DeepSeek V4 Pro launched, then its API prices jumped up to 1,100%, and that number isn't about the tokens you'd think

What actually shipped on 13 August

DeepSeek-V4-Pro-0813 came out of an April 2026 preview to reach general availability on 13 August, available through DeepSeek's website, mobile app and API, with weights published under the MIT license on Hugging Face. DeepSeek's own benchmark table shows large jumps over the preview build: Terminal-Bench climbing from 72.1 to a claimed 87.9, and SWE-bench Verified reaching a claimed 96.40%. Those are DeepSeek's numbers, not independently reproduced ones, and that distinction matters more than usual here.

Independent evaluators have not been kind to the self-reported figures. Artificial Analysis places V4-Pro's Intelligence Index at 53, level with Zhipu's GLM-5.2, four points behind GPT-5.6 Terra and seven behind Kimi K3, not a frontier-leading score. The South China Morning Post reported an independent Terminal-Bench rerun landing around 54.68%, far below DeepSeek's claimed 87.9%, and cited Vals AI ranking the model 12th on its own leaderboard, behind the previous-generation GPT-5.5. None of this makes V4-Pro a bad model. It makes DeepSeek's marketing table the same kind of unreliable that every vendor's self-reported benchmark table tends to be, and worth treating with the same skepticism you'd apply to any company grading its own homework.

The pricing change, exactly

Three days after V4-Pro's launch, at 16:00 UTC on 16 August 2026, DeepSeek's own changelog confirmed a shift from flat per-token pricing to a peak and off-peak structure. Peak hours are 01:00-04:00 and 06:00-10:00 UTC, which is 9am-noon and 2pm-6pm Beijing time, seven hours a day; the other 17 hours are off-peak, priced at exactly half the peak rate. DeepSeek's stated reason, in its own words, is "to allocate resources more reasonably" and to encourage developers to shift work into quieter hours. The company had already flagged a coming "significant increase" on its pricing page around 6 August, so this wasn't a total surprise to anyone watching closely.

Model / token type              Old (flat)    New off-peak   New peak    Peak increase
------------------------------------------------------------------------------------
V4-Pro  input, cache miss         $0.435/M       $0.66/M        $1.32/M        ~203%
V4-Pro  input, cache HIT          $0.0036/M      $0.022/M       $0.044/M      ~1,114%
V4-Pro  output                    $0.87/M        $1.98/M        $3.96/M        ~355%
V4-Flash  input, cache miss       $0.14/M        $0.22/M        $0.44/M        ~214%
V4-Flash  output                  $0.28/M        $0.66/M        $1.32/M        ~371%

That table is the whole story, and it's why headlines saying "prices quadrupled" and headlines saying "prices up 1,100%" are both technically accurate about different numbers. The output-token increase, the one that determines most people's monthly bill, is a 355-371% jump, roughly a 4x-4.7x multiplier. The eye-catching 1,100%-plus figure belongs to cached input tokens specifically, which is a narrower and more specialized cost: it's what you pay when a request reuses context DeepSeek's servers have already processed, common in long back-and-forth agent sessions and repeated-context workflows. That token type used to be priced at an unusually low $0.0036 per million, so even a large dollar jump produces a huge percentage swing. If your workflow doesn't lean on heavy prompt caching, you'll feel closer to the 4x number, not the 12x one.

Who actually felt this, and what they said

The clearest breakdown of the change circulating among developers came from a Hacker News thread discussing DeepSeek's own announcement. One widely upvoted comment put it plainly: "the four-digit percentage increase comes from cached input, which was priced at an unusually low $0.003625 per million tokens... Even after the increase, DeepSeek remains cheap relative to major frontier models... at peak rates, DeepSeek V4-Pro still costs about two-thirds as much as Sonnet on input and roughly 40% as much on output." Another developer admitted that even before the hike, the friction of maintaining a DeepSeek-based setup wasn't clearly worth it once a Claude subscription was already paid for: "I have been trying to stick with a DeepSeek + OpenCode setup for the whole week, but seems like my subsidized Claude Max is still quite relevant."

A third point from that thread is easy to miss and genuinely useful if you're deciding whether to switch providers over this: raw per-token pricing tells you less than it seems to, because it ignores how aggressively a provider's caching actually works and how many tokens a given model burns to finish the same task. A cheaper sticker price on a model that needs twice the tokens to solve a problem isn't actually cheaper. Some developers on the same thread reported monthly bills climbing from around $5 to $15 under the new structure, and noted that DeepSeek's peak hours line up with China's core workday, which structurally disadvantages US and EU developers who can't easily shift heavy workloads into DeepSeek's off-peak window without working unusual hours themselves.

