Leaderboard · updated 8 September 2026
Coding Harness Leaderboard
The highest accuracy on Terminal-Bench 2.1 is 87.4% ± 1.8, running GPT-6 Astra, as of September 2026. Terminus 2 has no published run on GPT-6 Astra, so its lift is read from GPT-5.5, where the harness adds +5.2 pp. Coding harnesses are ranked once per version, because two versions are two different task sets: 4 on Terminal-Bench 2.1 and 29 on Terminal-Bench 2.0. Beneath the tables, 9 are evaluation harnesses, shown but not ranked, 4 have only their own numbers, and 11 have none at all. Last change 8 September 2026.
We went through every coding harness on this board and took its score from the benchmark’s own published rows, one link a row. Nothing is re-run and nothing is re-scored here. The only figure worked out on this page is the lift, and it is subtraction between two published rows.
How we checked
- Every rank rests on an official row
- A score counts only when the benchmark itself has published the row, and the Source column on each entry is that row. A figure with only the vendor's own report behind it is shown beneath the tables and never ranked, however plausible it looks.
- Lift needs a baseline row on the same model
- Lift is the coding harness's accuracy minus the same model's accuracy under Terminus 2. Where the benchmark publishes no Terminus 2 row for that model, the entry carries no lift rather than borrowing a figure from another model.
- A version is a task set, not a revision
- Two versions of the benchmark are two different sets of tasks, so there is one table each and no combined order. A coding harness ranked on both appears on both, with its own row and its own source each time.
- Scaffolds are shown, never ranked
- Some entries are harnesses built to run the benchmark rather than products anyone installs, and there are 9 of them here. Ranking those against the tools would compare two different things, so they sit beneath the tables with their figures.
Every table and every list below, searched at once.
Terminal-Bench 2.1
latest version · 4 ranked
Terminal-Bench 2.1, 4 ranked on official rows. Lift is the accuracy minus the same model’s accuracy under Terminus 2, in percentage points. Data as of 8 September 2026.
What the columns mean
- Rank
- Where a harness sits on this table, ordered by its best accuracy on this version. A version is a task set of its own, so a rank here is never a rank against another table. Sorting by another column leaves these numbers alone, because the rank belongs to the harness and not to the row order.
- Harness
- The coding agent being measured: the tool you install and run, not the model inside it. The same model scores differently under different harnesses, and that difference is the thing this board exists to show. Open a harness to see every official row it holds on this version.
- Accuracy
- The figure on the official row, copied as published, with the plus or minus as published beside it. That second figure is the interval the benchmark reports across repeated runs, so two scores whose intervals overlap are not separated by the data. Nothing here recomputes either number.
- Lift vs model
- A row's accuracy minus the same model's accuracy under Terminus 2, the reference agent the benchmark runs every model through. Positive means the harness got more out of that model than the reference agent did. Beside a rank it is the harness's highest-scoring row that has a Terminus 2 run to subtract, which is not always the row in the Accuracy column; when it is a different model, that model is named underneath. It is blank only when no row of that harness has a Terminus 2 counterpart at all.
- Source
- The permanent row this figure was read from, on the leaderboard the benchmark maintainers publish. Every ranked number on this page links to one. A number with only a vendor's own report behind it is listed beneath the tables and never ranked.
| Rank | Harness | Accuracy | Lift vs model | Source |
|---|---|---|---|---|
| 1 | Codex CLICoding Agent | 87.4% ± 1.8GPT-6 Astra | +5.2 ppon GPT-5.5 | official row |
| GPT-6 Astra high | 87.4% ± 1.8 | — | source | |
| GPT-6 Astra medium | 87.0% ± 1.9 | — | source | |
| GPT-6 Astra low | 86.7% ± 1.8 | — | source | |
| GPT-6 Astra max | 86.7% ± 1.4 | — | source | |
| GPT-6 Astra xhigh | 85.8% ± 1.4 | — | source | |
| GPT-5.5 xhigh | 83.2% ± 2.2 | +5.2 pp | source | |
| GPT-5.6 Terra max | 78.4% ± 2.5 | — | source | |
| GPT-5.6 Luna max | 75.7% ± 2.6 | — | source | |
| 2 | Claude CodeCoding Agent | 83.8% ± 2.3Fable 5 | +3.3 pp | official row |
| Fable 5 xhigh | 83.8% ± 2.3 | +3.3 pp | source | |
| Opus 4.8 high | 78.9% ± 2.6 | — | source | |
| Sonnet 5 high | 74.6% ± 3.2 | — | source | |
| Opus 4.7 max | 68.9% ± 2.8 | +2.8 pp | source | |
| GLM-5.1 max | 58.6% ± 2.4 | — | source | |
| 3 | Cursor CLI | 79.3% ± 2.9Grok 4.5 | — | official row |
| Grok 4.5 high | 79.3% ± 2.9 | — | source | |
| 4 | Gemini CLICoding Agent | 65.8% ± 2.7Gemini 3 Pro | -8.1 pp | official row |
| Gemini 3 Pro high | 65.8% ± 2.7 | -8.1 pp | source | |
| Gemini 3.1 Pro high | 65.8% ± 3.3 | +0.2 pp | source | |
| No ranked harness here matches . | ||||
Terminal-Bench 2.0
earlier version · 29 ranked
Terminal-Bench 2.0, 29 ranked on official rows. Lift is the accuracy minus the same model’s accuracy under Terminus 2, in percentage points. Data as of 8 September 2026.
