AI Coding Harnesses
The same model can score 15 points apart on the same benchmark depending on which agent harness wraps it. This page tracks how the major coding harnesses (Claude Code, Cursor, Codex CLI, Aider, OpenHands, Devin, Cline, Windsurf, Amp, Continue, Roo Code) perform across SWE-bench Verified, Terminal-Bench, Aider Polyglot, and SWE-Lancer. Last updated 2026-10-28.
Machine-readable JSON/api/harnessesMost of the AI coding conversation in 2026 is about harnesses, not models. Claude Sonnet 4.6 in Claude Code scores ~71% on SWE-bench Verified. The same Sonnet 4.6 in Continue scores ~52%. The model is identical. The harness is doing the work: tool-use loop, retrieval, planning, the order it reads files in, when it decides to stop and run tests, how it backs off after a failed edit. The harness gap is real and it is the load-bearing thing in most production agent setups.
The matrix below collects the best vendor-published score for each harness × base-model combination across 7 benchmarks. Tabs above the table switch which benchmark drives the ranked leaderboard view. The full matrix is below that, and each harness name links to a detail page with the harness architecture, model story, and pricing model.
Live agentic-coding scores for the major harnesses come from the TerminalFeed federation board, with a per-result source_url to the upstream benchmark. Harnesses the board does not cover keep TensorFeed editorial scores. SWE-Lancer is no longer published by the federation board.
Terminal-Bench 4.0: The current Terminal-Bench revision, hosted by Harbor and the Laude Institute. Scores sit far below the 2.1 column because the tasks are harder, so the two are separate measurements and must not be compared across columns.
Scoring unit: % solved. Max: 100.
Terminal-Bench 4.0 Leaderboard
| Rank | Harness | Base Model | Vendor | Type | Score |
|---|---|---|---|---|---|
| #1 | Claude Code | Claude Opus 5.5 | Anthropic | cli | 64.8/ 100 |
| #2 | Claude Code | Claude Sonnet 5.5 | Anthropic | cli | 61.8/ 100 |
| #3 | Codex CLIOSS | GPT-6 Astra | OpenAI | cli | 58.2/ 100 |
| #4 | Codex CLIOSS | GPT-6.1 Sol | OpenAI | cli | 58.2/ 100 |
| #5 | Claude Code | Claude Fable 5.1 | Anthropic | cli | 57.9/ 100 |
| #6 | Claude Code | Claude Opus 5 | Anthropic | cli | 53.9/ 100 |
| #7 | Codex CLIOSS | GPT-6 Sol | OpenAI | cli | 49.4/ 100 |
| #8 | Claude Code | Claude Fable 5 | Anthropic | cli | 44.5/ 100 |
| #9 | Claude Code | GLM-5.3 | Anthropic | cli | 41.8/ 100 |
| #10 | Grok BuildOSS | Grok 4.7 | xAI | cli | 37.6/ 100 |
| #11 | Codex CLIOSS | GPT-5.6 Sol | OpenAI | cli | 37.3/ 100 |
| #12 | Claude Code | GLM-5.3 Flash | Anthropic | cli | 35.8/ 100 |
| #13 | Claude Code | Qwen3.8-Max-0902 | Anthropic | cli | 27.0/ 100 |
| #14 | Claude Code | Claude Opus 4.8 | Anthropic | cli | 23.6/ 100 |
| #15 | Codex CLIOSS | GPT-5.6 Terra | OpenAI | cli | 21.5/ 100 |
| #16 | Grok BuildOSS | Grok 4.6 | xAI | cli | 20.3/ 100 |
| #17 | Codex CLIOSS | GPT-5.6 Luna | OpenAI | cli | 17.3/ 100 |
| #18 | Codex CLIOSS | GPT-6 Luna | OpenAI | cli | 16.4/ 100 |
| #19 | Muse Code | Muse Spark 1.3 | Meta | cli | 14.5/ 100 |
| #20 | Grok BuildOSS | Grok 4.5 | xAI | cli | 12.4/ 100 |
| #21 | Claude Code | Claude Sonnet 5 | Anthropic | cli | 12.4/ 100 |
Full Matrix
Every harness × base-model combination across every tracked benchmark. Empty cells mean the vendor has not published a score on that benchmark for that model in that harness.
