Best LLM for coding in 2026: scores and price per task
Claude Opus 5.5 and GPT-6 Astra lead the coding benchmarks of late 2026, with Claude Sonnet 5.5 close behind, but the price of a task differs tenfold between them. The effort setting moves cost and score as much as the model does. The scores from 3 independent coding benchmarks, the cost of a task on each model, and how to test them on your own repository.
Tecyfy · Published · 9 min read
The best LLM for coding in late 2026 is Claude Opus 5.5 or GPT-6 Astra on raw score, and Claude Sonnet 5.5 for most teams once the bill is counted. The 3 sit within a few points of each other on the independent coding benchmarks, while the cost of one task differs by up to 10 times. And the effort setting, how long the model thinks before it acts, moves both score and cost as much as the choice of model.
All figures below come from 3 benchmarks run by others, not by the model makers, as collected in Epoch AI's benchmark data on 23 October 2026. None of them is your codebase, which is why the last section is a way to run the same test on your own repository.
Which models score highest on coding benchmarks?
Opus 5.5 leads on merge-quality fixes and everyday tasks; GPT-6 Astra leads on the longest projects. Each benchmark measures a different kind of work, so no single number settles it.
| Model | FrontierCode: merge-ready fixes | CursorBench: everyday tasks (best effort) | FrontierSWE: 20-hour projects |
|---|---|---|---|
| Claude Opus 5.5 | 54.6% | 57.8% | 62.3% |
| GPT-6 Astra | 53.3% | not tested | 65.5% |
| Claude Sonnet 5.5 | 52.1% | 55.5% | 61.9% |
| Claude Fable 5.1 | 50.9% | 51.8% | 56.3% |
| GPT-6.1 Sol | 50.2% | not tested | not tested |
| Gemini 4 Argon | 47.6% | 52.9% | 55.0% |
| Claude Haiku 5.5 | 46.4% | 48.4% | 43.8% |
| GPT-6 Luna | 42.4% | not tested | not tested |
| Gemini 3.8 Flash | 41.2% | 39.6% | 19.6% |
FrontierCode, from Cognition, gives an agent a real open-source issue and grades the patch against hidden tests and a rubric written by the project's maintainers; failing any blocker scores zero, so a patch that passes tests but fails a blocker criterion scores zero. Models run through agent harnesses such as Claude Code, Codex, mini-SWE-agent and Devin. CursorBench uses ambiguous, multi-file tasks taken from real Cursor sessions, and has not yet listed the GPT-6 models. FrontierSWE gives 34 hard engineering projects, such as a filesystem that survives crashes or an FFmpeg optimisation, a 20-hour budget each, 5 times over.
Read the gaps with care. FrontierSWE's scores swing about 9 points either side between a model's worst and best run, and Cursor says small differences on its board may not be statistically meaningful. A 1 or 2 point lead is a tie. SWE-bench Verified, the benchmark most articles still quote, is left out because none of these models has an independent result on it in Epoch's data.
What does a coding task cost on each model?
From 8 cents to over $1,000, depending on the model, the effort level and the size of the task. Price per million tokens is only the start: a model that takes more steps, or thinks longer at each, sends far more tokens for the same task. The chart has every model at every effort level CursorBench tested, plus FrontierSWE's long projects.
Everyday multi-file tasks from real Cursor sessions
Cheapest at 50% or more: Claude Sonnet 5.5 at extra high effort, $2.81 a task (it scores 53.1%). The next cheapest model is Claude Opus 5.5 at medium effort at $2.91, about the same.
CursorBench by Cursor, via Epoch AI's data (CC BY 4.0). Cost per task is each model's list price applied to the tokens it used. Cursor says small differences may not be statistically meaningful. GPT-6 models are not on it yet.
Four things stand out on CursorBench. Haiku 5.5 at max effort scores 48.4% for $1.12 a task, level with Sonnet 5.5 at high effort, 47.8% for $1.20. Sonnet 5.5 at extra high, 53.1% for $2.81, matches Opus 5.5 at medium, 52.5% for $2.91. So a smaller model thinking harder is a real alternative to a bigger model thinking less, at the same price.
Raising effort has a point past which it buys nothing. Opus 5.5 scores 56.0% at both high and extra high effort, but the extra high task costs $6.98 against $3.97, 76% more. Gemini 4 Argon moves only about 2 points from low to high, 50.6% to 52.9%, for $4.09 to $4.57, so its lower settings are the better buy. Claude Code's default for Opus 5.5, Sonnet 5.5 and Haiku 5.5 is medium; its models page shows how to change it with /effort.
The 4th is the cheap model that is not cheap. Gemini 3.8 Flash costs $0.75 per million input tokens until 31 December 2026, under half Sonnet 5.5's $2, yet its CursorBench tasks cost $4.70 at high effort against Sonnet 5.5's $1.20. Flash took 324 steps per task where Sonnet took 41, and every step of an agent re-reads the conversation so far. What an AI agent costs per month works through why the step count dominates.
Is GPT-6 Astra worth 10 times the price?
For most coding, no. Astra has the top FrontierSWE score, 65.5% against Opus 5.5's 62.3%, but the 2 ranges overlap for most of their width, and an Astra project cost $1,029.65 against $98.87 for Opus 5.5. On FrontierCode, Opus 5.5 at medium effort scores 54.6% and Astra at maximum effort 53.3%.
