DeepSeek Raised API Prices Up To 371%: What We Moved To

On 6 August 2026, DeepSeek told developers that API prices would rise significantly. It gave no percentage and no date, so most of the coverage stopped there. The rates landed on 16 August 2026 at 16:00 UTC, and they are worth reading closely, because the structure changed as well as the numbers.

What actually changed

Both tiers moved from flat pricing to peak and off-peak billing. Peak hours run 01:00 to 04:00 and 06:00 to 10:00 UTC. Off-peak is set at half of peak.

At peak, V4-Flash input rose 214% and output rose 371%. V4-Pro rose 203% and 355%. Those are the headline figures, and they are large enough that any cost model built before 16 August is now wrong.

The off-peak tier is not a discount

This is the part that gets misread. An off-peak rate at half of peak sounds like a concession, and several write-ups have treated it as one. It is not a discount against what you were paying. It is a discount against a number that did not exist two weeks ago.

Measured against the old flat rate, DeepSeek's cheapest available tier is 57% more expensive on input and 136% more expensive on output. There is no time of day at which V4-Flash costs what it cost on 15 August. If your workload already ran overnight, you did not dodge the increase, you got a smaller one.

The off-peak window also has an operational cost that rarely makes it into the spreadsheet. Shifting batch work to specific UTC hours means either your automation queues and waits, or someone maintains a scheduler that did not need to exist before. For a small team, that maintenance is often worth more than the tokens it saves.

The crossover almost nobody has mentioned

On 30 July 2026, OpenAI cut GPT-5.6 Luna by 80%, to $0.20 input and $1.20 output. Terra fell 20% to $2 / $12, and Sol kept its price while becoming roughly 2.5 times faster in the API. These were list price changes rather than a promotion, so there is no expiry date to plan around.

Put the two events together and the ordering flips. During DeepSeek's peak hours, Luna is 2.2 times cheaper on input and about 10% cheaper on output than V4-Flash. The mental model that a Western frontier lab is the expensive option and DeepSeek is the budget one stopped being true on 16 August, at least for the hours when most of Europe and Asia are actually working.

What we moved, and what we left alone

We run client automations across classification, extraction, summarisation, and a smaller amount of agentic work. The move was not wholesale, because the right answer differs by task.

• High-volume classification and extraction went to Gemma 4 26B A4B at $0.070 / $0.300. It is Apache 2.0, released 2 April 2026, carries a 262K context window and native function calling, and posts 85.2% on MMLU Pro and 84.3% on GPQA. For work where the output is a label or a JSON object, it is cheaper than DeepSeek at DeepSeek's best hour and the quality difference does not show up.
• Latency-sensitive user-facing calls went to GPT-5.6 Luna. The 80% cut made it competitive on price, and it removes the peak-hour scheduling problem entirely.
• Long-running batch jobs stayed on DeepSeek off-peak. At $0.22 / $0.66 it is still the cheapest place to put work that genuinely does not care when it runs.
• Nothing moved to a model we had not already load-tested on our own prompts. Benchmark scores do not predict behaviour on your specific extraction schema, and a migration that saves $40 a month and costs a day of debugging is a loss.

The reason a price table is not enough

Per-token price is the number everyone compares, and it is a weak predictor of what you actually pay. Models differ enormously in how many tokens they spend reaching the same answer. Research on reasoning efficiency has found two models hitting comparable accuracy while differing by more than 25 times in tokens generated, which swamps any price gap in the table above.

That means a model at half the per-token price can be more expensive per completed task, and the only way to know is to measure your own workload. We cover the method and the numbers in cost per token is the wrong metric, which is the piece to read before acting on any of the prices here.

What to do this week

Pull your last 30 days of usage and split it by input and output tokens. Output is where the increase bites hardest, and workloads that generate long responses are the ones that just got 136% to 371% more expensive. Reprice those first. If your token mix is output-heavy, the case for moving is much stronger than the input-side numbers suggest.

Then check whether you are on V4-Pro for work that V4-Flash or Gemma 4 would handle. In our experience most production automation sits well below the capability ceiling of the model behind it, and tier choice is a larger lever than provider choice. If you want that audit done properly, our AI and automation team runs it against your actual traffic rather than a benchmark.