Coding Agent Benchmarks
How AI coding agents compare on real tasks: how many they solve, what each task costs, how long it takes and how many tokens it burns. Results come from public leaderboards and refresh hourly. Filter the charts to the agents you care about, then save the view to your account.
Hard, real-world tasks done in a terminal: compiling code, training models, setting up servers. Each agent and model pair runs every task several times. Source: tbench.ai.
Showing 23 of 35 runs.
- Claude Code
- Codex
- Grok Build
- mini-SWE-agent
- Muse Code
Performance
Share of tasks each agent and model pair solved.
Score
Accuracy across all trials. Hover for the 95% confidence range.
- 1Opus 5.5 (max)Claude Code · Anthropic64.8%
- 2Sonnet 5.5 (max)Claude Code · Anthropic61.8%
- 3GPT-6 Astra (max)Codex · OpenAI58.2%
- 4GPT-6.1 Sol (max)Codex · OpenAI58.2%
- 5Fable 5.1 (xhigh)Claude Code · Anthropic57.9%
- 6Opus 5 (xhigh)Claude Code · Anthropic53.9%
- 7GPT-6 Sol (max)Codex · OpenAI49.4%
- 8Fable 5 (max)Claude Code · Anthropic44.5%
- 9GLM-5.3 (max)Claude Code · Z.ai41.8%
- 10Grok 4.7 (xhigh)Grok Build · xAI37.6%
- 11GPT-5.6 Sol (max)Codex · OpenAI37.3%
- 12GLM-5.3-FlashClaude Code · Z.ai35.8%
- 13Qwen3.8-Max-0902 (max)Claude Code · Alibaba27.0%
- 14Opus 4.8 (max)Claude Code · Anthropic23.6%
- 15GPT-5.6 Terra (max)Codex · OpenAI21.5%
Score vs. cost per task
Top-left is better: higher score for less money. Labelled runs are the best value at their price.
Token usage
How much the agent reads and writes to finish one task.
Tokens per task
Agents re-read their context every turn, so cached input usually dominates.
- Input
- Cached input
- Output
- 1Sonnet 5 (max)Claude Code · Anthropic65.3M
- 2Sonnet 5.5 (max)Claude Code · Anthropic58.8M
- 3Gemini 3.8 Flash (high)mini-SWE-agent · Google52.1M
- 4GPT-5.6 Luna (max)Codex · OpenAI35.0M
- 5Gemini 3.7 Flash (high)mini-SWE-agent · Google33.7M
- 6Muse Spark 1.3 (xhigh)Muse Code · Meta27.3M
- 7GLM-5.3 (max)Claude Code · Z.ai26.3M
- 8Opus 5.5 (max)Claude Code · Anthropic24.3M
- 9Opus 5 (xhigh)Claude Code · Anthropic20.9M
- 10Opus 4.8 (max)Claude Code · Anthropic19.4M
- 11GPT-5.6 Terra (max)Codex · OpenAI17.2M
- 12Grok 4.7 (xhigh)Grok Build · xAI15.8M
- 13GPT-6 Sol (max)Codex · OpenAI15.1M
- 14GPT-5.6 Sol (max)Codex · OpenAI13.4M
- 15Grok 4.6 (high)Grok Build · xAI12.1M
Score vs. tokens per task
Cost
Average API spend to attempt one task, at list prices.
Cost per task
Lower is better. Runs without a reported cost are left out.
- 1GPT-6 Luna (max)Codex · OpenAI$0.22
- 2GLM-5.3-FlashClaude Code · Z.ai$0.36
- 3GPT-5.6 Luna (max)Codex · OpenAI$1.05
- 4GPT-6.1 Sol (max)Codex · OpenAI$1.92
- 5Gemini 3.7 Flash (high)mini-SWE-agent · Google$3.82
- 6GPT-6 Sol (max)Codex · OpenAI$4.40
- 7GPT-5.6 Terra (max)Codex · OpenAI$5.25
- 8Gemini 3.8 Flash (high)mini-SWE-agent · Google$5.54
- 9Grok 4.5 (high)Grok Build · xAI$6.35
- 10GPT-5.6 Sol (max)Codex · OpenAI$7.70
- 11GLM-5.3 (max)Claude Code · Z.ai$8.27
- 12GPT-6 Astra (max)Codex · OpenAI$9.90
- 13Grok 4.6 (high)Grok Build · xAI$10.88
- 14Grok 4.7 (xhigh)Grok Build · xAI$11.16
- 15Qwen3.8-Max-0902 (max)Claude Code · Alibaba$14.05
Execution time
Wall-clock time for one attempt at a task.
Time per task
Lower is better.
- 1Gemini 3.7 Flash (high)mini-SWE-agent · Google28m
- 2Gemini 3.8 Flash (high)mini-SWE-agent · Google33m
- 3GPT-6 Luna (max)Codex · OpenAI38m
- 4Grok 4.6 (high)Grok Build · xAI39m
- 5GPT-5.6 Sol (max)Codex · OpenAI40m
- 6GPT-5.6 Terra (max)Codex · OpenAI42m
- 7GPT-6 Sol (max)Codex · OpenAI44m
- 8GPT-6 Astra (max)Codex · OpenAI47m
- 9Grok 4.5 (high)Grok Build · xAI53m
- 10GPT-6.1 Sol (max)Codex · OpenAI53m
- 11Fable 5.1 (xhigh)Claude Code · Anthropic53m
- 12GPT-5.6 Luna (max)Codex · OpenAI1.1h
- 13Fable 5 (max)Claude Code · Anthropic1.2h
- 14Muse Spark 1.3 (xhigh)Muse Code · Meta1.2h
- 15Opus 5 (xhigh)Claude Code · Anthropic1.3h
Score vs. time per task
All results
Every run that matches your filters.
