Choose a coding API by the changes it can complete and verify in your project. Use documented capabilities to shortlist candidates, then measure accepted patches and total work.
What matters for coding
Define the coding task narrowly: explain existing code, propose a patch, write a migration or investigate a failing check. These are different output contracts. A model that produces a helpful explanation may still fail to create a patch that applies and preserves existing behavior.
Use a safe repository fixture with a clear acceptance check. Include an ordinary edit and a case that requires recognizing missing context. Keep the same relevant files, instructions and allowed tools across candidates. Do not judge only the narrative around the patch.
Kimi’s catalog includes coding-focused variants, and Qwen Coder documents coding-oriented models with their own operating conditions. Official documentation.
Record the number of attempts and the work needed after the first answer. A lower output price can be useful, but repeated repair calls and invalid changes can change cost per accepted task. Include the application’s checking and retry policy in the evaluation design.
Ranked candidates
Only active candidates with documented values for this ranking are included. Prices retain the tier and deployment condition shown below.
| Model | Provider | Input USD / 1M | Output USD / 1M | Context tokens | Price condition | Cost |
|---|---|---|---|---|---|---|
| deepseek-flash | DeepSeek | $0.3 | $1.2 | 1,000,000 | peak | Estimate cost |
| deepseek-v4-pro | DeepSeek | $1.32 | $3.96 | 1,000,000 | peak | Estimate cost |
| qwen3.8-max | Alibaba Cloud | $2 | $6 | 1,000,000 | 0<Token≤1M · international | Estimate cost |
| Claude Opus 5 | Anthropic | $5 | $25 | 1,000,000 | Standard | Estimate cost |
| gpt-6-astra | OpenAI | $10 | $50 | 1,050,000 | Standard | Estimate cost |
Last verified · Source ↗
The list filters documented coding capability and exposes output price and context for comparison. It is not a leaderboard of patch correctness. Inspect missing feature evidence and the required request interface before adding a candidate to the evaluation.
Our three picks
Find your starting point
Your text and estimates stay in this browser. No API requests are sent to model providers.
Loading verified model records…
The cheapest role should pass the narrow coding fixtures before receiving routine edits. Balanced can be useful when a candidate handles a broader range of context without excessive repair. Strongest requires comparable task evidence; absent that, treat the third role as an explicitly unproven evaluation candidate.
Keep tool authorization outside the model’s prose. A model can propose a command or patch, while the application decides what is permitted and verifies the outcome. Preserve the result of each executed check so the final answer does not claim a validation that never occurred.
When each pick is wrong
The cheapest candidate is wrong when it reliably misses project constraints or produces changes that fail the checks. A larger candidate can be wrong when a narrow formatter or extractor already solves the task reliably with a smaller model. Match the capability to the actual work.
A long-context candidate is wrong when the application simply sends irrelevant files instead of identifying necessary context. Evaluate a focused request and a broader one with the same acceptance criteria. More material is not automatically better evidence.
Review Claude model choices, Kimi coding families and Qwen candidates as provider-specific investigation paths.
Estimate cost
Estimate your API costs
Your text and estimates stay in this browser. No API requests are sent to model providers.
Loading verified model records…
Use the coding workload calculator with repository input, generated patches and additional attempts needed for accepted changes.
Retain elapsed time and validation outcomes alongside returned usage. Choose a fallback only after it passes the same mandatory checks, then preserve the actual model used when recovery occurs.
Last verified · Source ↗
Frequently asked questions
Does coding-capable metadata prove a model will fix my bug?
What should a coding evaluation accept?
Why inspect output cost?
Does longer context always improve coding?
Who should authorize tool execution?
What belongs in the cost record?
Sources
Last verified · Source ↗