DeepSeek updates Flash and opens its Search and Fetch APIs

An old name now reaches a new model

The current API documentation maps deepseek-flash to Flash 1017. It still accepts deepseek-v4-flash and deepseek-v4-flash-vision-exp, but says their corresponding models are retired: requests are served by 1017 and billed at the Flash rate. Keeping an old identifier therefore does not preserve the old model’s behaviour.

DeepSeek’s release note also recommends DeepSeek Harness v0.2.1-rc.1 or later. The model card provides downloadable weights under the MIT licence. Its reasoning, coding and visual benchmark comparisons are supplier results; they do not establish independent performance superiority. The original announcement below introduces those claims and the confirmed rollout.

Model tokens and web retrieval are separate bills

Flash’s listed token prices vary by time of use. Peak windows are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays; all other hours, including weekends, are off-peak. The figures below are current rates, not a claim of a price cut.

Search finds pages; Fetch retrieves one

The Search and Fetch guide describes two standalone JSON APIs: Search returns structured web results, while Fetch retrieves one page as Markdown. Each costs $0.75 per 1,000 successful requests. Empty search results, failed requests and timeouts are not billed; asking for more search results does not change the per-request price.

There is a temporary distinction for server-side Search and Fetch tool calls through DeepSeek Harness, Claude Code and Codex when using the DeepSeek API. Those calls remain free until 31 October at 16:00 UTC, equivalent to 1 November at midnight Beijing time. The standalone endpoints already have their own listed charges.

The immediate job for an existing integration is to check its model routing and billing assumptions. A first personal coding test now offers a narrower look at the model in use.

A small coding test makes the harness visible

Ibragim Badertdinov, a research engineer at Nebius, published a personal leaderboard covering ten everyday coding tasks, with three attempts per configuration. His methodology describes automated checks and gives example tasks, but the full task set remains private. This is a small personal evaluation, with several configurations already reaching its ceiling.

His 17 October results compare Flash 1017 in Pi and DeepSeek Harness. Changing the surrounding agent and reasoning effort changes cost and time even when nine tasks pass all three attempts. The Pi high-effort run reports a 97.9% input-cache hit rate, so its bill should not be treated as a universal per-task price. Broader reliability and code quality still need tests beyond these ten tasks.