Checkpoints V4-Pro-0813 and V4-Flash-0731 · read 15 August 2026
V4-Flash vs V4-Pro
On the current checkpoints these two models are one point apart on capability, and 3.0× apart on price. Flash is also faster and allows 5.0× the concurrency. Stated that way the choice looks obvious, and for most workloads it is — but the rate card is only half of what you pay.
The other half is how many tokens each model spends getting to an answer, which no comparison table contains. This page puts a number on where that crossover sits: 3.1× the token consumption, and the cheaper model stops being cheaper.
Headline differences
Capability gap
53 vs 52
Pro ahead by 1 — inside the noise for most work
Price gap
3.0×
Pro costs that much more per output token
Break-even
3.1×
Token consumption at which the bills meet
Why most comparisons you’ll find are wrong
Both model names have meant different things at different times, and the ordering between them has flipped twice. Between 31 July and 13 August, Flash genuinely outscored the V4-Pro preview by a wide margin — which is when most of the published comparisons were written. V4-Pro-0813 closed that gap. Anything claiming a large lead in either direction is describing a snapshot that no longer holds.
-
2026-04-24
V4-Pro and V4-Flash ship together
Both released the same day under MIT with open weights, both with a 1M token context window. Flash was trained separately rather than distilled from Pro, which is why it was never simply a smaller copy of it.
-
2026-07-31
Flash is re-post-trained
The 0731 checkpoint sharpened agentic and tool-use behaviour at unchanged prices, and briefly took Flash past the V4-Pro preview on the Intelligence Index. Most published comparisons date from this window.
-
2026-08-13
V4-Pro-0813 reaches general availability
The production checkpoint adds selectable reasoning effort (low / high / max), a native OpenAI Responses API and DSpark speculative decoding — and moves Pro back in front, by one index point. DeepSeek Harness launched the same day.
-
2026-08-16
New rate card takes effect
Prices rise and time-of-day billing returns for both models. The ratio between them is roughly preserved, so the choice between the two is not what the increase changes.
Intelligence Index scores are reported at maximum reasoning effort. Artificial Analysis revised Flash-0731 upward from 50 to 52 when the index moved to v4.1.1, so figures quoted elsewhere may predate that revision.
Side by side
| Measure | DeepSeek V4-Pro | DeepSeek V4-Flash |
|---|---|---|
| Intelligence Index Artificial Analysis, max reasoning effort | 53 | 52 |
| Output speed Artificial Analysis median | 77.6 tok/s | 120.6 tok/s |
| Output price, off-peak Same ratio at peak | $1.98/M | $0.66/M |
| Concurrent requests Account-level cap; extra API keys do not raise it | 500 | 2,500 |
| Context window Identical | 1M | 1M |
| Parameters Trained separately — Flash is not a distillation of Pro | 1.6T total · 49B active | 284B total · 13B active |
Capability and speed from Artificial Analysis; prices, context and concurrency from DeepSeek’s own documentation. For scale, Claude Opus 5 (max) scores 63 on the same index.
Where the discount runs out
A smaller model can need more attempts, longer reasoning or extra tool calls to finish the same job — and one published hands-on comparison across three real coding tasks found exactly that: the cheaper model produced better work while consuming enough extra tokens that the two bills came out roughly level. So the question is not which rate is lower, it is how much more the cheaper model has to spend before the gap closes.
DeepSeek V4-Pro
—
The measured workload, unchanged
DeepSeek V4-Flash
—
—
Equal token use Break-even at 3.1×
—
Baseline is the measured run on this site: 155.9M of prompt at 98.1% cache hit rate and 754K of output, priced on the new card. The multiplier is applied to every meter at once, which assumes the cheaper model's extra work is spread across input and output alike.
The setting that outweighs the model choice
V4-Pro-0813 introduced selectable reasoning effort — low, high and max. Effort controls how many thinking tokens the model generates before answering, and thinking tokens are billed at the output rate, the most expensive meter on the card. You are charged for them and you never see them.
