benchr
This is the comparison that decides whether the open-weight tier has earned your production traffic. One point of SWE-bench separates these two — in DeepSeek's favor — while an order of magnitude separates the bills. The interesting question isn't which model is better. It's whether the things Sonnet still does better are the things your workload needs.
Side-by-side specs
| Dimension | DeepSeek V4-Pro | Claude Sonnet 4.6 |
|---|---|---|
| Input / 1M | $0.435 | $3.00 |
| Output / 1M | $0.87 | $15.00 |
| Cache-hit input / 1M | $0.0036 | $0.30 |
| Context window | 1,000,000 | 1,000,000 |
| Max output | 384,000 | 64,000 |
| SWE-bench Verified | 80.6% | 79.6% |
| GPQA Diamond | 90.1% | 89.9% |
| License | MIT (self-hostable) | Proprietary API only |
The bill, on a real pipeline
A document-processing pipeline that ingests 50M input tokens and produces 8M output tokens a month: Sonnet 4.6 charges $150 for input and $120 for output — $270/month versus $21.75 + $6.96 = $28.71 on V4-Pro. Multiply both by ten for a serious pipeline and the gap is $2,700 against $287. DeepSeek's cache-hit rate ($0.0036/1M) makes repeated-prefix workloads cheaper still — effectively free input on stable prompts.
What Sonnet still buys you
Three real things, none of them on the price sheet. First, agent-loop discipline: in long multi-step tool use, Claude recovers from errors with less scaffolding — the gap the open-weight survey calls a lead rather than a chasm, but a lead all the same. Second, the compliance story: Anthropic's API terms, vision support, and safety tuning fit enterprise review processes that DeepSeek's hosted API often doesn't pass. Third, max-context honesty under load — though DeepSeek now publishes its own long-context numbers (83.5% MRCR at 1M), which is more than most rivals do.
The self-host escape hatch
If the objection to DeepSeek is "our data can't go to that API," the MIT license is the counter: download the weights and serve V4 yourself. You trade per-token fees for GPU costs and ops work — the local-inference piece walks the math. Sonnet has no equivalent. For regulated environments, that one row of the table can decide the whole comparison.