Image: DeepSeek
DeepSeek has made it official. DeepSeek-V4.1-Flash-1017, the model developers spotted in its API earlier today, is now the default Flash model on DeepSeek's website, app and API, and the company has also released the weights for anyone to download.
That settles what we reported this morning, when DeepSeek watcher Teortaxes noticed the deepseek-flash model's fingerprint had changed and guessed the new version felt "smarter" and "terser." DeepSeek's own pitch lines up with that.
🚀 DeepSeek-V4.1-Flash-1017 is here!
— DeepSeek (@deepseek_ai) October 17, 2026
Stronger benchmark performance, faster task completion, and shorter outputs than the previous version—a Pareto improvement in performance and cost. Now live on our website, app, and API. pic.twitter.com/MiqGbnCiVz
In its announcement, DeepSeek says the update brings "stronger benchmark performance, faster task completion, and shorter outputs" than the original V4.1 Flash, which launched last month. Its changelog says nothing changes for API users: the price stays the same, and calling deepseek-flash now gets you the new model automatically.
The Hugging Face model card calls it a "minor update," but DeepSeek's numbers suggest a bigger jump than that. On Terminal-Bench 4.0, a test of coding in a command-line environment, the score rises from 31.2 to 42.3. ProgramBench climbs from 20.3 to 29.1. On the Humanity's Last Exam reasoning test without tools, it goes from 36.8 to 42.6.
DeepSeek's benchmark comparison. Image: DeepSeek
DeepSeek's table puts the update ahead of Claude Opus 5 on several agent-style tests, including DeepSWE, Automation-Bench and Apex Bench, while Opus 5 still leads on Terminal-Bench 4.0, ProgramBench and the reasoning test. These are DeepSeek's own results, run on test setups it chose, and nobody has independently checked them yet.
The more interesting chart is about cost. On DeepSeek's DeepSWE graph, the new model's top "max" setting scores 77% at roughly 20 cents per task. The old V4.1 Flash topped out around 74% at close to 50 cents. Even the new low setting, at about 11 cents a task, lands near the scores that some rival models in the chart reach at several dollars per task.
DeepSWE score vs cost per task. Image: DeepSeek
That fits the "shorter outputs" claim. With prices per token unchanged, a model that writes fewer tokens to finish the same job simply costs less to run.
One thing the "Flash" name hides: Hugging Face lists the model at 763 billion parameters. It's released under the permissive MIT licence with a 1-million-token context window, but this isn't something you'll run on a home PC. For most people, the app and API are the practical way in.
DeepSeek also opened its in-house Search and Fetch APIs to all developers, at $0.75 per 1,000 successful calls each. Search and Fetch calls made through tools like DeepSeek Harness, Claude Code and Codex are free for DeepSeek API users until billing starts on 1 November, Beijing time.
The quiet rollout does have a downside. Anyone who built on deepseek-flash got a different model today without changing a line of code. DeepSeek says that's an upgrade, but teams with carefully tuned prompts may want to test before they trust it.
Comments