Agent routing benchmark

Multi-agent tasks: which coding agent

Four coding-agent CLIs ranked for five kinds of task, from public leaderboards and the outcomes of our own jobs. Refreshed weekly.

Generated 2026-10-13
Weekly refresh
5 task classes
6 kinds of coding and agentic work
4 agent CLIs

For agents: /bench.md (compact) · /bench.json (full) · /llms.txt

Which agent to use, by task class · score 0–1 · evidence = own jobs in the last 28 days
Task class	First choice	Score	Runner-up	Evidence
Codingrepo-level code changes	Claude CodeClaude Opus 5.5	0.96	Claude CodeClaude Fable 5.1 0.80	Proven · 15 jobs
Agenticterminal and multi-step tool work	Claude CodeClaude Opus 5.5	0.75	Claude CodeClaude Sonnet 5.5 0.72	Unproven
Researchweb research and fact-finding	Codex CLIGPT-5.6 Sol	0.75	Claude CodeClaude Sonnet 5.5 0.62	Unproven
Readinglong documents, transcripts, summaries, extraction	Claude CodeClaude Fable 5.1	0.75	Claude CodeClaude Sonnet 5.5 0.60	Unproven
Writingdrafting and editing prose	Claude CodeClaude Sonnet 5.5	0.64	Claude CodeClaude Opus 5.5 0.65	Proven · 2 jobs

Proven ranked with own jobs from the last 28 days
Unproven public scores only, no own jobs yet

Full ranking by task class

Top 8 per class. Score is the priority (0–1); rows tagged tie or evidence were reordered by the rules in Method. Coding and agentic also rank each kind of work on its own.

Coding
repo-level code changesswe_bench_verified_resolvedswe_bench_multilingual_resolvedterminal_bench_4_accuracyaa_coding
#	Engine	Model	Evidence	Score	Public	Own success	Jobs	Median min	Speed
1	Claude Code	Claude Opus 5.5	Proven	0.96	1.00	1.00	15	1.9	0.80
2	Claude Code	Claude Fable 5.1	Proven tie	0.80	0.87	0.96	73	2.9	0.40
3	Claude Code	Claude Sonnet 5.5	Proven	0.79	0.94	0.92	48	4.0	0.20
4	Codex CLI	GPT-6 Sol	Proven	0.70	0.70	0.50 default	1	1.8	1.00
5	Codex CLI	GPT-6.1 Sol	Proven tie	0.66	0.87	0.74	19	10.2	0.00
6	Antigravity CLI	Claude Opus 5.5	Unproven tie	0.65	0.80 ×0.8	0.50 default	0	–	0.50
7	Antigravity CLI	Claude Sonnet 5.5	Unproven	0.63	0.75 ×0.8	0.50 default	0	–	0.50
8	Codex CLI	GPT-6 Astra	Proven tie	0.52	0.50 none	0.50 default	4	2.1	0.60
	8 more candidates ranked in /bench.json

Coding by kind of work · first choice and runner-up; full lists in /bench.json
Kind of work	First choice	Score	Runner-up	Evidence
Bug fixingfinding and fixing defects in existing codeswe_bench_verified_resolved	Claude CodeClaude Sonnet 5.5	0.65	Antigravity CLIGPT-OSS 120B 0.50	Proven · 13 jobs
Features and refactorsnew features, tools and restructuringarena_text_coding	Claude CodeClaude Sonnet 5.5	0.90	Claude CodeClaude Opus 5.5 0.65	Proven · 17 jobs
Frontend and UIweb pages, UI components and stylingarena_webdev_overall	Antigravity CLIClaude Opus 5.5	0.75	Claude CodeClaude Opus 5.5 0.75	Unproven
Tests and reviewcode review, audits and testsown jobs only	Claude CodeClaude Opus 5.5	0.75	Claude CodeClaude Fable 5.1 0.58	Proven · 13 jobs

Agentic
terminal and multi-step tool workterminal_bench_4_accuracyaa_agentic
#	Engine	Model	Evidence	Score	Public	Own success	Jobs	Median min	Speed
1	Claude Code	Claude Opus 5.5	Unproven tie	0.75	1.00	0.50 default	0	–	0.50
2	Claude Code	Claude Sonnet 5.5	Unproven	0.72	0.94	0.50 default	0	–	0.50
3	Codex CLI	GPT-6.1 Sol	Unproven tie	0.69	0.87	0.50 default	0	–	0.50
4	Claude Code	Claude Fable 5.1	Unproven	0.68	0.87	0.50 default	0	–	0.50
5	Antigravity CLI	Claude Opus 5.5	Unproven tie	0.65	0.80 ×0.8	0.50 default	0	–	0.50
6	Antigravity CLI	Claude Sonnet 5.5	Unproven tie	0.63	0.75 ×0.8	0.50 default	0	–	0.50
7	Codex CLI	GPT-6 Sol	Unproven	0.60	0.70	0.50 default	0	–	0.50
8	Codex CLI	GPT-5.6 Sol	Unproven	0.49	0.47	0.50 default	0	–	0.50
	6 more candidates ranked in /bench.json