What the columns mean
- Rank
- Where a harness sits on this table, ordered by its best accuracy on this version. A version is a task set of its own, so a rank here is never a rank against another table. Sorting by another column leaves these numbers alone, because the rank belongs to the harness and not to the row order.
- Harness
- The coding agent being measured: the tool you install and run, not the model inside it. The same model scores differently under different harnesses, and that difference is the thing this board exists to show. Open a harness to see every official row it holds on this version.
- Accuracy
- The figure on the official row, copied as published, with the plus or minus as published beside it. That second figure is the interval the benchmark reports across repeated runs, so two scores whose intervals overlap are not separated by the data. Nothing here recomputes either number.
- Lift vs model
- A row's accuracy minus the same model's accuracy under Terminus 2, the reference agent the benchmark runs every model through. Positive means the harness got more out of that model than the reference agent did. Beside a rank it is the harness's highest-scoring row that has a Terminus 2 run to subtract, which is not always the row in the Accuracy column; when it is a different model, that model is named underneath. It is blank only when no row of that harness has a Terminus 2 counterpart at all.
- Source
- The permanent row this figure was read from, on the leaderboard the benchmark maintainers publish. Every ranked number on this page links to one. A number with only a vendor's own report behind it is listed beneath the tables and never ranked.
| Rank | Harness | Accuracy | Lift vs model | Source |
|---|---|---|---|---|
| 1 | NexAU-AHE | 84.7% ± 2.1GPT-5.5 | — | official row |
| GPT-5.5 | 84.7% ± 2.1 | — | source | |
| 2 | Capy | 83.1% ± 2.1GPT-5.5 | +12.4 ppon Claude Opus 4.6 | official row |
| GPT-5.5 | 83.1% ± 2.1 | — | source | |
| Claude Opus 4.6 | 75.3% ± 2.4 | +12.4 pp | source | |
| 3 | Codex CLICoding Agent | 82.2% ± 2.2GPT-5.5 | +8.9 ppon GPT-5.2 | official row |
| GPT-5.5 | 82.2% ± 2.2 | — | source | |
| GPT-5.2 | 62.9% ± 3.0 | +8.9 pp | source | |
| GPT-5.1-Codex-Max | 60.4% ± 2.7 | — | source | |
| GPT-5 | 49.6% ± 2.9 | +14.4 pp | source | |
| GPT-5-Codex | 44.3% ± 2.7 | +0.9 pp | source | |
| GPT-5-Mini | 31.9% ± 3.0 | +7.9 pp | source | |
| GPT-5-Nano | 11.5% ± 2.3 | +3.6 pp | source | |
| 4 | TongAgents | 80.2% ± 2.6Gemini 3.1 Pro | +9.0 ppon Claude Opus 4.6 | official row |
| Gemini 3.1 Pro | 80.2% ± 2.6 | — | source | |
| Claude Opus 4.6 | 71.9% ± 2.7 | +9.0 pp | source | |
| 5 | WOZCODE | 80.2% ± 2.1Claude Opus 4.7 | — | official row |
| Claude Opus 4.7 | 80.2% ± 2.1 | — | source | |
| 6 | SageAgent | 78.4% ± 2.2GPT-5.3-Codex | +13.7 pp | official row |
| GPT-5.3-Codex | 78.4% ± 2.2 | +13.7 pp | source | |
| Gemini 3 Pro | 65.2% ± 2.1 | +8.3 pp | source | |
| 7 | Droid | 77.3% ± 2.2GPT-5.3-Codex | +12.6 pp | official row |
| GPT-5.3-Codex | 77.3% ± 2.2 | +12.6 pp | source | |
