DeepSeek V4-Pro Leaves Preview: What Changed

DeepSeek moved V4-Pro out of preview on 13 August 2026, ending a preview period that started in April. The GA release note in DeepSeek's own API docs lists three changes that matter if you are building on it: significantly upgraded agent capability, three selectable thinking effort levels, and native support for OpenAI's Responses API format. A pricing change lands separately, at 16:00 UTC on 16 August, moving the API to peak and off-peak rates where off-peak is half the peak price.

The release arrived with no launch event. It surfaced as a changelog entry, which is roughly how DeepSeek has shipped everything this year.

What actually changed

The model identifier does not change. Calls that already target deepseek-v4-pro pick up the GA version with no code edit, which is convenient and also the reason a lot of teams will not notice they switched.

The substantive changes:

The benchmark figures come from DeepSeek's own table. Treat them the way you would treat any vendor's self-reported numbers, which is to say as a claim to test rather than a fact to plan around. If you want a method for that, we wrote up how to spot an inflated AI benchmark claim.

Why the effort levels matter more than the benchmarks

The three-level thinking control is the part that will change your bill.

Reasoning tokens are billed as output tokens on every major provider, so a model that thinks harder costs more per identical question. Giving you low, high and max rather than an on-off switch means the cost of a request is now partly a routing decision you make, not a fixed property of the model you chose. A classification job that ran at full reasoning depth for the last four months was probably paying two to five times what it needed to.

DeepSeek is not alone here. OpenAI's reasoning guide describes effort values that "can include none, minimal, low, medium, high, xhigh, and max" depending on the model, and Anthropic's docs put depth control in an output_config.effort field on its newer models. Three vendors, three vocabularies, same underlying dial. We unpacked the whole pattern in what reasoning effort actually controls.

The pricing change is the part with a deadline

Peak and off-peak pricing is unusual among frontier API providers, and it rewards a specific kind of workload: batch jobs that do not care what hour they run.

If you run nightly document processing, scheduled summarisation, evaluation sweeps, or any queue that tolerates a delay, the off-peak rate halves that line item for the cost of a cron expression. If you serve interactive traffic during business hours in Europe or North America, you get no discount and should budget at peak.

Two practical notes. First, the rates take effect at 16:00 UTC on 16 August, so any cost model built before then is stale after tomorrow afternoon. Second, DeepSeek's release note states the off-peak-is-half-of-peak relationship but the specific per-token figures live in the pricing table on its docs, not in the announcement text, so check the live page rather than a secondary summary before you commit numbers to a spreadsheet.

Should you switch

Not on the strength of a changelog. The honest answer for most teams is that a GA label is a stability signal, not a quality signal, and the only evidence that matters is your own workload.

A reasonable sequence:

1. Run your existing eval set against deepseek-v4-pro at high effort and record cost, latency and pass rate.

2. Re-run the same set at low. If quality holds, that is your new default and the saving is immediate.

3. Only reach for max on the cases that failed at high, and check whether the extra tokens actually converted into a correct answer or just a longer one.

4. Compare against whatever you run today on the same eval, not on a public leaderboard.

That last point does most of the work. Our guide to testing a new AI model before switching has the fuller version, and if you are wondering how often this exercise is worth repeating, how often you should switch AI models argues for less frequently than the release cadence suggests.

What to watch next

The Responses API support is the quiet strategic move. When an independent lab ships native compatibility with a competitor's request format, it is lowering the cost of leaving that competitor. Model switching gets easier every time a provider does this, and easy switching is good for everyone who is not a frontier lab. Alibaba's Qwen3.8-27B open-weight release landed the same week, another reminder that open-weight models are shipping on a similarly fast cadence right now.

Sources for this piece: DeepSeek's GA release note and its API changelog, both dated 13 August 2026 and checked on 15 August 2026.