73.6%
of DeepSeek V4-Pro's generated tokens were thinking, at effort high
Measured across 613 requests on one machine. On that run, roughly 74% of the output bill bought reasoning rather than answers. The DeepSeek V4-Flash requests in the same logs sat at 57.2%, but they ran different tasks — treat that as a second data point, not a head-to-head.
The practical consequence: before switching models to save money, check what effort you left the current one on. It is one line of config, it acts directly on the priciest meter, and it does not change which model’s answers you are getting.
Choosing
Default to DeepSeek V4-Flash
- High-volume production traffic. A 1-point index gap does not survive contact with most real workloads, and you are paying 3.0× for it.
- Anything latency-sensitive. 1.6× the output speed is a user-visible difference in a way the index score is not.
- Parallel or batch work. 2,500 concurrent requests against 500 — a 5.0× ceiling difference that decides throughput long before capability does.
Escalate to DeepSeek V4-Pro
- Long autonomous agent loops. Small per-step error rates compound over many steps, and a retry costs more than the capability premium would have.
- Hard multi-file refactors and frontier reasoning. The workloads where the top of the index is genuinely being asked for.
- Accuracy-critical output where a wrong answer costs more than the difference between the two rate cards.
Routing by difficulty — Flash by default, Pro on escalation — beats picking one, and both share a rate card structure so the arithmetic stays simple.
Whichever you pick, the largest single lever on the bill is neither model nor effort: it is your cache hit rate. On the measured run it removed $97.56 from a $104.39 bill — more than any model choice could.
Model choice questions
Does DeepSeek charge extra to write to the cache?
No. Caching is automatic and there is no separate write charge — you are billed at the hit rate or the miss rate, and nothing else.
This is a real difference from some competitors rather than a technicality. Anthropic bills a premium for writing to its prompt cache, which has to be earned back through reads before caching pays for itself. On DeepSeek the first request simply costs the miss rate, as it would have anyway.
The trade-off is that you get no guarantee in return. DeepSeek describes its caching as best-effort, and unused entries are cleared after a few hours to a few days.
What is the difference between V4-Pro and V4-Flash?
Both carry a 1M token context window and the same maximum output length. They differ on price and on how many requests you may run at once: DeepSeek V4-Pro allows 500 concurrent requests, DeepSeek V4-Flash allows 2,500.
DeepSeek V4-Flash output costs $1.32 per million at peak against $3.96 for DeepSeek V4-Pro. Both models moved to the new card on the same date and share the same peak windows.
Is V4-Flash actually cheaper than V4-Pro in practice?
The rate card says 3.0× on output. Your bill may not agree, because price per token is only half of what you pay — the other half is how many tokens the model spends reaching an answer, and the smaller model does not always spend fewer. One published hands-on comparison ran both across three real coding tasks and found the cheaper model produced better work while consuming enough extra tokens to make the two bills roughly equal.
A second factor is usually larger than the model choice: reasoning effort. In the measured run on this site, 73.6% of all generated tokens were thinking tokens — invisible in the response, billed in full at the output rate, which is the most expensive meter on the card.
So the honest answer is that neither the headline discount nor a benchmark table settles it. Run your own workload on both, compare the total token counts rather than the rates, and check what reasoning effort you left it on.
Is DeepSeek still cheaper than Claude, OpenAI and Gemini after the increase?
For most workloads, yes — but by a smaller margin than before, and the gap narrows further against the cheaper tiers of each provider once you account for cache hits.
The right comparison depends on your own mix of input, output and cache rate, which is what the calculator is for: it applies your volumes to every rate card at once instead of comparing headline numbers that assume a workload you may not have.
Benchmark figures are third-party measurements of specific checkpoints and are revised over time; the date each was read is on this page. Prices are from DeepSeek’s published card. Neither is a substitute for running your own workload on both models and comparing the token counts.