Benchmark table (as of Oct 18, 2026)
Best-scoring tested variant per harness on the Artificial Analysis Coding Agent Index v1.5, an equal-weight composite of DeepSWE v1.1 (113 tasks), Terminal-Bench 4.0 (66 tasks) and SWE-Atlas-QnA (124 tasks), each the average pass@1 over three attempts.
| Harness | Model tested | Index | Terminal-Bench 4.0 | Avg API cost / task | Avg time / task |
|---|---|---|---|---|---|
| Claude Code | Sonnet 5.5 (max) | 68.4 | 66.2% | $14.19 | 87.4 min |
| Devin Fusion CLI | Claude Opus 5.5 Max + GPT-6 Luna High | 68.1 | 62.1% | $9.01 | 54.8 min |
| Muse Code | Muse Wonder Full (max) | 68.0 | 59.1% | not shown | 28.8 min |
| Antigravity CLI | Gemini 4 Argon (high) | 63.8 | 56.1% | $5.84 | 34.5 min |
| Codex | GPT-6.1 Sol (xhigh) | 62.9 | 54.5% | $1.04 | 15.5 min |
| Grok Build | Grok 4.7 (xhigh) | 56.3 | 33.3% | $8.82 | 39.2 min |
| OpenCode | GLM-5.3 (max) | 53.6 | 39.9% | $4.24 | 48.1 min |
| Kimi Code CLI | Kimi K3 | 51.9 | 21.2% | $5.05 | 61.5 min |
Source: artificialanalysis.ai/agents/coding-agents, read Oct 18, 2026 (42 variants listed). Cost is AA's pay-per-token API estimate, not subscription pricing; AA notes many people use these harnesses through plans instead.
How to read it
- The model moves the score as much as the harness. Claude Code alone ranges from 28.2 (Haiku 5.5, low) to 68.4 (Sonnet 5.5, max) in the same table.
- Effort settings trade cost for score. Claude Code with Sonnet 5.5 scores 68.4 at max effort for $14.19 a task, and 62.9 at xhigh for $3.33. Codex with GPT-6.1 Sol scores 61.4 at medium for $0.70.
- Missing isn't losing. goose, Aider, OpenHands and Gemini CLI aren't in AA's coding-agent list, so there's no number to compare.
Open or closed
| Harness | License (GitHub, Oct 18, 2026) | Runs local models? |
|---|---|---|
| OpenCode | MIT | Yes, documented (Ollama, LM Studio, llama.cpp) |
| Codex CLI | Apache-2.0 | Yes, built-in ollama / lmstudio providers |
| goose | Apache-2.0 | Yes, built-in Ollama provider |
| Aider | Apache-2.0 | Yes, ollama_chat/ models |
| Gemini CLI | Apache-2.0 | Not covered on this site yet |
| OpenHands | MIT | Not covered on this site yet |
| Claude Code | No license file detected | Not supported by Anthropic |
Picks by job
| Job | Start with | Why | Recipe |
|---|---|---|---|
| Daily coding on a Claude plan | Claude Code | Top of the index; subagents, hooks and permission modes | vs OpenCode |
| Cheapest strong runs | Codex | 61.4 index at $0.70 a task (GPT-6.1 Sol, medium) | CI job |
| Local or private models | OpenCode | Documented Ollama path, MIT | Local LLM rigs |
| Reviewable small edits | Aider | Every change is a git commit | vs OpenCode |
| Repeatable non-coding tasks | goose | Parameterized YAML recipes | Literature review |
| Long documents | Aider or Claude Code | Per-change commits; read-only editor subagent | Writing |
FAQ
What is the best agent harness in 2026?
On Artificial Analysis's Coding Agent Index (Oct 18, 2026), the top tested pair is Claude Code with Sonnet 5.5 at max effort (68.4), with Devin Fusion CLI (68.1) and Muse Code (68.0) close behind. The best harness for you depends on your models, budget and job.
What is the best open-source agent harness?
Among harnesses AA tested, OpenCode (MIT) is the open-source entry, at 53.6 with GLM-5.3. Codex CLI's code is Apache-2.0 and scored 62.9 with GPT-6.1 Sol. goose and Aider are open source but untested there.
Are these scores for the harness or the model?
Both. Every row is a harness + model pair, so a gap between rows mixes the two. Compare rows that share a model when you can.
Get new rigs by email
New job recipes and harness changes, a few times a month. Each one with copy-ready config and dated sources.
Sources
- Artificial Analysis: Coding Agent Index v1.5 · checked Oct 18, 2026
- GitHub repo metadata (license, last push), via the GitHub API · checked Oct 18, 2026
- OpenCode 1 docs: providers · checked Oct 18, 2026
- Codex docs: config reference · checked Oct 18, 2026
- goose docs: providers · checked Oct 18, 2026
- Aider docs: Ollama · checked Oct 18, 2026
- Claude Code docs: LLM gateway · checked Oct 18, 2026
Spotted something out of date? Email sternberg@mail.grokbot.com.