DeepSeek further strengthens its Flash model and pricing grid
DeepSeek released V4.1-Flash-1017 without event or press release on October 17. The model improves on several benchmarks, while its Search API opens up and prices remain aggressive against American giants.
DeepSeek released V4.1-Flash-1017 on October 17, an update to its V4.1-Flash model launched five weeks earlier, without a keynote or press release. The version improves its scores on 11 benchmarks announced by the publisher, with a gain from 74.2 to 77.0 on DeepSWE according to MarkTechPost, a test measuring the ability to solve real coding tasks. At the same time, the group maintains a very aggressive pricing policy, which directly matters for teams deploying AI agents at large scale.
A discreet but significant update
DeepSeek has distributed the new weights on Hugging Face and ModelScope under the MIT license, a formula that allows developers to reuse the model more freely. The model retains its backbone of 552 billion parameters, but only a part is activated in use: 8 billion to read a prompt and 16 billion to generate a response, according to the provided information. This causal encoder-decoder architecture aims to reduce execution costs. In other words, DeepSeek cuts the bill where it hurts most, without making too much noise. On the benchmarks reported by the company, V4.1-Flash-1017 performs better than the previous version on the 11 cited tests, with differences ranging from 2.0 to 11.1 points.
The most closely watched point remains the model's concrete scope. DeepSWE, a benchmark highlighted by MarkTechPost, evaluates end-to-end software engineering task resolution, not just the ability to neatly answer a multiple-choice questionnaire. For teams building agents capable of writing code, calling tools, and completing a task, this is more meaningful than a showroom score. And this is precisely where DeepSeek pushes its advantage.
The price, still the most aggressive weapon
The pricing grid for V4.1-Flash remains very low. At peak hours, the API is priced at $0.30 per million input tokens and $1.20 per million output tokens; during off-peak hours, rates drop to $0.15 for uncached input and $0.60 for output, according to reported information. This is low compared to premium competing models like those from OpenAI or Anthropic, generally much higher at comparable performance on certain agentic uses.
This strategy is not new. V4.1-Flash was launched on September 10 to compete with V4-Pro at a lower cost, and DeepSeek initially considered redirecting V4-Pro API traffic to Flash before reversing course on September 15, facing demand. In this context, the company's logic seems clear: gain users by lowering the cost per task, not by multiplying announcements. No need for fanfare when the bill is enough to attract attention.
The new piece of the puzzle: the Search API
BigGo Finance also indicates that this release marks the first public opening of DeepSeek's proprietary Search API. This detail matters because so-called agentic uses often rely on information retrieval, reasoning, then action. By providing access to its own search infrastructure, DeepSeek reduces dependence on third-party components and simplifies developers' tech stack.
The model also handles images and context windows up to 1 million tokens, still according to available information. This broadens the range of use cases, notably for applications that need to ingest large documents, code repositories, or very long conversation histories. A major caveat remains: the figures advanced by DeepSeek have not been independently verified, and the model remains behind the best closed models on some more difficult reasoning tests, such as Humanity’s Last Exam [to be verified].
What this changes for the market
The signal sent to competitors is less spectacular than a high-profile product launch but more troubling in the long term. By slightly improving its model while maintaining very low prices, DeepSeek forces OpenAI, Anthropic, and Google to defend both their margins and their scores. For startups, the challenge is immediate: an improvement of a few points on a benchmark like DeepSWE can be enough to reduce the cost of an agent running millions of times per month. The AI model market is therefore not only about raw quality; it is also about the real cost of a solved ticket, a generated line of code, or a completed task.
Since its DeepSeek-R1 moment, which shook markets in January 2025, the company has progressed through rapid and low-key iterations. It is less photogenic than a stage demonstration but ruthlessly effective in establishing a price reference. And in AI, prices often speak louder than slides.
By the numbers
- 552 billion parameters — backbone of V4.1-Flash-1017, according to DeepSeek.
- 8 billion parameters — activated when reading the prompt.
- 16 billion parameters — activated when generating a response.
- 1 million tokens — announced maximum context size.
- $0.30 / $1.20 — API rates per million tokens, input/output at peak hours.