DeepSeek V4.1 Flash

What DeepSeek says. The pricing page is a feature and price table rather than a positioning statement, so DeepSeek does not spell out target use cases for Flash versus Pro. The Thinking Mode guide says thinking mode outputs chain-of-thought reasoning first "to improve the accuracy of the final response", and its sample code uses deepseek-flash throughout. The docs also carry setup guides for coding agents including Claude Code, Codex and OpenCode.

Our editorial judgment. Flash looks like DeepSeek's default for volume: five times the concurrency of V4 Pro, a lower price on every line, and vision, which Pro lacks. It suits batch summarization, extraction, chat back ends and agent subtasks, especially if you can schedule jobs outside peak hours. We would test Pro alongside it for hard reasoning or long coding sessions. For finished chat products rather than raw APIs, see our AI chatbots guide.

• Peak hours double the price. Peak is 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays. Everything else, including weekends, is off-peak at half price.
• Alias repointing. DeepSeek retired the models behind deepseek-v4-flash and deepseek-v4-flash-vision-exp; those names now route to V4.1 Flash. Pin your code to deepseek-flash and re-run your tests when the version behind a name changes.
• Thinking is on by default. Unless you disable it, every call produces reasoning content that you pay for as output.
• Ignored sampling settings. In thinking mode, temperature, presence_penalty and frequency_penalty have no effect, and top_p is clamped to 0.95 to 1.0.
• Tool calls need reasoning passed back. With tools in the request, you must return reasoning_content from earlier turns or the API returns a 400 error.
• Concurrency caps. Above 2,500 concurrent requests per account you get HTTP 429. DeepSeek offers free capacity expansion on request.
• Long waits. If inference has not started after 10 minutes, the server closes the connection.
• Prices can move. DeepSeek states it reserves the right to adjust prices.

DeepSeek lists two API models on one table. Both share the 1M context, 384K maximum output, thinking modes and feature list apart from vision.

At list rates, Pro costs 4.4 times as much on cache-miss input and 3.3 times as much on output. Outside DeepSeek, Google lists Gemini 3.5 Flash-Lite at USD 0.30 input and USD 2.50 output per 1M tokens, and Gemini 3.8 Flash at USD 0.75 and USD 3.75 through 2026.

1. Gather 30 to 100 real prompts with known good outputs, including some image inputs if you plan to use vision.
2. Run them on Flash in thinking mode at low and high effort and in non-thinking mode, logging cache hits, misses and output tokens.
3. Run the hardest subset on V4 Pro with identical settings to see where the cheaper model falls short.
4. Price each run at peak and off-peak rates, and test whether your workload can move into off-peak windows.
5. Load test near your expected concurrency to confirm you stay under the 2,500 cap or need an expansion request.

We read DeepSeek's Models & Pricing page, Thinking Mode guide and Rate Limit & Isolation page on 15 October 2026. We have not run DeepSeek V4.1 Flash, and nothing here is a test result or ranking. Prices and limits change; check DeepSeek's pages before committing budget. See our ranking method for how HumanRank scores tools and the model directory for other profiles.

What is DeepSeek V4.1 Flash? It is the model version DeepSeek serves under the API name deepseek-flash, with a 1M context, 384K maximum output, thinking mode and vision.

How much does DeepSeek V4.1 Flash cost? It costs USD 0.30 per 1M cache-miss input tokens and USD 1.20 per 1M output tokens at peak, and half that off-peak.

What happened to deepseek-v4-flash? Its model was retired; the name is still accepted, but requests are served by DeepSeek V4.1 Flash and billed at the Flash price.

When are DeepSeek off-peak hours? Off-peak is any time outside 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, plus all weekends and Chinese public holidays.

Does DeepSeek V4.1 Flash support images? Yes, Vision is listed as supported for Flash, unlike DeepSeek V4 Pro.

Can I turn off thinking mode? Yes, set {"thinking": {"type": "disabled"}}; the Thinking Mode guide also lists a reasoning effort of none as a way to disable it.