34%
#1 LiveBench Agentic Coding · #13 LMArena Agent Task Outcome
Moonshot AI · coverage 34%
#1 LMArena Agent Tool Hallucination · #7 LiveBench Agentic Coding
Z.ai · coverage 34%
#1 LMArena Agent Tool Hallucination · #19 LMArena Agent Task Outcome
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.
- 01
Claude Opus 5.5Anthropic · $4 in / $20 out per M tokens · 1M context
Leads Terminal-Bench 4.0 with 64.85% and ranks #1 on LMArena's agentic arena (20.4) and explicit task-outcome signal (21.4), demonstrating strong performance on reproducible terminal tasks and human-preferred multi-step tool use.
Best when: Use for terminal-based agent workflows where Claude Code integration is available, achieving 64.85% on reproducible terminal tasks.More tips
- Deploy when minimizing tool hallucinations matters; ranks #10 with score 0.6 on LMArena's tool-hallucination signal.
- Select for executable code generation tasks where it scores 71.72% on LiveBench Agentic Coding (#2 of 59).
No practitioner reports are linked yet, so no cautions are listed. Absence of reports is not evidence of reliability.
- 02
Claude Sonnet 5.5Anthropic · $2 in / $10 out per M tokens · 1M context
Ranks #2 on LMArena Agent's explicit task-outcome signal (20.7) and #5 on the agentic arena (12.8), with a practitioner selecting it as the sole model for all agent roles in a production application.
Best when: Adopt for multi-role agent systems where a single model handles observer, tutor-editor, crop, and generator roles, as validated in production.More tips
- Use for terminal tasks where it achieves 61.82% on Terminal-Bench 4.0 (#2 of 23).
- 03
Claude Fable 5.1Anthropic · $10 in / $50 out per M tokens · 1M context
Claude Fable 5.1 ranks #1 of 48 on LMArena Agent's tool-hallucination signal (score 0.7), where a higher score means fewer fabricated tool calls.
Best when: Its strongest result for this task is on LMArena Agent tool hallucination, which measures calling only tools that exist, with real arguments.No practitioner reports are linked yet, so no cautions are listed. Absence of reports is not evidence of reliability.
- 04
GPT-6 AstraOpenAI · $10 in / $50 out per M tokens · 1.1M context
GPT-6 Astra ranks #1 of 48 on LMArena Agent's tool-hallucination signal (score 0.7), where a higher score means fewer fabricated tool calls.
Best when: Its strongest result for this task is on LMArena Agent tool hallucination, which measures calling only tools that exist, with real arguments.No practitioner reports are linked yet, so no cautions are listed. Absence of reports is not evidence of reliability.
- 05
GPT-6.1 SolOpenAI · $2 in / $10 out per M tokens · 1.1M context
GPT-6.1 Sol ranks #1 of 48 on LMArena Agent's tool-hallucination signal (score 0.7), where a higher score means fewer fabricated tool calls.
Best when: Its strongest result for this task is on LMArena Agent tool hallucination, which measures calling only tools that exist, with real arguments.No practitioner reports are linked yet, so no cautions are listed. Absence of reports is not evidence of reliability.
- 06
DeepSeek V4.1 Flashdeepseek · $0.056 in / $0.12 out per M tokens · 1.0M context · open-weight
Leads LiveBench Agentic Coding with 77.27% (#1 of 59) and offers cost-per-solve at $4.32 versus Opus 5.5's $9.05, though LMArena agentic rankings are mid-tier.
Best when: Prioritize for executable code generation where it achieves 77.27% on LiveBench Agentic Coding, outperforming all other models.More tips
- Use for cost-sensitive agent deployments where cost-per-solve is $4.32 compared to $9.05 on Opus 5.5 high effort.
Watch out for
- Expect higher tool-hallucination rates with score 0.5 (#20 of 48) versus 0.7 from top performers.
- Watch for token budget exhaustion when using reasoning_effort 'low' with tight output limits, as the model may consume the entire budget during thinking.
- Verify prompt labeling consistency, as the model may interpret messages as addressed to it when they are not, due to mismatched @-mention detection between layers.
- 07
Kimi K3Moonshot AI · $0.24 in / $11.94 out per M tokens · 1.0M context · open-weight
Kimi K3 ranks #1 of 48 on LMArena Agent's tool-hallucination signal (score 0.7), where a higher score means fewer fabricated tool calls.
Best when: Its strongest result for this task is on LMArena Agent tool hallucination, which measures calling only tools that exist, with real arguments.No practitioner reports are linked yet, so no cautions are listed. Absence of reports is not evidence of reliability.
- 08
GLM 5.2Z.ai · $0.62 in / $2.42 out per M tokens · 1.0M context · open-weight
Resolves 62.88% on SWE-rebench (#4 of 12) with tools and achieves top tool-hallucination score (0.7), though LiveBench Agentic Coding is moderate at 51.77% (#31 of 59).