| Harness | Base Model | SWE-bench Verified | Terminal-Bench | Terminal-Bench 2.1 | Terminal-Bench 3.0 | Terminal-Bench 4.0 | Aider Polyglot | SWE-Lancer |
|---|---|---|---|---|---|---|---|---|
| Claude Code | Claude Opus 5.5 | - | - | - | - | 64.8 | - | - |
| Claude Code | Claude Sonnet 5.5 | - | - | - | - | 61.8 | - | - |
| Codex CLI | GPT-6 Astra | - | - | 87.4 | - | 58.2 | - | - |
| Codex CLI | GPT-6.1 Sol | - | - | - | - | 58.2 | - | - |
| Claude Code | Claude Fable 5.1 | - | - | - | - | 57.9 | - | - |
| Claude Code | Claude Opus 5 | - | - | - | - | 53.9 | - | - |
| Codex CLI | GPT-6 Sol | - | - | - | - | 49.4 | - | - |
| Claude Code | Claude Fable 5 | - | - | 83.8 | 34.0 | 44.5 | - | - |
| Claude Code | GLM-5.3 | - | - | - | 32.4 | 41.8 | - | - |
| Grok Build | Grok 4.7 | - | - | - | - | 37.6 | - | - |
| Codex CLI | GPT-5.6 Sol | - | - | - | 34.6 | 37.3 | - | - |
| Claude Code | GLM-5.3 Flash | - | - | - | - | 35.8 | - | - |
| Claude Code | Qwen3.8-Max-0902 | - | - | - | - | 27.0 | - | - |
| Claude Code | Claude Opus 4.8 | - | - | 78.9 | 21.1 | 23.6 | - | - |
| Codex CLI | GPT-5.6 Terra | - | - | 78.4 | 20.8 | 21.5 | - | - |
| Grok Build | Grok 4.6 | - | - | - | 26.5 | 20.3 | - | - |
| Codex CLI | GPT-5.6 Luna | - | - | 75.7 | 14.3 | 17.3 | - | - |
| Codex CLI | GPT-6 Luna | - | - | - | - | 16.4 | - | - |
| Muse Code | Muse Spark 1.3 | - | - | - | - | 14.5 | - | - |
| Grok Build | Grok 4.5 | - | - | - | - | 12.4 | - | - |
| Claude Code | Claude Sonnet 5 | - | - | 74.6 | 14.6 | 12.4 | - | - |
| Devin | SWE-1.7 Lightning | - | - | - | 18.6 | - | - | - |
| Cursor Agent | Grok 4.5 | - | - | 79.3 | 15.7 | - | - | - |
| Claude Code | GLM-5.2 | - | - | - | 4.6 | - | - | - |
| Codex CLI | GPT-5.5 | - | - | 83.2 | - | - | - | - |
| Claude Code | Claude Opus 4.7 | - | - | 68.9 | - | - | - | - |
| Gemini CLI | Gemini 3 Pro | - | - | 65.8 | - | - | - | - |
| Gemini CLI | Gemini 3.1 Pro | - | - | 65.8 | - | - | - | - |
| Claude Code | GLM-5.1 | - | - | 58.6 | - | - | - | - |
| Aider | Claude Opus 4.7 | - | 31.2 | - | - | - | 84.2 | - |
| Aider | GPT-5.5 | - | 28.5 | - | - | - | 81.8 | - |
| Aider | DeepSeek V4 Pro | - | 19.7 | - | - | - | 73.4 | - |
| OpenHands | Claude Sonnet 4.6 | 65.8 | 30.1 | - | - | - | - | 28.4 |
| OpenHands | GPT-5.5 | 64.2 | 29.6 | - | - | - | - | - |
| Cline | Claude Sonnet 4.6 | 63.4 | - | - | - | - | - | - |
| Windsurf Cascade | GPT-5.5 | 64.1 | 37.8 | - | - | - | - | - |
| Windsurf Cascade | SWE-1 (Codeium) | 58.2 | 30.4 | - | - | - | - | - |
| Amp | Claude Sonnet 4.6 | 70.8 | - | - | - | - | - | - |
| Continue | Claude Sonnet 4.6 | 52.4 | - | - | - | - | - | - |
| Roo Code | Claude Sonnet 4.6 | 57.3 | - | - | - | - | - | - |
Harness Directory
Every harness in the matrix above, with a link to the detail page.
Anthropic's official terminal agent. Native MCP, hooks, slash commands, subagent orchestration, and CLAUDE.md project memory.
VS Code fork with a multi-file agent (Composer) and a hosted background agent. Largest paid install base of any AI IDE.
OpenAI's open-source terminal agent. Sandboxed code execution, OpenAI Apps SDK plug-ins, MIT license.
xAI's open-source terminal agent, run as the grok command and powered by Grok 4.7. Parallel subagents in their own worktrees, plan mode, skills, hooks, and MCP servers, Apache 2.0 license.
Google's open-source terminal agent, now maintained for enterprise Gemini Code Assist. MCP servers, GEMINI.md context files, headless JSON output, Apache 2.0 license.
Google's terminal agent for Antigravity, run as agy. Shares its agent engine with the Antigravity 2.0 app, with background subagents in git worktrees, MCP, and skill bundles. Took over free individual Gemini access from Gemini CLI on June 18, 2026.
Meta's terminal and CI coding agent, run as muse and co-trained with Muse Spark. Multi-agent by default, with an OS sandbox and approvals from the first run, computer use, and MCP. The CLI is a native binary; its session SDK is MIT licensed.
Open-source CLI. Edit-by-diff over whole-file rewrites; runs on any OpenAI-compatible model. Maintains the Polyglot leaderboard.
Formerly OpenDevin. Open-source autonomous SWE agent with sandboxed runtime, browser tool, and microservice agent architecture.
Hosted autonomous SWE agent with persistent VM workspaces, Slack and IDE integrations, and DeepWiki repo retrieval.
Most-installed open-source VS Code agent. Plan-and-act loop with explicit human approval, MCP support, BYOK pricing.
Standalone IDE with Cascade multi-step agent loop. Backed by either frontier APIs or Codeium's own SWE-1 model family.
Sourcegraph's VS Code and JetBrains agent. Anchored on Sonnet 4.6, layered on a code-graph retrieval system that scales to monorepos.
Open-source VS Code and JetBrains agent. First-class local model support (Ollama, LM Studio), per-task model routing.
Open-source VS Code agent forked from Cline. Specialized modes (Code, Architect, Ask, Debug), MCP support.
For agents: the same data is served as JSON at /api/harnesses. Free, no auth, cached 5 minutes.