benchr
One point separates the cited SWE-bench figures, while the listed API prices differ by roughly an order of magnitude. Neither difference identifies a universal winner: the relevant question is how each model performs on a held-out sample of your workload under the same prompts, tools, retry policy, and latency budget.
Side-by-side specs
| Dimension | DeepSeek V4-Pro | Claude Sonnet 4.6 |
|---|---|---|
| Input / 1M (peak) | $1.32 | $3.00 |
| Input / 1M (off-peak) | $0.66 | $3.00 |
| Output / 1M (peak) | $3.96 | $15.00 |
| Output / 1M (off-peak) | $1.98 | $15.00 |
| Cache-hit input / 1M | $0.044 | $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 |
Monthly cost for an example pipeline
At the rate-card snapshot used here, a document pipeline with 50M input tokens and 8M output tokens a month would cost $150 for Sonnet input and $120 for output — $270/month — versus $21.75 + $6.96 = $28.71 on V4-Pro. At ten times that volume, the arithmetic becomes about $2,700 versus $287 before discounts, failed calls, retries, infrastructure, or taxes. DeepSeek lists a lower cache-hit input rate for eligible repeated prefixes; actual savings depend on cache eligibility and hit rate.
What the price sheet cannot decide
SWE-bench and list prices do not measure long-running tool reliability, recovery from malformed calls, latency under load, or fit with an organization's controls. Run both models through the same agent traces and score task completion, recovery behavior, retries, and cost per accepted result. Separately review each vendor's current data-processing terms, retention controls, deployment regions, security documentation, and multimodal support. Do not infer safety or compliance from the model's country, license, or hosting arrangement alone.
Self-hosting considerations
DeepSeek publishes V4-Pro weights under the MIT license, so an organization can evaluate self-hosting instead of sending prompts to the hosted API. That replaces per-token fees with hardware, security, monitoring, and operations work — the local-inference piece walks through the cost categories. Self-hosting can support some data-residency goals, but it does not by itself establish regulatory compliance.