DeepSeek Called It a 'Small Update' — It Beat Claude on Coding at One-Fortieth the Price
A “minor update” that moved the numbers that matter
DeepSeek has a habit of underselling the releases that embarrass everyone else’s pricing. On Friday, October 17, the Chinese lab shipped DeepSeek-V4.1-Flash-1017 — a point update to a model it only released in September, with a version number that sounds like a patch note. Hugging Face co-founder and chief science officer Thomas Wolf captured the reaction on X: “It is truly funny that DeepSeek calls it a ‘small update’ after boosting Terminal Bench performance from 31.2 to 42.3 points.”
That jump — roughly a third better on a hard agentic coding benchmark, in about a month — is the story in miniature. The update improved on all 11 of DeepSeek’s reported benchmarks, beat OpenAI’s GPT-5.6 Sol and Anthropic’s Claude Opus 5 in four of those 11 categories in DeepSeek’s own evaluation, and did it while cutting the cost of a typical coding task by 60%. The weights are open under an MIT license. And the API price is not a rounding error away from the frontier labs — it is roughly one-fortieth of Claude’s.
Call it a small update if you like. The economics are not small.
What actually shipped on October 17
The release is broader than a new checkpoint:
- The model itself: V4.1-Flash-1017 is available immediately on DeepSeek’s website, in its app, and through its API, as an update to the September V4.1-Flash release, with enhanced coding, agent and reasoning performance.
- Open weights, permissive license: the weights are published under the MIT License through DeepSeek’s official repositories on Hugging Face and ModelScope — meaning any team can download, inspect, self-host and build commercial products on them.
- Vision built in: the model handles text and images, with reported gains on visual benchmarks including BabyVision with tools (89.6 to 92.5) and Chartography with tools (78.9 to 83.0).
- Search tools opened up: alongside the model, DeepSeek opened its Search and Fetch APIs to all developers — the plumbing agents need to pull live information, not just generate text.
- Tuned for agent harnesses: DeepSeek says the model is optimised for its own DeepSeek Harness and for Tencent’s CodeBuddy and WorkBuddy environments — a sign the target workload is not chat, but multi-step coding agents.
AI researcher Nathan Lambert summed up the positioning: although it carries a minor version number and date, “the magnitude of performance improvement is significant,” adding that with built-in vision “it may prove to be a stronger model than GLM 5.3. Has the king of open weights returned?”
The benchmark jump, in DeepSeek’s own numbers
Every figure here is from DeepSeek’s published evaluation — company-reported, not yet independently verified, a caveat we will come back to. Against its own previous version, the gains are consistent rather than cherry-picked:
- DeepSWE v1.1 (coding): 77.0, up from 74.2 — and ahead of the 74.0 DeepSeek attributes to Claude Opus 5 on the same test.
- Terminal-Bench 4.0 (agentic terminal work): 42.3, up from 31.2 — the single biggest jump in the table.
- ProgramBench: 29.1, up from 20.3.
- Automation-Bench v1.0.6: 60.5, up from 54.8.
- Agents’ Last Exam: 35.4, up from 31.8.
- Humanity’s Last Exam (no tools): 42.6, up from 36.8.
- ZeroBench-main (vision, with tools): 51.0, up from 49.0.
The pattern matters more than any single score. The largest gains are exactly where enterprises are spending: agents that operate terminals, run automation workflows and write production code over many steps. A model that gets a third better at agentic coding, in a month, at this price, is a budget line item.
The price is the headline
Benchmarks can be argued with. Invoices cannot.
DeepSeek’s off-peak API output price is $0.60 per million tokens. OpenAI’s GPT-5.6 Sol charges $20 for the same million. Claude Opus 5 charges $25. DeepSeek’s output is therefore about one-forty-second the price of Claude’s — and this is before the efficiency gains. DeepSeek also reports that the cost per coding task in its internal evaluation fell 60%, from $0.50 to $0.20, because the new model completes tasks faster and generates fewer output tokens to get the same work done. The company has not published an exact figure for the time and token savings, so treat the direction as claimed and the per-task cost as the concrete number.