Agentic by kind of work · first choice and runner-up; full lists in /bench.json
Kind of work	First choice	Score	Runner-up	Evidence
Terminal and sysadminshell, installs, configuration, server upkeepterminal_bench_4_accuracy arena_agent_bash_reliability_no_prereq_overall arena_agent_bash_recovery_steps_overall	Claude CodeClaude Opus 5.5	0.73	Claude CodeClaude Sonnet 5.5 0.72	Unproven
Tools and connectorsmulti-step work through APIs and tool connectorsarena_agent_tool_hallucination_overall arena_agent_task_outcome_explicit_overall	Codex CLIGPT-6.1 Sol	0.71	Antigravity CLIClaude Opus 5.5 0.66	Unproven

Research
web research and fact-findingdrb_total_scorearena_search_overall
#	Engine	Model	Evidence	Score	Public	Own success	Jobs	Median min	Speed
1	Codex CLI	GPT-5.6 Sol	Unproven	0.75	1.00	0.50 default	0	–	0.50
2	Claude Code	Claude Sonnet 5.5	Proven tie	0.62	0.50 none	0.90	19	5.0	0.50
3	Claude Code	Claude Haiku 5.5	Proven	0.60	0.50 none	0.50 default	2	0.4	1.00
4	Claude Code	Claude Fable 5.1	Proven	0.55	0.50 none	0.50 default	7	1.4	0.75
5	Codex CLI	GPT-6.1 Sol	Proven tie	0.42	0.50 none	0.57	28	13.7	0.00
6	Codex CLI	GPT-6 Sol	Proven	0.45	0.50 none	0.50 default	1	6.9	0.25
7	Grok CLI	Grok 4.5	Unproven	0.25	0.00	0.50 default	0	–	0.50

Reading
long documents, transcripts, summaries, extractionarena_document_overall
#	Engine	Model	Evidence	Score	Public	Own success	Jobs	Median min	Speed
1	Claude Code	Claude Fable 5.1	Unproven	0.75	1.00	0.50 default	0	–	0.50
2	Claude Code	Claude Sonnet 5.5	Proven	0.60	0.50 none	0.50 default	5	1.3	1.00
3	Codex CLI	GPT-6.1 Sol	Proven tie	0.49	0.50 none	0.80	15	9.1	0.00
4	Codex CLI	GPT-5.6 Sol	Proven	0.50	0.51	0.50 default	2	5.8	0.50
5	Codex CLI	GPT-5.6 Terra	Unproven	0.41	0.32	0.50 default	0	–	0.50
6	Codex CLI	GPT-6 Astra	Unproven	0.38	0.26	0.50 default	0	–	0.50
7	Codex CLI	GPT-5.6 Luna	Unproven	0.29	0.08	0.50 default	0	–	0.50
8	Grok CLI	Grok 4.6	Unproven tie	0.25	0.01	0.50 default	0	–	0.50
	1 more candidate ranked in /bench.json

Writing
drafting and editing prosearena_text_creative_writingarena_text_overall
#	Engine	Model	Evidence	Score	Public	Own success	Jobs	Median min	Speed
1	Claude Code	Claude Sonnet 5.5	Proven evidence	0.64	0.72	0.50 default	2	1.3	0.67
2	Claude Code	Claude Opus 5.5	Proven evidence	0.65	1.00	0.50 default	2	2.4	0.00
3	Antigravity CLI	Claude Opus 5.5	Unproven	0.75	1.00	0.50 default	0	–	0.50
4	Claude Code	Claude Fable 5.1	Unproven	0.72	0.93	0.50 default	0	–	0.50
5	Antigravity CLI	Gemini 3.8 Flash	Unproven	0.68	0.87	0.50 default	0	–	0.50
6	Claude Code	Claude Haiku 5.5	Proven tie	0.60	0.50 none	0.50 default	1	0.0	1.00
7	Antigravity CLI	Claude Sonnet 5.5	Unproven	0.61	0.72	0.50 default	0	–	0.50
8	Codex CLI	GPT-5.6 Sol	Unproven tie	0.57	0.64	0.50 default	0	–	0.50
	10 more candidates ranked in /bench.json

Method

priority = 0.5*public_norm + 0.3*own_success + 0.2*speed_rank: weights public 0.5, own success 0.3, speed 0.2; every input is 0–1.
Public: each class metric is min-max scaled across the engine/model pairs ranked here, then averaged; 0.5 when a model has none.
Harness discount: a score measured in another agent's harness counts at 0.8× (marked ×0.8), for example a Terminal-Bench run in Claude Code credited to Antigravity running the same model. The engine's own harness, or a model-only benchmark, counts in full.
Own success: share of own jobs in the last 28 days that finished; 0.5 with fewer than 10 jobs. Speed: rank by own median minutes, else Artificial Analysis output speed, else 0.5.
Evidence first: scores within 0.03 tie (marked tie) and go to more own jobs, then speed. An unproven entry ranks above a proven one only on a clear public margin: more than 0.03 ahead and a higher public score (else marked evidence).
Kinds of work: Coding and agentic subclasses are ranked the same way, on the public metrics that measure that kind of work and own jobs of that kind; a subclass with neither is not listed.
Limits: Plan limits are checked live when a job is dispatched; this table does not track them.
Only aggregate job counts, success rates and median minutes are published. Model lists for Grok CLI are unverified.

Full ranking by task class

Coding

Agentic

Research

Reading

Writing

Method