DeepSeek V4 Pro launched cheap, then quadrupled its price three days later

What happened

On August 13, 2026, DeepSeek moved its flagship V4 Pro model out of preview and into general availability as the "0813" checkpoint, following an April preview release. Launch pricing was $0.435 per million input tokens and $0.87 per million output tokens, with cached input at $0.003625 per million — well below what OpenAI, Anthropic, or xAI charge for frontier-class output.

It did not stay there. On August 16 at 16:00 UTC, three days after GA, DeepSeek switched to a peak/off-peak billing structure. Off-peak rates moved to $0.66 per million input and $1.98 per million output. Peak rates hit $1.32 input and $3.96 output. Depending on token type and time of day, that's an increase of anywhere from 50% to over 1,000% on the launch price, according to reporting from Tech Times and Enterprise DNA.

What is genuinely new

The benchmark gains, if they hold, are real jumps versus the April preview. DeepSeek reports DeepSWE moving from 12.8 to 62.7, CyberGym from 52.7 to 83.3, and Terminal Bench 2.1 from 72.1 to 87.9. Those are agentic coding and terminal-use benchmarks, not general trivia tests, and a jump that size suggests real work went into the tool-use and multi-step reasoning that agent builders actually care about. The model also ships with support for the OpenAI ChatCompletions format and the Anthropic Messages format alongside DeepSeek's own API, which lowers the switching cost if you're already building against one of those.

Scale is the other genuinely new part of this story. DeepSeek was reportedly second only to Anthropic in raw API token volume in July, per usage data cited by Wccftech, and is closing the gap. This isn't a niche open-weight release anymore — it's a model with real production traffic behind the benchmark chart. Community estimates put it at 1.6 trillion total parameters with 49 billion active, though DeepSeek hasn't confirmed a parameter count or published a system card, so treat that figure as circulating, not official.

What it means for a business owner

If you're weighing which model powers a coding agent, a document-processing pipeline, or any workflow that burns a lot of output tokens, V4 Pro's benchmark-to-price ratio is worth evaluating even at the new, higher rate — it still undercuts the frontier labs by a wide margin. The billing structure is what actually deserves your attention, though. Off-peak hours are defined against UTC in a way that happens to line up with the US and European business day, so an automation that runs during the American workday may land in the cheaper tier by default. Run something as an overnight batch job, or run it from a business whose working hours map onto China's peak window, and the math changes considerably.

The honest caveat

Two things should slow you down before you commit. First, the benchmark numbers are DeepSeek's own. Nothing here has been reproduced by an independent evaluation yet, and a 50-point jump on a single benchmark between two checkpoints is the kind of result that deserves your own test before it touches production traffic. Second, and more practically: a lab that quadruples its price three days after a GA launch has told you something about how stable its pricing will be. The number that got this model attention this week is not the number you should build a twelve-month cost model around.

What to do about it

Don't budget off the launch price or the vendor benchmark chart. If V4 Pro is a candidate for something you're automating, run your own eval against your actual task, price it against your actual usage pattern across both peak and off-peak windows, and revisit that number in a month. A model that looks ten times cheaper on day one can look very different by the time you've built a pipeline around it.