Part of the gap is the price list. Astra costs $10 per million input tokens and $50 per million output, 2.5 times Opus 5.5's $4 and $20, and above 272,000 input tokens its price rises to $20 and $75. A long agent session can pass that size, and from then on every step pays the higher rate. The rest is how many tokens each model spends on a task.
There is a case for Astra on a multi-hour job where its slightly higher score is worth more than the bill, and for teams already in Codex. Inside OpenAI's own line, GPT-6.1 Sol scores 50.2% on FrontierCode at $2 and $10 per million, the same token price as Sonnet 5.5. Claude Opus 5.5 vs GPT-6 Astra has every price of the 2 side by side.
Where do Fable 5.1 and the Gemini models fit?
Fable 5.1 is Anthropic's model for the largest, longest tasks, according to Claude Code's docs, and costs $10 and $50 per million tokens, the same as Astra. On all 3 coding benchmarks it scores below Opus 5.5: 51.8% against 57.8% at max effort on CursorBench, at $17.28 a task against $13.43. For coding today, Opus 5.5 is the better buy.
Gemini 4 Argon is 3rd on CursorBench score: 52.9% on CursorBench for $4.57, near Sonnet 5.5 at extra high. Its FrontierSWE cost of $129 is inflated, Proximal notes, by a cache setting that is not used in production. Its API price doubles on 1 December 2026 to $4 and $20 per million, the same as Opus 5.5, which the flagship comparison works through. Gemini 3.8 Flash trails on every coding benchmark here.
Which model should you pick for your coding agent?
In Claude Code, keep Sonnet 5.5 for day-to-day work and switch to Opus 5.5 at high effort for a hard bug or a large refactor; that matches the sonnet and opus advice in Claude Code's own docs and the benchmarks above. Raise Sonnet to extra high before you reach for Opus, and try Haiku 5.5 at max effort on routine edits.
In Codex, the choice is inside OpenAI's line: GPT-6.1 Sol for most work, Astra when one long job justifies its price. In Cursor, which runs models from several makers, CursorBench's own cost column is the guide. Claude Code vs Cursor vs Codex compares the 3 tools themselves.
On a subscription, cost per task becomes how fast you reach your limit. Why Claude Code eats your limit explains how.
How to find the best model for your own repository
Score per dollar on your own work is the number that decides, and you can measure it in an afternoon. Take 10 issues your team has already closed, with the fix that was merged. Run each through your agent with 2 or 3 models at 2 effort levels, starting from the commit before the fix, and record whether the result would have passed your review and what the run cost.
Divide the cost of all runs by the number of fixes you would have merged, and you have cost per merged fix for each setting. The cheapest setting that reaches your bar wins, just as in the chart above. Rerun it when a new model comes out; which model to use and the agent cost simulator estimate the bill before you run anything.
Questions
- What is the best LLM for coding in 2026?
- Claude Opus 5.5 and GPT-6 Astra lead the independent coding benchmarks in October 2026, with Claude Sonnet 5.5 close behind. On FrontierCode, which grades whether a patch is good enough to merge, Opus 5.5 scores 54.6%, GPT-6 Astra 53.3% and Sonnet 5.5 52.1%. On FrontierSWE's multi-hour projects GPT-6 Astra scores 65.5% and Opus 5.5 62.3%, but Astra's tasks cost about $1,030 each against Opus 5.5's $99.
- Which model should I use in Claude Code?
- Sonnet 5.5 for most work and Opus 5.5 at high effort for hard problems. On CursorBench, Opus 5.5 at high effort scores 56.0% for $3.97 a task, the same score as its extra high setting at $6.98, and Sonnet 5.5 at extra high scores 53.1% for $2.81. Fable 5.1 scores lower than Opus 5.5 on every coding benchmark here and costs more per task.
- Is GPT-6 Astra worth it for coding?
- For most coding, no. It has the top score on FrontierSWE, 65.5% against Claude Opus 5.5's 62.3%, but the 2 ranges overlap by most of their width and an Astra task costs about 10 times as much. On FrontierCode, Opus 5.5 at medium effort scores slightly higher than Astra at maximum effort. Astra costs $10 per million input tokens and $50 per million output, and its input price doubles to $20 above 272,000 tokens.
- Does the reasoning effort setting matter for coding?
- As much as the model. On CursorBench, Claude Haiku 5.5 at max effort scores 48.4% for $1.12 a task, about what Sonnet 5.5 scores at high effort for $1.20. Raising Opus 5.5 from high to extra high adds nothing to its score and 76% to its cost. Gemini 4 Argon moves only about 2 points from low to high effort, so its lower settings are the better buy.
- Is a cheap model cheaper for coding?
- Not always. Gemini 3.8 Flash costs $0.75 per million input tokens until 31 December 2026, against Claude Sonnet 5.5's $2, yet on CursorBench a Flash task at high effort costs $4.70 and a Sonnet 5.5 task at high effort $1.20, because Flash took 324 steps a task against Sonnet's 41. Each step re-reads the conversation, so price per task, not per token, is the number that counts.
- Which benchmark is best for comparing coding models?
- Use more than one, because each measures a different job. FrontierCode grades whether a fix to a real open-source issue is good enough to merge. CursorBench runs ambiguous multi-file tasks from real Cursor sessions and reports cost per task at each effort level. FrontierSWE gives hard engineering projects a 20-hour budget.
Sources