This is why the release is landing with practitioners even where it loses on paper. Early user discussion has converged on a pattern familiar to anyone running agents in production: use the fast, cheap model as the default for the bulk of the work, and keep an expensive frontier model on tap for the hardest steps. When the default is 40 times cheaper and, on several coding tests, as good or better, the expensive model stops being the workhorse and becomes the specialist. That is a structural change in how AI budgets get spent, not a leaderboard shuffle.
The caveats are not footnotes
Three cautions deserve the same prominence as the scores:
- These are internal evaluations. DeepSeek ran the comparisons, chose the harnesses, and published the table. Wolf himself added that he is curious whether the Terminal-Bench figure holds outside DeepSeek’s own environment. Independent replication will take days to weeks.
- Four of eleven is not eleven of eleven. The new Flash model beat GPT-5.6 Sol and Claude Opus 5 in a minority of categories. On the rest, the frontier models still lead. “Outperforms US models” is true with an asterisk the size of the claim.
- Benchmarks are a weak proxy for your workload. As we argued earlier this month, standard coding benchmarks have largely stopped predicting real-world fit. A 77.0 on DeepSWE tells you the model is strong on DeepSeek’s chosen software tasks; it tells you nothing about your codebase, your latency requirements, or your compliance constraints.
There is also a geopolitical wrinkle worth noticing. This MIT-licensed, download-it-anywhere release landed one day after China proposed adding large models themselves to its export control catalogue, with pre-export registration for listed technologies. The proposal is a catalogue and a comment period, not a ban — but the juxtaposition is striking: Beijing is drafting rules to register frontier model exports at the same moment its best-known open lab is giving one away to the entire world, free, under the most permissive license in the industry. Which of those two impulses wins is one of the defining questions of the next two years.
The market has seen this film before
DeepSeek is the lab that, with its R1 reasoning model last year, knocked roughly 17% — about $589 billion in market value — off Nvidia in a single day, by raising the obvious question: if this can be done this cheaply, why is everyone spending this much? The question has not gone away. The Nasdaq fell about 2.1% on October 16 amid renewed doubts about AI investment profitability, and because DeepSeek announced over the weekend, the market’s verdict on this release has not yet been delivered.
The counterargument is Jevons’ paradox: cheaper AI tends to expand total usage rather than shrink total spend, because work that was uneconomic at $25 per million tokens gets done at $0.60. Both things can be true at once — per-task economics collapse while total compute demand grows — and for most businesses, the per-task number is the one that matters.
What builders and buyers should do with this
- Benchmark on your own tasks, this week. The weights are free and the API is cheap. Run your real coding and agent workloads through it, measure cost per completed task — not per token — and compare against what you pay now. That experiment costs less than the meeting to discuss it.
- Adopt the two-tier pattern deliberately. A cheap, fast default model with a frontier model reserved for hard cases is now the rational architecture. Design your routing, fallbacks and quality checks around it instead of paying frontier prices for routine steps.
- Count the open-weights option honestly. MIT-licensed weights mean self-hosting, data control and no per-token bill — but also your own infrastructure, security review and evaluation burden. For regulated or cost-sensitive workloads, run the full comparison, not just the API price.
- Do not rip up contracts on a vendor’s benchmark table. Switch the default where your own tests justify it, keep the frontier option wired in, and let independent evaluations land before making exclusivity decisions.
The bottom line
DeepSeek did not release a new frontier model this weekend. It did something arguably more disruptive: it made last month’s model a third better at agentic coding, cut the cost of a coding task to twenty cents, published the weights under MIT, and priced output at one-fortieth of the US frontier. When improvements like this arrive as patch notes, the price of capable AI is no longer set by the labs at the top of the leaderboard. It is set by whoever is willing to give the capability away and charge almost nothing for the rest.