DeepSeek Ships V4.1-Flash-1017: Same Price, Faster Tasks, and the Web Search Engine Opens Up
DeepSeek released V4.1-Flash-1017 on Saturday, October 17 — a “minor update” to its fast Flash model that lands as anything but minor in the numbers: Terminal-Bench 4.0 jumps from 31.2 to 42.3, DeepSWE v1.1 climbs from 74.2 to 77.0, and Agents’ Last Exam rises from 31.8 to 35.4. The company calls it a “Pareto improvement in performance and cost” — benchmarks up, task time down, output length shorter, price unchanged. According to the official release notes, the update is live on the website, the app, and the API, and it ships alongside something arguably bigger: DeepSeek’s in-house search engine, previously an internal tool, is now an open developer API.
The model card on Hugging Face spells out what changed under the hood. V4.1-Flash-1017 was trained with reinforcement learning using self-feedback — the model critiquing and improving its own rollouts — with the goal of completing entire projects end to end, from game creation to 3D modelling. The release notes showcase generated examples that go well past file-level autocomplete: a 3D model of the Forbidden City, a playable web racing game, and a recreation of Portal 2 built in Godot. On OpenRouter, the snapshot is served under the deepseek-flash name with a 1M-token context window, and the weighted average price customers actually pay — after caching and discounts — comes to about one cent per million input tokens ($0.01071) and $0.59 per million output tokens.
The search API is the quiet part
The headline numbers will get the attention, but the search API may matter more. DeepSeek says its in-house engine indexes hundreds of billions of pages and serves hundreds of millions of queries a day, tuned — as Trending Topics reports — to filter out SEO spam and low-quality results. Both Search and Fetch endpoints are priced at $0.75 per 1,000 calls, and tool-driven calls made through DeepSeek Harness, Claude Code, or Codex stay free for DeepSeek API users until October 31, 2026, 4:00 PM UTC, after which billing starts.
That pricing language is worth pausing on. Claude Code and Codex are competitors’ agents — Anthropic’s and OpenAI’s respectively — and DeepSeek is making its search layer free inside them until the end of the month. It is a land-grab move dressed as a courtesy: get your retrieval layer into every serious coding agent, then switch on the meter once it’s load-bearing.
The Pareto playbook, round three
This is the third time this year DeepSeek has played the same card. In September, the lab cut the flagship’s price by 75 percent permanently; earlier, the V4 line shipped as open weights that matched frontier models at a fraction of the cost. The January 2025 shock was a performance surprise. The October 2026 version is a logistics surprise — no single number looks spectacular, but the dev.to community writeup makes the point bluntly: a fast/cheap model beating models that cost ten times more on the benchmarks that matter for agents.
The “talks less” change is doing more work than it appears. DeepSeek reports output lengths shortened across the board, which reduces per-task cost more than any per-token discount would — you pay for fewer tokens to get a finished task. Combined with the shorter execution time, the cost-per-completed-task curve bends down twice.
What it means for the price war
OpenAI’s GPT-5.6 and Anthropic’s Opus 5 still lead the frontier table — Opus 5 posts 51.8 on Terminal-Bench 4.0 against DeepSeek’s 42.3, and both retain a clear edge on the hardest reasoning suites. But the frontier in late 2026 is not the frontier most buyers live on. The majority of agentic workloads need the 30-80 band: good enough to complete the task, cheap enough to run thousands of times a day. DeepSeek has just moved that band’s cost down again while keeping its weights open under MIT on Hugging Face and ModelScope.
The updated release notes are optimistic for their own harness too; the notes recommend DeepSeek Harness v0.2.1-rc.1 or later for the best experience, and the model was optimised for Tencent’s CodeBuddy alongside it. This ties into the company’s broader full-stack strategy — the model, the harness, and now the search layer — of which the search API is the newest piece. When DeepSeek released V4 as open weights earlier this year, matching GPT-5.5 in the process, the pattern was the same: performance at the same or lower price, pushed out to everyone who downloads it.
For its own part, the company has been here before — when the original V4.1-Flash shipped in September, two days before Amodei asked labs to slow down, we wrote that it was the argument against a slowdown in concrete form. The 1017 snapshot strengthens that argument. Efficiency-first model updates are arriving monthly now, and each one resets the floor for what a task costs.
New Zealand angle: with the price gap between “frontier-adjacent” and “actually affordable” now this wide, the practical calculus for local developers building on open agents shifts again. A thousand automated document-review tasks that cost dollars under Western lab pricing cost cents under the DeepSeek snapshot — and with weights on Hugging Face, none of it requires trusting a hosted API at all.
