Home / Best AI models / Best AI for coding
The best AI for coding: models, tools and cost per task
Three independent coding tests put two models on top: Claude Opus 5.5 leads FrontierCode and CursorBench, and GPT-6 Astra leads DeepSWE. The bigger lesson is in the cost: on CursorBench, Opus 5.5 at high effort scored almost as well as at max effort for under a third of the price per task. Here are the results, what each task costs and the coding tools that use these models.
Benchmark results compiled by Epoch AI, checked October 18, 2026
FrontierCode: would the code be merged?
FrontierCode, from Cognition, the company behind Devin, checks whether an AI agent’s code changes would be merged.
| Model (effort) | Score |
|---|---|
| Claude Opus 5.5 (medium) | 54.6% |
| Claude Opus 5 (max) | 53.4% |
| GPT-6 Astra (max) | 53.3% |
| Claude Sonnet 5.5 (extra high) | 52.1% |
| Claude Fable 5.1 (medium) | 50.9% |
| GPT-6.1 Sol (medium) | 50.2% |
CursorBench: real coding sessions, with cost per task
CursorBench, from the makers of the Cursor editor, uses vague, multi-file tasks taken from real Cursor sessions and reports what each task cost.
| Model (effort) | Score | Cost per task |
|---|---|---|
| Claude Opus 5.5 (max) | 57.8% | $13.43 |
| Claude Opus 5.5 (high) | 56% | $3.97 |
| Claude Sonnet 5.5 (max) | 55.5% | $7.05 |
| Claude Sonnet 5.5 (extra high) | 53.1% | $2.81 |
| Gemini 4 Argon (high) | 52.9% | $4.57 |
| Claude Fable 5.1 (max) | 51.8% | $17.28 |
| Claude Haiku 5.5 (max) | 48.4% | $1.12 |
| Grok 4.7 (extra high) | 46.3% | $6.01 |
OpenAI’s newest models aren’t in the CursorBench results yet.
DeepSWE: hard software tasks
DeepSWE, from Datacurve, runs every model with the same simple agent, so it compares models rather than tools. Google’s low-cost Gemini 3.8 Flash comes within a point of GPT-6 Astra at about half the cost per task.
| Model (effort) | Pass rate | Cost per task |
|---|---|---|
| GPT-6 Astra (extra high) | 74.1% | $4.43 |
| Gemini 3.8 Flash (high) | 73.8% | $2.36 |
| Claude Opus 5 (max) | 73.6% | — |
| GPT-6 Astra (medium) | 72.8% | $3.08 |
| GLM 5.3 (max) | 69% | $3.99 |
| GLM 5.3 Flash (max) | 63.4% | $0.48 |
Anthropic’s 5.5 models aren’t in the DeepSWE results yet. All three tables: results compiled by Epoch AI (CC BY 4.0) from each benchmark’s leaderboard.
What the results mean for your bill
- Max effort rarely pays. Opus 5.5 scored 57.8% on CursorBench at max effort for $13.43 a task, and 56.0% at high effort for $3.97.
- Cheaper models come close. Sonnet 5.5 at extra-high effort scored 53.1% for $2.81, and Haiku 5.5 48.4% for $1.12. On DeepSWE, Gemini 3.8 Flash nearly matched GPT-6 Astra for $2.36 a task.
- The top tier can cost more for less. On CursorBench, Claude Fable 5.1 at max effort scored 51.8% for $17.28 a task, below Opus 5.5 at medium effort, 52.5% for $2.91.
The best AI coding assistants and what they cost
| Tool | Free | Paid | Models |
|---|---|---|---|
| Claude Code | No | With Claude Pro $20, Max from $100 | Claude |
| Codex | Limited, on ChatGPT Free and Go | With ChatGPT Plus $20; expanded on Pro | OpenAI |
| GitHub Copilot | Limited chat and agent use | Pro $10 ($15 of AI credits), Pro+ $39 ($70), Max $100 ($200) | Several makers’ |
| Cursor | — | With Grok’s plans from $8; cloud agents from $30 | Grok included; others at API list price |
| Devin | Light quota | Pro $20, Max $200; teams $80 plus $40 a seat | OpenAI, Anthropic, Google, SpaceXAI and its own SWE-2 |
| Google’s Jules | — | Higher limits with Google AI Pro $19.99 | Gemini |
Sources: Claude, ChatGPT, GitHub Copilot, Grok, Devin and Google AI plans. More: Claude Code pricing and Cursor pricing.
Questions
What is the best LLM for coding?
Claude Opus 5.5 leads two of the three independent coding tests here, FrontierCode and CursorBench, and GPT-6 Astra leads the third, DeepSWE. For the price, Opus 5.5 at high effort and Sonnet 5.5 give most of the result for much less.
What is the best free AI for coding?
GitHub Copilot’s free plan, Devin’s free plan and the limited Codex in ChatGPT’s free plan all help with code; Claude’s free plan can write and run code but doesn’t include Claude Code.
Is Claude or ChatGPT better for coding?
On these tests, Claude Opus 5.5 leads FrontierCode and CursorBench and GPT-6 Astra leads DeepSWE. Both are included at $20 a month, as Claude Code with Claude Pro and Codex with ChatGPT Plus; see Claude vs ChatGPT.
What is the best AI coding assistant?
For most developers, Claude Code or Codex through a $20 plan, or GitHub Copilot from $10 if you want several makers’ models in your editor.
Sources
- Cognition: FrontierCode, Cursor: CursorBench and Datacurve: DeepSWE, as compiled by Epoch AI (CC BY 4.0)
- Plans: Claude, ChatGPT, GitHub Copilot, Grok, Devin, Google AI plans