| # | Agent | Model | Score | Cost / task | Time / task | Tokens / task | Date |
|---|---|---|---|---|---|---|---|
| 1 | Claude Code | Opus 5.5 (max) · Anthropic | 64.8% ±3.1 | $14.30 | 1.3h | 24.3M | 2026-09-22 |
| 2 | Claude Code | Sonnet 5.5 (max) · Anthropic | 61.8% ±2.9 | $22.22 | 2.1h | 58.8M | 2026-09-28 |
| 3 | Codex | GPT-6 Astra (max) · OpenAI | 58.2% ±2.8 | $9.90 | 47m | 4.6M | 2026-09-03 |
| 4 | Codex | GPT-6.1 Sol (max) · OpenAI | 58.2% ±3.1 | $1.92 | 53m | 4.4M | 2026-09-29 |
| 5 | Claude Code | Fable 5.1 (xhigh) · Anthropic | 57.9% ±3.4 | $14.76 | 53m | 7.1M | 2026-09-01 |
| 6 | Claude Code | Opus 5 (xhigh) · Anthropic | 53.9% ±3.2 | $18.44 | 1.3h | 20.9M | 2026-07-24 |
| 7 | Codex | GPT-6 Sol (max) · OpenAI | 49.4% ±3.2 | $4.40 | 44m | 15.1M | 2026-09-22 |
| 8 | Claude Code | Fable 5 (max) · Anthropic | 44.5% ±3.9 | $22.02 | 1.2h | 11.5M | 2026-06-09 |
| 9 | Claude Code | GLM-5.3 (max) · Z.ai | 41.8% ±3.2 | $8.27 | 1.6h | 26.3M | 2026-08-14 |
| 10 | Grok Build | Grok 4.7 (xhigh) · xAI | 37.6% ±3.5 | $11.16 | 1.6h | 15.8M | 2026-09-21 |
| 11 | Codex | GPT-5.6 Sol (max) · OpenAI | 37.3% ±3.8 | $7.70 | 40m | 13.4M | 2026-06-26 |
| 12 | Claude Code | GLM-5.3-Flash · Z.ai | 35.8% ±3.5 | $0.36 | 1.7h | 8.8M | 2026-08-26 |
| 13 | Claude Code | Qwen3.8-Max-0902 (max) · Alibaba | 27.0% ±3.8 | $14.05 | 2.8h | 9.0M | 2026-09-02 |
| 14 | Claude Code | Opus 4.8 (max) · Anthropic | 23.6% ±3.6 | $19.64 | 1.4h | 19.4M | 2026-05-28 |
| 15 | Codex | GPT-5.6 Terra (max) · OpenAI | 21.5% ±3.3 | $5.25 | 42m | 17.2M | 2026-06-26 |
| 16 | Grok Build | Grok 4.6 (high) · xAI | 20.3% ±3.1 | $10.88 | 39m | 12.1M | 2026-08-12 |
| 17 | mini-SWE-agent | Gemini 3.8 Flash (high) · Google | 19.1% ±3.4 | $5.54 | 33m | 52.1M | 2026-09-02 |
| 18 | Codex | GPT-5.6 Luna (max) · OpenAI | 17.3% ±2.9 | $1.05 | 1.1h | 35.0M | 2026-06-26 |
| 19 | Codex | GPT-6 Luna (max) · OpenAI | 16.4% ±2.7 | $0.22 | 38m | 11.8M | 2026-09-22 |
| 20 | Muse Code | Muse Spark 1.3 (xhigh) · Meta | 14.6% ±2.9 | — | 1.2h | 27.3M | 2026-09-02 |
| 21 | Grok Build | Grok 4.5 (high) · xAI | 12.4% ±2.6 | $6.35 | 53m | 10.3M | 2026-07-16 |
| 22 | Claude Code | Sonnet 5 (max) · Anthropic | 12.4% ±3.1 | $29.10 | 1.8h | 65.3M | 2026-06-30 |
| 23 | mini-SWE-agent | Gemini 3.7 Flash (high) · Google | 11.2% ±2.5 | $3.82 | 28m | 33.7M | 2026-08-13 |
Frequently asked questions
What is a coding agent?
A coding agent is a program that pairs an AI model with tools such as a terminal, a file editor and a test runner, then works on a task over many steps. Claude Code, Codex and Grok Build are agents; Opus, GPT and Grok are the models inside them. The same model can score differently in different agents.
Where does this data come from?
Terminal-Bench results come from the public leaderboard at tbench.ai, which reports score, total cost, token counts and average run time for each agent and model pair. SWE-bench Verified results come from the public leaderboard at swebench.com. Both are fetched automatically and refreshed every hour.
How is cost per task worked out?
For Terminal-Bench, it is the total API cost the leaderboard reports divided by the number of trials. SWE-bench publishes an average cost per issue directly. Both use list API prices, so a subscription plan or a discount changes what you would actually pay.
Why is cached input so large?
An agent sends its whole conversation back to the model on every step. Most of that text was sent before, so providers serve it from a cache at a much lower price. That is why agents can read billions of tokens while the bill stays modest.
How do I save my view?
Your filters and chart settings are remembered on your device automatically. Create a free account to save them to your profile, so the page opens exactly as you left it on any device. Reset to defaults restores the standard view at any time.
Comparing raw model prices instead? Try the AI model comparison or the AI API cost calculator.