DeepSeek opens a million-token Flash beta at low API prices

DeepSeek released V4 Flash 0731 as a public beta on July 31. The company said it retained the architecture and size of the earlier preview while adding a new post-training pass. The release was available through the API and did not replace DeepSeek's Pro model in the consumer app, web service or API.

The model supports a one-million-token context window and a maximum output of 384,000 tokens, according to DeepSeek's documentation. It also includes Responses API and Codex-oriented adaptations. Those specifications open long-document and coding use cases, but the ability to accept a very large context does not show how well the model uses every part of it.

Official prices per million tokens were listed at $0.0028 for cache-hit input, $0.14 for cache-miss input and $0.28 for output. The large gap between cached and uncached input makes workload shape central to any comparison. A headline rate based on cache hits will not describe a job that repeatedly sends new context.

Raw token prices also leave out latency, reasoning tokens, retries, failed calls and output quality. A model can be inexpensive per token but costly per accepted result if the task needs repeated correction. Comparisons with GPT, Claude, Gemini or Kimi should use the same inputs and report cache behavior and completion criteria.

Because the release is a beta, model identifiers and prices may change. Developers should confirm the active API entry before deployment and keep cost records tied to a date. A reproducible cost-per-task test will provide more lasting information than a launch table that assumes every token is billed at the lowest rate.