| Claude Opus 4.6 | 69.9% ± 2.5 | +7.0 pp | source | |
| GPT-5.2 | 64.9% ± 2.8 | +10.9 pp | source | |
| Claude Opus 4.5 | 63.1% ± 2.7 | +5.3 pp | source | |
| Gemini 3 Pro | 61.1% ± 2.8 | +4.2 pp | source | |
| 8 | CodeBrain-1.5 | 75.8% ± 2.0GPT-5.3-Codex | +11.1 pp | official row |
| GPT-5.3-Codex | 75.8% ± 2.0 | +11.1 pp | source | |
| Gemini 3 Pro | 62.2% ± 2.6 | +5.3 pp | source | |
| 9 | Codelia | 75.7% ± 2.2GPT-5.3-Codex | +11.0 pp | official row |
| GPT-5.3-Codex | 75.7% ± 2.2 | +11.0 pp | source | |
| 10 | Mux | 74.6% ± 2.5GPT-5.3-Codex | +9.9 pp | official row |
| GPT-5.3-Codex | 74.6% ± 2.5 | +9.9 pp | source | |
| Claude Opus 4.6 | 66.5% ± 2.5 | +3.6 pp | source | |
| GPT-5.2 | 60.7% | +6.7 pp | source | |
| Claude Opus 4.5 | 58.4% | +0.6 pp | source | |
| 11 | MAYA-V2 | 72.1% ± 2.2Claude 4.6 Opus | +9.2 pp | official row |
| Claude 4.6 Opus | 72.1% ± 2.2 | +9.2 pp | source | |
| Claude 4.5 Sonnet | 42.7% | -0.1 pp | source | |
| 12 | Ante | 69.4% ± 2.1Gemini 3 Pro | +12.5 pp | official row |
| Gemini 3 Pro | 69.4% ± 2.1 | +12.5 pp | source | |
| 13 | IndusAGI Coding Agent | 69.1% ± 2.3GPT-5.3-Codex | +4.4 pp | official row |
| GPT-5.3-Codex | 69.1% ± 2.3 | +4.4 pp | source | |
| MiniMax M2.7 | 45.1% | — | source | |
| 14 | Crux | 66.9%Claude Opus 4.6 | +4.0 pp | official row |
| Claude Opus 4.6 | 66.9% | +4.0 pp | source | |
| GPT-5.1-Codex | 57.8% ± 2.9 | +20.9 pp | source | |
| GPT-5.1-Codex-Mini | 43.1% ± 3.0 | — | source | |
| MiniMax M2.1 | 36.6% ± 2.9 | +7.4 pp | source | |
| GLM 4.7 | 33.3% ± 2.5 | -0.1 pp | source | |
| 15 | Deep AgentsAgent Harness | 66.5% ± 3.1GPT-5.2-Codex | — | official row |
| GPT-5.2-Codex | 66.5% ± 3.1 | — | source | |
| 16 | clnkr | 66.1% ± 2.5GPT-5.5 | — | official row |
| GPT-5.5 | 66.1% ± 2.5 | — | source | |
| 17 | Junie CLI | 64.3% ± 2.8Gemini 3 Flash | +12.6 pp | official row |
| Gemini 3 Flash | 64.3% ± 2.8 | +12.6 pp | source | |
| 18 | II-Agent | 61.8% ± 2.8Gemini 3 Pro | +4.9 pp | official row |
| Gemini 3 Pro | 61.8% ± 2.8 | +4.9 pp | source | |
| 19 | hookele | 61.6% ± 1.9GPT-5.1-Codex-Mini | — | official row |
| GPT-5.1-Codex-Mini | 61.6% ± 1.9 | — | source | |
| 20 | Gemini CLICoding Agent | 61.4% ± 4.1Gemini 3.1 Pro | -4.3 ppon Gemini 3 Flash | official row |
| Gemini 3.1 Pro | 61.4% ± 4.1 | — | source | |
| Gemini 3.1 Pro | 59.4% ± 4.2 | — | source | |
| Gemini 3 Flash | 47.4% ± 3.0 | -4.3 pp | source | |
| Gemini 2.5 Pro | 19.6% ± 2.9 | -13.0 pp | source | |
| Gemini 2.5 Flash | 15.4% ± 2.3 | -1.5 pp | source | |
| 21 | Letta Code | 59.1% ± 2.4Claude Opus 4.5 | +1.3 pp | official row |
| Claude Opus 4.5 | 59.1% ± 2.4 | +1.3 pp | source | |
| Gemini 3 Pro | 56.0% ± 3.0 | -0.9 pp | source | |
| GPT-5.1-Codex | 53.5% ± 2.8 | +16.6 pp | source | |
| 22 | Claude CodeCoding Agent | 58.0% ± 2.9Claude Opus 4.6 | -4.9 pp | official row |
| Claude Opus 4.6 | 58.0% ± 2.9 | -4.9 pp | source | |
| Claude Opus 4.5 | 52.1% ± 2.5 | -5.7 pp | source | |
| Claude Sonnet 4.5 | 40.1% ± 2.9 | -2.7 pp | source | |
