DeepSeek Just Released V4.1 Flash 1017, Scores 77.0 on DeepSWE, Opens Search API

Justy A cheaper coding model plus built-in web tools under one key is the kind of thing that makes people rewrite a routing table before lunch.

Cody Yeah. It matters because the annoying part of agent stacks is still the glue. DeepSeek updated V4.1-Flash to this 1017 checkpoint, kept it behind the deepseek-flash name, and then opened Search and Fetch to everyone. Therefore the story isn't just score up, cost down. It's fewer vendors in the loop.

Justy Right, and that's the user story immediately. If you already have something doing coding or repo work, you don't want a grand migration. You want the same model name, your app keeps running, and suddenly your agent can search the web and pull a page as Markdown without you wiring three extra services together.

Cody Mm-hm.

Justy Also, very funny that episode ten ninety-three of this extremely normal show is once again about the layer around the model deciding whether anybody actually uses it.

Cody You mean our long and glamorous career of reading pricing tables was not in vain.

Justy There it is. That's the catch-up beat. We touched the stove and we're back.

Cody What they actually shipped is pretty specific. V4.1-Flash is a multimodal mixture-of-experts model with forty transformer layers, split into a twenty-layer causal encoder and a twenty-layer decoder. And for this checkpoint, their claim is reinforcement learning with self-feedback for end-to-end project work, with examples ranging from game creation to 3D modeling.

Justy Yeah.

Cody I don't hate that framing, actually. It's at least aimed at a real workload shape. But the benchmark table is still their own table. They say 77.0 on DeepSWE, they say it leads on Automation-Bench, Agents' Last Exam, Apex Bench. Fine. BUT Claude Opus 5 still leads their own chart on Terminal-Bench four point zero, ProgramBench, N L 2 Repo, and H L E, which is a useful reminder that this is not some clean sweep.

Justy And that's where you save me from getting too shiny about it. Because the win isn't "mystery solved, best model." The win is more practical than that. If you're an agent builder, you can buy a fast model at thirty cents per one million cache-miss input tokens and a dollar twenty per one million output tokens at peak, half off off-peak, then bolt on first-party search and fetch for web grounding. That's a pretty legible bundle.

Cody Right.

Justy And the migration path is low drama. OpenAI- or Anthropic-format endpoints means you're not rewriting your whole client. DeepSeek even tells you which harness version they want, zero point two point one release candidate one or later. That's the kind of upgrade note I trust more than a chest-thumping benchmark graphic.

Cody Same. Though I'd still separate two things. One is the model got better. Probably true. The other is the model got shorter in output and faster in task completion, which they call a Pareto improvement. Maybe! But shorter can also mean it learned to stop talking, which is good for cost and sometimes bad for thoroughness. You only know after you run your own tasks.

Justy No, that's fair. Shorter is not automatically smarter. It's just VERY appealing when you're paying the bill and your agent currently writes a novella before opening a file.

Cody The world's most expensive preamble. "I have carefully considered your request" — please don't. Just patch the bug.

Justy Exactly. And opening Search plus Fetch is the other half here. Search gives ranked web results for a query. Fetch grabs one page and returns Markdown. That's not glamorous, but for coding agents and RAG-ish workflows it's a clean primitive set. Search finds. Fetch reads. Move on.

Cody Okay okay. The zoom-out beat is this: open-weight pressure keeps getting weirder in a good way. The competitive frame in front of us says DeepSeek's own table has this thing above Claude Opus 5 on DeepSWE, seventy-seven to seventy-four, while another open-vs-closed snapshot says the best open-weight models are months behind, not years, and one source pegs the best open model winning forty-four percent of blind votes against the best closed one.

Justy Yeah, and this one steals a job buyers actually recognize. Not "be impressed by my algebraic paper name." More like, here's a coding model, here's web access, here's the bill, here's the endpoint shape you already support.

Cody Please don't summon Harness cubed.

Justy It'll optimize the font and call itself agentic typography.

Cody My honest reservation is the search side, not the model. "Hundreds of millions of requests daily" is nice social proof, but developer-facing search quality lives or dies on weird edge cases, rate limits, and whether fetch returns clean enough Markdown on ugly sites. The docs matter there more than the launch post.

Justy Totally. And I think that's why the one-key part matters so much. Even if the search is just decent, teams will try it because procurement friction is near zero. You don't need a new vendor story. You just turn on the web layer next to the model you were already testing.

Cody Mm-hm.

Justy My near-term bet: seventy-thirty that by November first, at least one major agent framework or gateway beyond DeepSeek surfaces Search and Fetch as a first-class integration, not just "here are raw endpoints." Vercel already moved on the model side. The web-tool side feels next.

Cody I'd take the under a little, like sixty-forty, only because tool integrations lag docs. And I should probably be less cocky on open stuff right now because I've been eating my vegetables and still missing some of these calls.

Justy Growth. Maturity. A calibrated king.

Cody Don't do this to me.

Justy Fine. Go read the docs, run your ugliest real task, and if the agent starts writing less fan fiction, maybe keep it. That's enough for me, Cody.