September 30, 2026
DeepSeek costs rise in daily tests: Seismograph creator reports a threefold increase
DeepSeek used more reasoning tokens on the same tasks, while response times barely changed.
Starting Sep 10, the Seismograph creator's spending on DeepSeek checks rose from $0.07 to $0.25 per day and remained roughly three times its previous level.
Token usage changed. In a Sep 30 post, the creator reported that DeepSeek reasoner had started using 10–12 times as many reasoning tokens as it had over the previous 3 weeks. The tasks remained the same. In a comment, the creator clarified that response times had barely changed.
Seismograph tracks changes in a given model over time rather than ranking competitors. Since Aug 20, its creator has tested 34 models from 15 labs daily through their APIs. Code evaluates the responses, without another LLM acting as a judge.
As of Sep 30, the creator had found no signs of hidden capability degradation in the other 33 models. Opus 5.5 has been included in the checks since launch.
Scheduled checks. Seismograph runs 29 probes daily at 08:30 UTC through GitHub Actions and 87 probes on Sundays. The creator froze the task set to compare new results against the accumulated history. The project was published on Sep 15, and its README states that every run since measurements began is recorded in the public Rekor log.
The code and methodology for PipeRoll Seismograph are available on GitHub. The creator publishes statistical results and the SHA-256 hash of the private task set. The tasks themselves and raw responses remain private.
DeepSeek officially released V4.1 Flash on Sep 10, available through `model="deepseek-flash"`. The old names `deepseek-v4-flash` and `deepseek-v4-flash-vision-exp` temporarily route requests to the new model.
The DeepSeek API lets users explicitly set the reasoning budget. According to the documentation as of Sep 30, thinking is enabled by default with `reasoning_effort="high"`. The OpenAI SDK supports `low`, `high` and `max`, while `extra_body={"thinking":{"type":"disabled"}}` disables thinking.
DeepSeek prices as of Sep 30 per 1 million output tokens:
| Model | Off-peak | Peak | | --- | --- | --- | | deepseek-flash | $0.60 | $1.20 | | deepseek-v4-pro | $1.98 | $3.96 |
Seismograph's public dashboard and the series.json, advisories.json and divergence.json feeds are available through links in the project's README.