| Claude Opus 4.1 | 34.8% ± 2.9 | -3.2 pp | source | |
| Claude Haiku 4.5 | 27.5% ± 2.8 | -0.8 pp | source | |
| 23 | Grok CLI | 57.3%Grok 4.20 Reasoning | — | official row |
| Grok 4.20 Reasoning | 57.3% | — | source | |
| 24 | GooseCoding / General Agent | 54.3% ± 2.6Claude Opus 4.5 | -3.5 pp | official row |
| Claude Opus 4.5 | 54.3% ± 2.6 | -3.5 pp | source | |
| Claude Sonnet 4.5 | 43.1% ± 2.6 | +0.3 pp | source | |
| Claude Haiku 4.5 | 35.5% ± 2.9 | +7.2 pp | source | |
| 25 | Simplai Agent | 53.4% ± 2.8Claude Sonnet 4.6 | — | official row |
| Claude Sonnet 4.6 | 53.4% ± 2.8 | — | source | |
| 26 | OpenHandsCoding Agent / Agent Platform | 51.9% ± 2.9Claude Opus 4.5 | -5.9 pp | official row |
| Claude Opus 4.5 | 51.9% ± 2.9 | -5.9 pp | source | |
| GPT-5 | 43.8% ± 3.0 | +8.6 pp | source | |
| Claude Sonnet 4.5 | 42.6% ± 2.8 | -0.2 pp | source | |
| Claude Opus 4.1 | 36.9% ± 2.7 | -1.1 pp | source | |
| GPT-5-Mini | 29.2% ± 2.8 | +5.2 pp | source | |
| Grok 4 | 27.2% ± 3.1 | +4.1 pp | source | |
| Kimi K2 Instruct | 26.7% ± 2.7 | -1.1 pp | source | |
| Qwen 3 Coder 480B | 25.4% ± 2.6 | +1.5 pp | source | |
| Gemini 2.5 Flash | 16.4% ± 2.4 | -0.5 pp | source | |
| Gemini 2.5 Pro | 16.4% ± 2.8 | -16.2 pp | source | |
| Claude Haiku 4.5 | 13.9% ± 2.7 | -14.4 pp | source | |
| GPT-5-Nano | 9.9% ± 2.1 | +2.0 pp | source | |
| 27 | OpenCodeCoding Agent | 51.7%Claude Opus 4.5 | -6.1 pp | official row |
| Claude Opus 4.5 | 51.7% | -6.1 pp | source | |
| 28 | CAMEL-AI | 46.5% ± 2.4Claude Sonnet 4.5 | +3.7 pp | official row |
| Claude Sonnet 4.5 | 46.5% ± 2.4 | +3.7 pp | source | |
| 29 | Dakou Agent | 27.2% ± 2.6Qwen 3 Coder 480B | +3.3 pp | official row |
| Qwen 3 Coder 480B | 27.2% ± 2.6 | +3.3 pp | source | |
| No ranked harness here matches . | ||||
Evaluation harnesses, not ranked (9)
Scaffolds built to run the benchmark rather than tools you install. Their rows are official and are shown here, but ranking them against products would compare two different things.
- Meta-Harness76.4% ± 2.4official rowTerminal-Bench 2.0 · Claude Opus 4.6
- mini-SWE-agent76.2% ± 2.4official rowTerminal-Bench 2.1 · Muse Spark 1.1
- Simple Codex75.1% ± 2.4official rowTerminal-Bench 2.0 · GPT-5.3-Codex
- Terminus-KIRA74.8% ± 2.6official rowTerminal-Bench 2.0 · Gemini 3.1 Pro
- spoox-o-m71.5% ± 2.5official rowTerminal-Bench 2.0 · GPT-5.3-Codex
- Harness Agent42.9% ± 2.9official rowTerminal-Bench 2.0 · MiniMax M2.7 Highspeed
- cchuter42.7% ± 2.8official rowTerminal-Bench 2.0 · minimax-m2.5
- little-coder24.6% ± 3.2official rowTerminal-Bench 2.0 · Qwen3.6-35B-A3B
- Bash Agent19.3% ± 2.0official rowTerminal-Bench 2.0 · TermiGen-32B
Vendor-reported, not ranked (4)
- Cline73.0%vendorTerminal-Bench 2.0 · GPT-5.3-Codex
- Kilo Code74.2%vendorTerminal-Bench 2.0 Completion · GPT-5.5
- Kimi Code CLI88.3%vendorTerminal-Bench 2.1 · Kimi K3
- Pi61.8%vendorTerminal-Bench 2.1 · Kimi K2.7 Code
No published benchmark (11)
Aider, AutoHand, Continue, ForgeCode / Forge, Groq Code CLI, Mistral Vibe, Nanocoder, Neovate Code, Qwen Code, Roo Code, VT Code.
