✦ Our pick · Build a new service or API
Claude Sonnet 5.5 at xhigh effort, $11.85 per finished task, is the cheapest you can get today. Within our measurement error it ties with GPT-6.1 Sol, Claude Opus 5.5, MiMo-V2.6-Pro and DeepSeek V4.1 Flash: equally good bets, worth trying two.
Updated Oct 13, 2026, every 3 hours. See it live with your plans and costs.
The top 10 for build a new service or api
Brand-new code from a spec: an API, a service, a module or a whole app. Cost per finished task: the tokens at each model's best effort level for this work, plus your time fixing failed attempts ($93 an hour). Success is the share of such tasks a model gets right the first time, on the benchmarks below.
| # | Model | Effort | Success | $/task |
|---|---|---|---|---|
| 1 | Gemini 4 Argon Google | high | 76.8% partly measured | $11.56 |
| 2 | Claude Sonnet 5.5 Anthropic | xhigh | 76.9% measured | $11.85 |
| 3 | GPT-6.1 Sol OpenAI | medium | 74.4% partly measured | $11.96 |
| 4 | Claude Opus 5.5 Anthropic | medium | 76.5% measured | $12.10 |
| 5 | MiMo-V2.6-Pro Xiaomi | default | 71.4% partly measured | $12.64 |
| 6 | DeepSeek V4.1 Flash DeepSeek | max | 71.1% partly measured | $12.86 |
| 7 | GPT-6 Sol OpenAI | high | 72.6% partly measured | $14.00 |
| 8 | Claude Haiku 5.5 Anthropic | max | 68.8% partly measured | $14.41 |
| 9 | GPT-6 Luna OpenAI | max | 68.6% partly measured | $14.45 |
| 10 | Muse Spark 1.3 Meta | xhigh | 71.1% partly measured | $14.46 |
Best within your tool
| Tool | Model | Effort | $/task |
|---|---|---|---|
| Cursor | Claude Sonnet 5.5 | xhigh | $11.85 |
| Claude Code | Claude Sonnet 5.5 | xhigh | $11.85 |
| Codex | GPT-6.1 Sol | medium | $11.96 |
| Antigravity | Claude Sonnet 5.5 | xhigh | $11.85 |
| GitHub Copilot | Claude Sonnet 5.5 | xhigh | $11.85 |
If tokens come with your plan
Tokens included in a plan cost you nothing extra, so the pick is the most reliable model in it, at the leanest effort level within our error of the best (the top levels only use up the plan sooner).
| Plan | Model | Effort |
|---|---|---|
| Claude Pro | Claude Opus 5.5 | high |
| Claude Max | Claude Opus 5.5 | high |
| ChatGPT Plus or Business | GPT-6.1 Sol | high |
| ChatGPT Pro | GPT-6.1 Sol | high |
| Google AI Pro | Claude Opus 5.5 | high |
| Google AI Ultra | Claude Opus 5.5 | high |
Measured on
- DeepSWE v1.1 · Datacurve; models it hasn't run filled in from Artificial Analysis (agent runs). 113 original tasks across 91 real repositories in 5 languages, each requiring a large, novel change (not taken from public commits); pass@1 on hidden tests. Measured for 36 of the 42 ranked models, results as of Oct 13, 2026.
- FrontierSWE · its authors (via Epoch AI). Long, open-ended engineering projects (building, speeding up and researching real software), hours each; mean of 5 runs in one harness. Measured for 18 of the 42 ranked models, results as of Oct 13, 2026.
- CursorBench 4.0 · Cursor. Ambiguous multi-file tasks from real Cursor sessions: edit, refactor, investigate, follow intent and design. Measured for 16 of the 42 ranked models, results as of Oct 13, 2026.
- Vibe Code Bench · Vals AI. Build a complete web application from a spec, end to end; scored by automated tests of the running app. Measured for 38 of the 42 ranked models, results as of Oct 12, 2026.
How every number is worked out, with the errors of our estimates: the method.