benchr
The table below lists selected public token rates from the two providers. The rows are not asserted to be equivalent quality tiers; model choice also depends on measured task success, latency, reliability, features, governance, and support.
| Model | Provider | Input / 1M | Output / 1M | Cached Input / 1M |
|---|---|---|---|---|
| GPT-5.5 | OpenAI | $5.00 | $30.00 | $0.500 |
| GPT-5 | OpenAI | $1.25 | $10.00 | $0.125 |
| GPT-5 Mini | OpenAI | $0.250 | $2.00 | $0.025 |
| DeepSeek V4-Pro (peak) | DeepSeek | $1.32 | $3.96 | $0.044 |
| DeepSeek V4-Pro (off-peak) | DeepSeek | $0.660 | $1.98 | $0.022 |
| DeepSeek V4-Flash (peak) | DeepSeek | $0.440 | $1.32 | $0.014 |
| DeepSeek V4-Flash (off-peak) | DeepSeek | $0.220 | $0.660 | $0.007 |
DeepSeek now charges by the clock
Since 16:00 UTC on August 16, 2026, DeepSeek bills two rates. Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday; everything else is half price. OpenAI has no equivalent time-of-day rule, so a like-for-like comparison now has to say which DeepSeek rate it means. The peak rate is the honest default for anything user-facing, because a product serves users when users are awake. The off-peak rate is the honest default for batch work you control the schedule of.
Price, benchmark, and operational caveats
Before changing providers, separate documented rates from provider benchmark claims and from measurements you run yourself:
- Provider-reported coding results: DeepSeek's V4-Pro model card reports 80.6% on SWE-bench Verified; OpenAI's GPT-5 system card reports 74.9%. These figures come from separate provider publications and may use different settings or evaluation conditions. They are not an independent matched result and do not, by themselves, prove that one model outperforms the other on production work.
- Latency and reliability: This article has no controlled primary-source comparison of time to first token, throughput, or availability between the providers. Measure those properties under your own concurrency and region, and review the applicable service terms or SLA.
- Prompt caching: GPT-5's listed $0.125 cached-input rate is 90% below its $1.25 standard input rate. DeepSeek V4-Pro's cached input is $0.044 at peak, about 97% below its $1.32 peak input rate, and V4-Flash's is $0.014, about 97% below $0.44. Both cached rates halve off-peak. The proportional discount survived the August repricing; the absolute rate rose about twelvefold for V4-Pro and fivefold for V4-Flash. Real savings depend on cache eligibility and the hit rate achieved.
For more detailed rankings, see our AI Model Rankings or compute your exact monthly cost with the Cost Calculator.
What the price gap does not show
The DeepSeek/OpenAI comparison is not only a token-price question. The choice also depends on procurement rules, data handling, latency, ecosystem fit, observability, and whether your team already uses OpenAI features such as structured outputs, tool calling, or internal eval infrastructure.
For greenfield, cost-sensitive workloads, DeepSeek's price advantage is large enough to test seriously. For enterprise workloads already tied to OpenAI governance, the switching cost may outweigh the token savings unless the workload is large, repetitive, and easy to evaluate automatically.
How to run the comparison fairly
Do not compare one polished OpenAI prompt against a first-draft DeepSeek prompt. Give both providers the same examples, the same tool schemas, the same output contract, and the same evaluation rubric. Then measure accepted outputs, not just model preference.
Also include operational checks: SDK maturity, error handling, streaming behavior, rate limits, observability, and how easily your team can debug failed calls. Token savings matter, but a brittle integration can erase them through engineering time.
Finally, compare support paths. A cheaper model is easier to adopt when failures are visible and recoverable. If the workload needs vendor support, audit logs, fine-grained access controls, or preapproved compliance language, include those requirements in the scorecard before the token price decides the winner.
Use a staged rollout rather than a hard switch. Send a small share of low-risk traffic to the cheaper provider, compare accepted outputs and operational incidents, then expand only where the evidence holds. This avoids turning a pricing experiment into a reliability incident.
For finance teams, the cleanest comparison is a spreadsheet with three columns: public token price, measured acceptance rate, and operational overhead. If DeepSeek wins all three, the decision is easy. If it wins only token price, the migration should stay limited to the workflows where the evidence is strongest.