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 | $0.435 | $1.25 |
| Output / 1M | $0.87 | $10.00 |
| Cache-hit input / 1M | $0.0036 | — |
| 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 |
Output price is the whole ballgame
Code generation is output-heavy by nature: small prompt, long completion. Take an agent that generates code at 4K input / 12K output per task, 50,000 tasks a month. GPT-5: $250 input + $6,000 output = $6,250/month. V4-Pro: $87 + $522 = $609/month. The output line is 96% of the GPT-5 bill. That 384K max-output ceiling matters here too — V4-Pro can emit an entire module in one call where GPT-5's 128K cap forces splitting on the largest generations. Check your own mix in the cost calculator.
Where GPT-5 keeps the job
Ecosystem gravity is real. Most agent frameworks, eval harnesses, and SaaS integrations are built and tested against OpenAI's API first, and GPT-5's function-calling behavior has had a year of hardening that shows in fewer malformed tool calls. If your platform's failure mode is "an agent got stuck at 2 a.m.," the cheaper model that needs more retry scaffolding isn't cheaper anymore. And for organizations that can't route data to DeepSeek's hosted API, the choice is GPT-5 or self-hosting V4's MIT weights — a real option, but an infrastructure project, not an API swap. The DeepSeek vs OpenAI pricing deep-dive covers the quality-per-dollar curve across both lineups.
The honest framing
GPT-5 isn't overpriced for what it is — it's the most integration-friendly model at its tier, as the benchr review lays out. But on the two axes this page measures, price and published coding benchmark, V4-Pro wins both, and that's a sentence that wasn't true of any open-weight model a year ago. The question to answer before switching is operational, not statistical: who's on call when the loop breaks, and where is your data allowed to go?