DeepSeek released V4.1-Flash-1017 on Friday, and the announcement was three sentences long. A new checkpoint, the web app and API move with it, and pricing stays the same. "A minor update," DeepSeek called it on social media, according to Bloomberg. Then the benchmark table landed, and every one of the eleven rows moved up.

Terminal-Bench 4.0 went from 31.2 to 42.3 — an eleven-point jump on an agentic coding benchmark for a release DeepSeek barely announced. ProgramBench rose from 20.3 to 29.1, Apex Bench from 64.6 to 72.1, and the model held its ground on DeepSWE v1.1 at 77.0. Faster task completion, shorter outputs, same API price. The gap between the word "minor" and those numbers is the whole story.

The independent results matter more than the self-reported ones, and they're arriving too. Arena.ai's Code Arena put the October 17 build at 1,670 points on its WebDev leaderboard — the highest-ranked open model on the board, displacing Moonshot's Kimi K3 (Max) while costing roughly a tenth as much to run. And Browser Use's own v2.1 benchmark, run on October 18, called it "the new pareto frontier": 75.0 at $0.22 per task against Claude Opus 5.5's 75.0 at $2.35 — equal score, more than ten times cheaper.

That cost-per-task ratio is the part worth watching. DeepSeek isn't competing on who wins Terminal-Bench anymore. It's competing on what happens when a model is good enough at a tenth of the cost: developers pair the cheap model for the routine stuff with the expensive brain on tap, and the expensive brain gets used less and less.

There's a second, quieter change in this release: DeepSeek also opened its in-house search engine to developers as a paid Search and Fetch API, per AI News Round. The model is free to use under an MIT license on Hugging Face and ModelScope; the tooling around it is starting to be a business. Open weights for reach, paid APIs for revenue — the same funnel playbook the Western labs ran, now running in reverse.

As Bloomberg notes, competition between Chinese open-weight models and U.S. closed models is intensifying as DeepSeek ships in quick succession. The uncomfortable question for the closed labs: if "minor updates" keep arriving at this pace and this price, how much headroom does a $2.35-per-task frontier model actually have?

— via Bloomberg