Best when: Use for software engineering tasks with tool use where it achieves 62.88% on SWE-rebench.More tips
- Select when tool-hallucination resistance matters; ties for #1 with score 0.7.
Watch out for
- Expect lower executable code performance with 51.77% on LiveBench Agentic Coding (#31 of 59).
- Note weak multi-step tool-use rankings with #22 (2.1) on agentic arena and #19 (3.8) on explicit task-outcome.
Frequently asked
- What is the best model for AI Agents?
- Claude Opus 5.5 ranks first as of October 20, 2026. Leads Terminal-Bench 4.0 with 64.85% and ranks #1 on LMArena's agentic arena (20.4) and explicit task-outcome signal (21.4), demonstrating strong performance on reproducible terminal tasks and human-preferred multi-step tool use.[1][2][3]
Sources
- 1
“Within the 23 catalog models matched to this evaluation: Scores 64.85% ± 3.11 (95% interval) on Terminal-Bench 4.0 (#1 of 23) at max reasoning effort, using Claude Code, measuring a model-and-agent configuration on reproducible terminal tasks.”
Terminal-Bench · Benchmark · Oct 20, 2026 - 2
“Within the 48 catalog models matched to this evaluation: Ranks #1 of 48 on LMArena Agent's explicit task-outcome signal (score 21.4), measuring tasks users marked as completed.”
LMArena Agent task outcome · Benchmark · Oct 19, 2026 - 3
“Within the 48 catalog models matched to this evaluation: Ranks #1 of 48 on LMArena's agentic arena (score 20.4), measuring tool-use and multi-step task performance from human preference.”
LMArena agentic arena · Benchmark · Oct 19, 2026 - 4
“Within the 48 catalog models matched to this evaluation: Ranks #2 of 48 on LMArena Agent's explicit task-outcome signal (score 20.7), measuring tasks users marked as completed.”
LMArena Agent task outcome · Benchmark · Oct 19, 2026 - 5
“Within the 48 catalog models matched to this evaluation: Ranks #5 of 48 on LMArena's agentic arena (score 12.8), measuring tool-use and multi-step task performance from human preference.”
LMArena agentic arena · Benchmark · Oct 19, 2026 - 6
“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 - 7
“Within the 59 catalog models matched to this evaluation: Scores 77.27% on LiveBench Agentic Coding (#1 of 59), covering executable JavaScript, TypeScript, and Python tasks.”
LiveBench Agentic Coding · Benchmark · Oct 18, 2026 - 8
“Completed and merged into main in pull request https://github.com/Wladefant/super-board/pull/342 (commit 3c360188711ef811452a2643c9c4265c4ce5db5f). Authoritative deliverable: https://github.com/Wladefant/super-board/blob/3c360188711ef811452a2643c9c4265c4ce5db5f/docs/research/agent-benchmarks-2026-09.md Comprehensive benchmark research delivered: Artificial Analysis Terminal-Bench 4.0 analysis, DeepSeek V4.1 Flash worker tier evaluation ($4.32/solved vs $9.05 on Opus high), cost-per-solve me…”
Wladefant · GitHub · Sep 29, 2026 - 9
“Within the 48 catalog models matched to this evaluation: Ranks #10 of 48 on LMArena Agent's tool-hallucination signal (score 0.6), where a higher score means fewer fabricated tool calls.”
LMArena Agent tool hallucination · Benchmark · Oct 19, 2026 - 10
“Within the 59 catalog models matched to this evaluation: Scores 71.72% on LiveBench Agentic Coding (#2 of 59), covering executable JavaScript, TypeScript, and Python tasks.”
LiveBench Agentic Coding · Benchmark · Oct 18, 2026 - 11
“Within the 23 catalog models matched to this evaluation: Scores 61.82% ± 2.94 (95% interval) on Terminal-Bench 4.0 (#2 of 23) at max reasoning effort, using Claude Code, measuring a model-and-agent configuration on reproducible terminal tasks.”
Terminal-Bench · Benchmark · Oct 20, 2026 - 12
“Within the 48 catalog models matched to this evaluation: Ranks #1 of 48 on LMArena Agent's tool-hallucination signal (score 0.7), where a higher score means fewer fabricated tool calls.”
LMArena Agent tool hallucination · Benchmark · Oct 19, 2026 - 13
“Within the 48 catalog models matched to this evaluation: Ranks #1 of 48 on LMArena Agent's tool-hallucination signal (score 0.7), where a higher score means fewer fabricated tool calls.”
