benchr
Most model comparisons are trade-offs. This one looks like a steal at first glance: the cheaper model also posts the higher coding score. That's exactly why it deserves a careful read — when a price sheet looks too good, the costs have usually moved somewhere the sheet doesn't measure.
Side-by-side specs
| Dimension | DeepSeek V4-Pro | GPT-5 |
|---|---|---|
| Input / 1M (peak) | $1.32 | $1.25 |
| Input / 1M (off-peak) | $0.66 | $1.25 |
| Output / 1M (peak) | $3.96 | $10.00 |
| Output / 1M (off-peak) | $1.98 | $10.00 |
| Cache-hit input / 1M | $0.044 | — |
| Context window | 1,000,000 | 400,000 |
| Max output | 384,000 | 128,000 |
| SWE-bench Verified | 80.6% | 74.9% |
| License | MIT (self-hostable) | Proprietary API only |
How output volume changes the bill
For an output-heavy example, take an agent using 4K input and 12K output tokens per task across 50,000 monthly tasks. GPT-5 totals $250 input + $6,000 output = $6,250/month. V4-Pro totals $87 + $522 = $609/month. Output accounts for 96% of the GPT-5 bill in this scenario.
V4-Pro also lists a 384K maximum output, compared with GPT-5's 128K limit. Whether that removes useful splitting depends on the workload. Check your own mix in the cost calculator.
Reasons to evaluate GPT-5
Many agent frameworks, evaluation harnesses, and SaaS integrations support OpenAI API shapes. This page does not cite a controlled comparison of malformed tool calls, retries, or multi-step reliability, so test both models against your schemas and failure handling.
Also review deployment and data-residency requirements. Self-hosting V4's MIT-licensed weights is an infrastructure project, not an API swap. The DeepSeek vs OpenAI pricing guide covers both lineups.
A practical evaluation plan
On the two axes quantified here, V4-Pro has the lower listed price and the higher cited coding score. The decision still depends on integration effort, failure handling, support, and where your data is allowed to go. Use the GPT-5 review for its broader documented context.