Best AI Model for Coding by Price: DeepSWE Scores vs Cost
Independent coding-benchmark scores next to the monthly cost of running a coding agent, plus when a flat-rate coding tool beats paying per token.
Updated September 19, 2026
Coding agents burn tokens faster than any other workload: they read files, re-read them after edits and keep a long history in context. That makes the price per token matter as much as the model’s skill. This guide ranks the models we track by their score on DeepSWE, a benchmark of real software-engineering tasks, and puts the cost of running an agent next to it.
Coding score against agent cost
The monthly figure prices 10,000 agent steps a month, ~40,000 input tokens each (70% read from the prompt cache) and ~2,000 output tokens. Scores come from Epoch AI, which runs the benchmarks independently of the model vendors.
| Model | Provider | DeepSWE | Blended price | Agent cost / month |
|---|---|---|---|---|
| GPT-6 Astra | OpenAI | 74.1% | $20.00 | $2,480 |
| Gemini 3.8 Flash | 73.8% | $1.50 | $186.00 | |
| Claude Opus 5 | Anthropic | 73.6% | $10.00 | $1,240 |
| GPT-5.6 Sol | OpenAI | 72.7% | $8.00 | $992.00 |
| GPT-5.6 Terra | OpenAI | 69.6% | $4.50 | $536.00 |
| Grok 4.6 | xAI | 67.5% | $3.00 | $500.00 |
| GPT-5.6 Luna | OpenAI | 67.2% | $0.45 | $53.60 |
| Claude Sonnet 5 | Anthropic | 53.8% | $4.00 | $496.00 |
| Gemini 3.5 Flash | 37.4% | $3.375 | $402.00 | |
| Gemini 3.1 Pro | 11.7% | $4.50 | $536.00 |
The short answer
- Highest score: GPT-6 Astra at 74.1%, for $2,480 a month on this workload.
- Most score per dollar: GPT-5.6 Luna — 67.2% at $0.45 per 1M tokens blended, or $53.60 a month.
Why the cheapest model often wins here
Agent work is repetitive. Most steps are reading a file, running a test or applying a small patch, and a mid-tier model handles those as well as a flagship. The expensive model earns its price on the hard step: an unfamiliar codebase, a subtle bug, a refactor across many files. Many teams run a cheap model by default and keep a flagship for the steps that fail.
Prompt caching matters more than the sticker price
An agent sends the same file contents over and over. With caching, that repeated context is billed at the cached rate, which is a fraction of the normal input price — the workload above assumes 70% of input comes from the cache. Without caching the same agent costs several times more, so check that your framework actually enables it before comparing providers.
API or a coding subscription?
If one developer runs an agent all day, a flat-rate plan is often cheaper than tokens. The coding tools page lists what Cursor, GitHub Copilot, Claude Code, Codex and the rest charge per month, so you can compare that against the table above.