LMArena Agent tool hallucination · Benchmark · Oct 19, 2026 - 14
“Within the 48 catalog models matched to this evaluation: Ranks #1 of 48 on LMArena Agent's tool-hallucination signal (score 0.7), where a higher score means fewer fabricated tool calls.”
LMArena Agent tool hallucination · Benchmark · Oct 19, 2026 - 15
“Within the 48 catalog models matched to this evaluation: Ranks #20 of 48 on LMArena Agent's tool-hallucination signal (score 0.5), where a higher score means fewer fabricated tool calls.”
LMArena Agent tool hallucination · Benchmark · Oct 19, 2026 - 16
思考模型在较小的输出 token 预算下可能无法生成标题
Original-language excerpt:“是否存在已有相关 issue？ - [x] 我已搜索现有 issue 是否能在默认主题（日间/夜间）下重现该问题？ - [x] 我能在默认主题下重现该问题 问题是否可能是由扩展引起的？ - [x] 我已排除扩展导致问题的可能性。 问题描述 AI Agent 会在收到首条消息后生成会话标题。当前标题请求限制50tokens，但未关闭模型思考。 部分思考模型会将 completion token 预算同时用于思考和正文，可能在输出标题前耗尽预算，最终返回空正文和 finish_reason: "length"，此时会使用思考内容作为标题导致标题错乱。 响应示例： 重现步骤 使用 deepseek-v4.1-flash 测试时： - reasoning_effort: "none" 和 50 tokens 可以直接生成标题 - reasoning_effort: "low" 和 50 tokens 会耗尽预算且无法生成标题，标题为输出的思考内容 - 省略 reasoning_effort 并使用 50 tokens 同样可能耗尽预算 预期…”
fengwind2006 · GitHub · Sep 15, 2026 - 17
[Bug] 「@我」标签与门控判据不同源：@别人 / @全体成员 都被标成「点名」，模型据此误回
Original-language excerpt:“本 issue 由 DeepSeek V4.1 flash 发现并提出。 > 关联说明：与 #5（私聊无回应）、#13（初印象）均无重合。判据层与门控层的差异见文末。 现象 - 机器人在没有被艾特的运行里，把消息当成「点名」来回。 - 真实案例：群友发 @某人 顶你顶你顶你，机器人回「诶？顶我干嘛啦」，会话思考里写着「@了我和某人」。 - 同一次运行的 contextReason 是 随机命中(1%) —— 门控认为没被艾特，提示词里却标了「（@我）」。 - 触发一次之后不会自愈：只要群里有人用 @ 开头说话，就继续误判。 期望行为 「@我」标签应与门控使用同一套判据（isAtMe：@群名片 / @昵称 / @机器人自己的 QQ 号），只有真的叫机器人才标。 实际行为 标签层与门控层是两套判据：门控说「没被艾特」（本次由随机命中触发），提示词里却写着「（@我）」。模型看到自相矛盾的上下文，就按「被点名」来回话。 10 秒复现（直接跑仓库自带代码，零改动） 实测输出（在上游 HEAD 72f537f 上原样运行，未改动…”
pomelo-neko · GitHub · Sep 20, 2026 - 18
“Within the 48 catalog models matched to this evaluation: Ranks #1 of 48 on LMArena Agent's tool-hallucination signal (score 0.7), where a higher score means fewer fabricated tool calls.”
LMArena Agent tool hallucination · Benchmark · Oct 19, 2026 - 19
“Within the 12 catalog models matched to this evaluation: Resolves 62.88% ± 1.19 standard error on SWE-rebench (#4 of 12) using tools, comparing the same task window (2026-05-15 to 2026-07-01) and agent setup across models.”
SWE-rebench · Benchmark · Oct 20, 2026 - 20
“Within the 48 catalog models matched to this evaluation: Ranks #1 of 48 on LMArena Agent's tool-hallucination signal (score 0.7), where a higher score means fewer fabricated tool calls.”
LMArena Agent tool hallucination · Benchmark · Oct 19, 2026 - 21
“Within the 59 catalog models matched to this evaluation: Scores 51.77% on LiveBench Agentic Coding (#31 of 59), covering executable JavaScript, TypeScript, and Python tasks.”
LiveBench Agentic Coding · Benchmark · Oct 18, 2026 - 22
“Within the 48 catalog models matched to this evaluation: Ranks #22 of 48 on LMArena's agentic arena (score 2.1), measuring tool-use and multi-step task performance from human preference.”
LMArena agentic arena · Benchmark · Oct 19, 2026 - 23
“Within the 48 catalog models matched to this evaluation: Ranks #19 of 48 on LMArena Agent's explicit task-outcome signal (score 3.8), measuring tasks users marked as completed.”
LMArena Agent task outcome · Benchmark · Oct 19, 2026
Rankings synthesized from community evidence and open benchmarks. See methodology. Not driven by vendor marketing.