DeepSeek pushed DeepSeek-V4.1-Flash-1017 live on its website, app, and API on October 17, under the same deepseek-flash identifier already wired into existing integrations. The company says the checkpoint scores better on benchmarks, completes faster, and produces shorter outputs than its predecessor. Alongside it, DeepSeek opened Search and Fetch APIs priced at $0.75 per 1,000 successful calls, and published MIT-licensed weights for the model. Calls routed server-side through Harness, Claude Code, or Codex run free until October 31 at 16:00 UTC.
The identifier reuse matters more than the version bump. Anyone pinned to deepseek-flash, rather than a pinned snapshot hash, gets the new checkpoint automatically, with no opt-in and no staged rollout mentioned in the announcement. That is a silent swap. Teams that built evaluation suites against the prior model's output length or latency profile may see those baselines shift without a corresponding alert, since DeepSeek frames the changes as improvements but gives no migration notice or deprecation timeline for the old weights.
DeepSeek's announcement confirms the release date, the pricing for Search and Fetch, the free window through October 31, and the MIT license on the weights. It does not disclose rate limits for the new endpoints, error-handling behavior on failed Fetch calls, or what happens to billing once the free period ends for Harness, Claude Code, and Codex traffic specifically. Anything about regional availability, latency SLAs, or rollback options is absent from the facts DeepSeek published. Treat those gaps as open questions, not settled defaults, until DeepSeek's API docs or status page says otherwise.
For platform operators, the immediate task is inventory. Check which services call deepseek-flash without a version pin, and check whether Search or Fetch calls are already flowing through billing pipelines that assume the old per-call cost or a zero-cost tier. A $0.75-per-1,000 rate is cheap at low volume but compounds fast in any pipeline doing bulk retrieval or agentic search loops, and nothing in the announcement caps usage during the free window, which could mean usage spikes right before October 31 as teams front-load free calls.
The free period is short, eight days from the source event date. Any load testing against Search or Fetch needs to happen now, because post-cutoff pricing applies without a stated grace period. On-call runbooks that reference deepseek-flash output length for timeout thresholds should be revisited, since DeepSeek's own claim of shorter outputs could trip downstream parsers expecting the old response shape.
October 31, 16:00 UTC is the date the free tier for server-side Harness, Claude Code, and Codex calls ends, and no extension has been announced.