Photo: Google DeepMind / Pexels
The Hangzhou startup's V4.1-Flash-1017 update is billed as faster and leaner, with benchmark gains and unchanged API pricing
DeepSeek has shipped another update to its AI lineup, and this one is built around a familiar pitch: more performance for less money.
The Hangzhou-based startup released DeepSeek-V4.1-Flash-1017 on October 17, 2026. It is designed to be faster and leaner as it competes with Silicon Valley heavyweights like OpenAI and Anthropic, plus a growing crowd of Chinese rivals.
The timing matters. The model it replaces only arrived on September 10, 2026, so this refresh landed roughly five weeks later.
What changed in the new model #
DeepSeek describes the update as a “Pareto improvement” in both performance and cost. That economics term means one thing gets better without anything else getting worse.
Applied here, the model is reported to post stronger benchmark scores, finish tasks faster and produce shorter outputs. API pricing stays the same, and developers reach the model through the existing deepseek-flash endpoint.
Shorter outputs may sound like a minor detail. They are not, because AI usage is typically billed by the volume of text a model generates. A model that reaches the same answer in fewer words is cheaper to run, even when the sticker price does not move.
The headline benchmark figure comes from DeepSWE v1.1, a coding test. The new model scored 77.0%, up from 74.2% for the previous version.
Benchmarks indicate that figure also tops Claude Opus 5 from Anthropic, which scored 74.0%. Benchmarks further suggest DeepSeek outperforms larger models, Claude Opus 5 among them, on reasoning and coding tasks.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
Cost is where the gap looks widest. Effective per-task costs for DeepSeek’s model are often under $0.20-$0.25, compared with $2-$10+ for competitor models.
Agents, coding and a new set of APIs #
Much of the improvement centers on what the industry calls agentic capability. DeepSeek says it strengthened these agentic and coding skills using reinforcement learning with self-feedback.
The practical upshot is end-to-end project building. Users can reportedly ask the model to create things like games and 3D models, not just snippets of code.
The update also expands what DeepSeek sells. For the first time, the company is opening its Search & Fetch APIs to outside developers. Those tools are priced at $0.75 per 1,000 calls. Some of them are being offered free until November 1, 2026.
The model itself is available on DeepSeek’s website, app and API. Its weights are open-sourced under the MIT license on platforms including Hugging Face and ModelScope.
Background: a faster release cadence #
DeepSeek has been picking up speed throughout 2026. Since mid-year, the company has moved to monthly updates, with a consistent focus on efficient architectures and open weights.
That strategy sets it apart from its biggest US competitors. OpenAI and Anthropic largely keep their most advanced model weights closed and sell access through paid products and APIs.
DeepSeek also faces pressure at home. Chinese rivals include Moonshot’s Kimi and the GLM models from Z.ai, both competing for the same developer attention.
What this means #
For developers, the calculation is straightforward. If the benchmark claims hold up in real-world use, DeepSeek offers competitive coding performance at a much lower cost per task. Free access to some search tools until November 1, 2026 lowers the barrier to experimenting further. The caveat is that benchmarks are not production environments. A 77.0% score on DeepSWE v1.1 is a meaningful data point, but teams will want to test reliability, security and compliance requirements against their own workloads before switching.
The things to watch are whether independent testing confirms DeepSeek’s benchmark numbers and how quickly developers adopt the new Search & Fetch APIs once the free period ends.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our