DeepSeek updates low-cost AI model with better test scores

DeepSeek released DeepSeek-V4.1-Flash-1017 on October 17, 2026, an upgrade to the earlier V4.1 version, according to the company's own update log. DeepSeek says the new version delivers significant improvements across benchmarks. A benchmark is a standardized test that compares how well different AI models handle the same tasks.

The model is used with the model parameter deepseek-flash. According to the log, pricing and API methods are unchanged. An API is the technical connection companies and developers use to hook the AI model into their own systems.

The figures are DeepSeek's own

DeepSeek reports these scores for the new version: 42.6 on HLE, 42.3 on Terminal-Bench 4.0 and 77.0 on DeepSWE v1.1. These are DeepSeek's own numbers, and they have not been independently confirmed.

According to DeepSeek, task runtime and response length have been reduced. DeepSeek says users can therefore solve tasks faster and at lower cost.

Its own search engine

The documentation also states that DeepSeek's web and API search features are powered by an in-house DeepSeek search engine. It refers readers to a usage guide.

Background: the September version

V4.1 Flash was originally released in September 2026. According to the independent outlet The Next Web, the model has 552 billion parameters, of which 8 billion are active when reading input and 16 billion when generating output. Parameters can roughly be thought of as the model's adjustable dials, which it has learned during training.

The Next Web lists a price of $0.60 per million output tokens off-peak and $0.003 for cached input. A token is a piece of text, typically a word or part of one. According to the outlet, Bloomberg Intelligence estimates a price cut of up to 32%, and the outlet describes it as a reversal after a price increase in August.

According to DeepSeek's own tests, as reported by The Next Web, the September version is on par with Claude Opus 5 on software development tests (74.2 versus 74.0) and leads on cybersecurity tests with 88.1. On Humanity's Last Exam, however, it scores 36.8 versus 56.3 for Opus 5.

The model's weights, meaning the learned numbers that make up the model itself, have been released under an MIT license on Hugging Face. This means others can freely download and use them.

Disagreement over V4 Pro

The sources disagree about DeepSeek's older V4 Pro. The Next Web writes that requests to V4 Pro are being redirected to V4.1-Flash from September 14. DeepSeek's own log, by contrast, says that after an update on September 10, the company decided to continue the API service for V4 Pro after September 14, 2026, with billing unchanged.