Top pick
Claude Opus 5
$199 / month
Lower bill among the top cited coding scores: SWE-Bench Pro 79.2 and SWE-Bench Verified 96 (anthropic.com).
Score 101.0 · $5.00 input · $25.00 output per 1M · context 1,000,000 · 39 providers
Job
Coding agents
Which model should I use for a coding agent, and what will it cost each month?
DRAFT — placeholder editorial, not approved. The score uses cited agentic-coding benchmarks from the lab or the benchmark owner, plus tool-calling signals. Models within a few points are treated as one flagship band and the lower bill is the top pick. A human has not signed off on it.
How this pick is chosen
- Outputs text. Embeddings and image-only models are not in the catalog.
- Lab is one of: openai, anthropic, google, meta, mistral, deepseek, xai, alibaba, moonshotai, zhipuai, cohere, amazon, nvidia, microsoft, ai21.
- Listed by at least 5 providers. A model on only the lab and a reseller does not clear this.
- Not deprecated, and not an alpha release. Beta releases stay in.
- Not a specialized model: OCR, embeddings, rerankers, speech, safeguard, or moderation, matched on the name.
- Supports tool calls.
- Supports reasoning.
- Context window of at least 128,000 tokens.
- Has a numeric input and output price.
- Reference price is not $0. A zero in Models.dev is often an unpublished price.
- A benchmark counts only when its source is the model's lab (including that lab's Hugging Face org) or the benchmark owner. Artificial Analysis, OpenRouter, and LLM Stats are dropped. When one benchmark has several cited scores, the median is used.
- Agentic-coding score: the percentile of the hardest cited software-engineering benchmark the model has, in this order: SWE-Bench Pro, SWE-Bench Verified, Terminal-Bench, DeepSWE, Aider Polyglot, then SWE-Atlas Codebase QnA. A fallback bench is capped (0.94, 0.86, 0.80, 0.80, 0.78) so a lone terminal score cannot beat a high SWE-Bench Pro. Other cited coding benches add up to 6 points. Structured output adds 3 tool-calling points, an explicit false removes 3, and a cited Toolathlon or MCP Atlas result adds 2. Preview and snapshot names lose 5. A name with no version digit loses 8.
- The top pick is the highest job score. On coding agents, models within 4 points of that leader are treated as a flagship tie: the lower monthly bill is the top pick and the higher bill is the premium pick. Otherwise the premium pick is the best upgrade that costs more than the top pick. The budget pick is the highest job score that costs less than the top pick and sits at or below the 25th cost percentile. Output rates above $50 per 1M tokens stay out of the premium card. Summarization uses raw quality, not quality per dollar, when it chooses that upgrade. Remaining ties break to more providers, then a lower monthly cost, then a newer release date.
A working day of an agent: 60 turns, a long repo in the prompt, and short patches back. 40% of the input is assumed cached.
154 models cleared the bar.303 failed the lab, provider, or specialty bar.61 failed a capability filter.1 listed at $0 were left out.0 had no usable price.
Budget pick
GPT-5.6 Luna
$8.86 / month
Lower bill with a cited coding result: SWE-Bench Pro 62.7 and Terminal-Bench 84.7 (openai.com).
Score 79.0 · $0.20 input · $1.20 output per 1M · context 1,050,000 · 44 providers
Premium pick
Claude Fable 5
$398 / month
Higher bill at the top of the cited coding table: SWE-Bench Pro 80.3 and Terminal-Bench 88 (anthropic.com).
Score 104.1 · $10.00 input · $50.00 output per 1M · context 1,000,000 · 43 providers
When not to use it
DRAFT — placeholder. Do not treat this as a reason to avoid the pick. The real caveat, once written, should say when a cheap tool-calling model is the wrong agent, for example when you need a lab's best coding quality and will pay for it.
DRAFT — placeholder. This note will say who the workload is for. Until then, change the calculator inputs to your own traffic.
The three cards follow the inputs. Fit, and who holds each role, stays on the reference workload.
Everyone who cleared the bar, best fit first
| Score | Model | Monthly | Input | Output | Context | Providers |
|---|---|---|---|---|---|---|
| 104.1 | Claude Fable 5 | $398 | $10.00 | $50.00 | 1,000,000 | 43 |
| 101.0 | Claude Opus 5 | $199 | $5.00 | $25.00 | 1,000,000 | 39 |
| 95.8 | Claude Opus 4.8 | $199 | $5.00 | $25.00 | 1,000,000 | 49 |
| 93.1 | Qwen3.8 Max | $62.10 | $2.00 | $6.00 | 1,000,000 | 32 |
| 92.8 | Muse Spark 1.3 | $41.00 | $1.25 | $4.25 | 1,048,576 | 16 |
| 92.0 | GPT-5.6 Sol | $159 | $4.00 | $20.00 | 1,050,000 | 45 |
| 90.3 | Grok 4.5 | $62.64 | $2.00 | $6.00 | 500,000 | 26 |
| 88.2 | Kimi K3 | $119 | $3.00 | $15.00 | 1,048,576 | 80 |
| 86.7 | Qwen3.8 Max Preview | $62.10 | $2.00 | $6.00 | 1,000,000 | 8 |
| 85.0 | GPT-5.6 Terra | $88.56 | $2.00 | $12.00 | 1,050,000 | 42 |
| 80.7 | DeepSeek V4 Pro 0813 | $19.84 | $0.66 | $1.98 | 1,000,000 | 42 |
| 80.6 | Claude Opus 4.7 | $199 | $5.00 | $25.00 | 1,000,000 | 41 |