How DeepSeek compares now

The honest framing: DeepSeek didn't stop being cheap, it stopped being absurdly cheap. At peak pricing, V4-Pro's output rate of $3.96 per million tokens is still a fraction of what the Western flagships charge.

Model                                  Input $/M      Output $/M
--------------------------------------------------------------
DeepSeek V4-Pro (off-peak)               $0.66          $1.98
DeepSeek V4-Pro (peak)                   $1.32          $3.96
DeepSeek V4-Flash (peak)                 $0.44          $1.32
Claude Sonnet 5                          $2.00          $10.00
Claude Opus 5                            $5.00          $25.00
Claude Fable 5                          $10.00          $50.00
GPT-5.6 Luna                            ~$0.20          ~$1.20
GPT-5.6 Terra                            $2.00          $12.00
GPT-5.6 Sol                              $5.00          $30.00
Gemini 3.7 Flash (intro, through 2026)   $0.75           $3.75

V4-Pro at peak still undercuts Claude Sonnet 5 on output price and comes in at roughly a quarter of Opus 5's rate and an eighth of GPT-5.6 Sol's. The gap that mattered to price-sensitive developers a month ago, DeepSeek being 10 to 30 times cheaper than a Western flagship, has narrowed to something closer to 4 to 8 times. That's still a real advantage. It's just no longer the kind of advantage that makes the pricing itself the whole pitch, especially against GPT-5.6 Luna, which OpenAI cut 80% in price on 30 July and now undercuts even off-peak DeepSeek V4-Flash on output.

What this means if you're actually building something

• Free web/app chat users: nothing changes. This price shift is API-only. If you're just using deepseek.com or the mobile app to chat, your access is unaffected.
• Batch or non-interactive workloads: shift what you can into the 17 off-peak hours (01:00-04:00 and 06:00-10:00 UTC) to keep roughly the old 2x multiplier instead of the new peak rates.
• Cache-heavy agent workflows (long-running sessions that repeatedly reuse context) are the hardest hit in relative terms, because that's exactly the token type that jumped ~12x at peak. Budget accordingly or reconsider whether the workload needs to run at peak hours at all.
• Before switching providers over this alone, price a real workload end to end, not just the per-token rate. A model that needs more tokens or more retries to finish the same job can erase a cheaper sticker price fast, a point the Hacker News thread above makes directly.

For readers of this site whose actual job is reading and summarizing documents rather than building agent pipelines, this news is more interesting than directly relevant. Gemini Notebook isn't priced per API token at all; it's a flat subscription (free, or $4.99-$199.99/month depending on tier) with source and query limits instead of a metered bill, which is a genuinely different cost model than anything in the table above. If you're evaluating AI tools generally rather than specifically shopping for the cheapest API, our comparison of Gemini Notebook alternatives covers where DeepSeek and the rest actually fit.

Is DeepSeek still cheaper than ChatGPT, Claude and Gemini?

Yes, but by a smaller margin than before. At peak API rates, DeepSeek V4-Pro's output price ($3.96/million tokens) is still roughly a quarter of Claude Opus 5's and an eighth of GPT-5.6 Sol's, down from the 10-30x gap DeepSeek used to have. Off-peak, the gap widens back out closer to the old margins.

Does the price increase affect DeepSeek's free chat app?

No. The 16 August 2026 pricing change applies only to the API, the interface developers use to build on top of the model. DeepSeek's free website and mobile chat app are unaffected.

Why did the price go up by 1,100% for some numbers but not others?

The 1,100%-plus figure applies specifically to cached input tokens, which were priced unusually low before the change ($0.0036/million) and rose to as much as $0.044/million at peak, a roughly 12x increase in dollar terms even though the absolute cost is still small. Output tokens, which determine most people's actual bill, rose a separate and smaller 355-371%.

Is DeepSeek V4 Pro actually a better model than the previous version?

DeepSeek's own published benchmarks show large gains, but independent testing has not confirmed them. The South China Morning Post reported an independent Terminal-Bench rerun scoring around 54.68% against DeepSeek's claimed 87.9%, and Vals AI's leaderboard reportedly ranks the model 12th, behind the previous-generation GPT-5.5. Treat DeepSeek's own numbers as a starting point, not a settled fact.