How the rank is computed
The rule, in full
One table per Terminal-Bench version, latest first. Within a table, a coding harness is ranked by its best accuracy on an official row, and the record cites that row. Lift is that accuracy minus the same model’s accuracy under Terminus 2. It is what the harness adds, and it is the number this board exists to show.
This board ranks tools you can install, and under each one lists every official row it has on that version. Scaffolds built to run the benchmark are shown beneath the tables with their figures, never ranked, because ranking them against products would compare two different things. A coding harness nobody has classified yet waits there too, rather than ranking unseen.
Vendor numbers are listed below the tables and never ranked.
Left out2
- Benchmarks with no official board: DeepSWE, SWE-Atlas-QnA, and vendor-defined metrics such as “Terminal-Bench 2.0 Completion”.
- Rows that are not a score: cost figures, configuration diffs, aggregates.
Changes
8 September 2026 · Rebuilt from the official catalogue, four harnesses ranked4
- The catalogue is rebuilt from every row the official source publishes: 190 results across 60 tools, against 63 across 22. Terminal-Bench 2.1 ranks 4 and 2.0 ranks 3.
- The board ranks Claude Code, Gemini CLI, Codex CLI and Cursor CLI. Every other tool with an official row is listed beneath the tables with its figure, never dropped.
- Each ranked tool now shows every official row it holds on that version, with the model, the effort level, the accuracy and the lift for each.
- Terminal-Bench 3.0 and 4.0 are not ranked. No Terminus 2 row is published for either, so lift cannot be computed on them.
7 September 2026 · Goose ranked on Terminal-Bench 2.01
- Goose: Goose was listed as having no published benchmark. It has three official rows on Terminal-Bench 2.0: 54.3% on Claude Opus 4.5, 43.1% on Claude Sonnet 4.5, and 35.5% on Claude Haiku 4.5. All three are now ranked. Its lift over Terminus 2 is +1.3 pp, averaged over the three.
3 September 2026 · First publication3
- The Coding Harness Leaderboard is published. Results are ranked per Terminal-Bench version, never across versions: two on 2.1, six on 2.0. Four more appear with vendor-reported numbers only; twelve are listed without a published benchmark.
- Codex CLI: GPT-5.5 · Terminal-Bench 2.1 corrected from 83.1% to 83.2% to match the official board.
- Gemini CLI: Gemini 3.1 Pro · Terminal-Bench 2.1 corrected from 67.1% to 65.8%, and its model baseline from 70.7% to 65.6%, to match the official board. The lift moves from −3.6 pp to +0.2 pp.
Questions
What is the highest accuracy on Terminal-Bench right now?
The best official row on Terminal-Bench 2.1 shows 87.4% ± 1.8 using GPT-6 Astra. The full order is in the table above.
What counts as an official row?
A permanent row on the Terminal-Bench leaderboard the benchmark maintainers publish. The link in the Source column is that row. A number with only the vendor’s own report behind it is shown, never ranked.
What is accuracy?
The figure on the official row, copied as published. The Terminal-Bench maintainers describe it as the resolution rate of the tasks: the share of runs in which the agent completed the task. A task is run more than once, and the ± figure is the 95% confidence interval across those runs. Nothing here recomputes it.
What is lift?
A row’s accuracy minus what the same model reaches under Terminus 2, the reference agent the Terminal-Bench maintainers run every model through, in percentage points. Positive means the harness gets more out of that model than the reference agent does. Negative means less. The figure beside a rank is one row’s measurement, taken from that harness’s highest-scoring row that has a Terminus 2 run to subtract. That is often not the row in the Accuracy column, because the strongest model on a board is frequently one Terminus 2 has not been run on, so the model it was measured on is named beneath it.
Why are the versions in separate tables?
A Terminal-Bench version is a task set. Ranking a score from one against a score from another would compare two different exams. A new version arrives as a new table on top.
How often does this update?
There is no fixed schedule. The date at the top is the last time the data changed, not the last time the page was built.