DeepSeek V4.1 Flash 1017 arrived on Saturday, October 17, with unchanged API prices and MIT-licensed weights that anyone can download. For US developers building coding agents and document-processing tools, that pairing is the news: a long-context model that costs a fraction of what American frontier labs charge per token.
What DeepSeek V4.1 Flash 1017 changed
On October 17, DeepSeek upgraded V4.1 Flash to V4.1-Flash-1017 across its web, app and API products. According to AICoder, developers call it as deepseek-flash, using the same call and the same price. The company also published the weights under the MIT license on Hugging Face and ModelScope.
DeepSeek describes the update as a checkpoint with stronger reasoning, coding and visual performance. NewsBytes reports a score of up to 42.6 on Humanity's Last Exam, up from 39.1 for the previous version. That figure is an upper bound, and the reporting did not publish the evaluation settings behind it.
The model carries a 1 million token context window and allows up to 384,000 output tokens, according to Requesty. Requesty describes it as a 552-billion-parameter mixture-of-experts model with native vision, which means only part of the network activates for any given request. Developers Digest says it replaces the earlier V4 Flash model.
The context window may matter more than any single benchmark. A team can send a large codebase or a stack of long contracts in one request instead of splitting them into chunks, which simplifies pipelines that otherwise chunk and reassemble documents.
Official pricing versus third-party hosts
Pricing is the central pitch, so the rates need careful reading. The table separates DeepSeek's own rate card from third-party hosts. All figures are per million tokens.
| Endpoint | Input | Output | Notes |
|---|---|---|---|
| DeepSeek official, off-peak | $0.15 | $0.60 | Cached input $0.003, half the peak rate |
| DeepSeek official, peak | $0.30 | $1.20 | Cached input $0.006 |
| OpenRouter | $0.08 | $0.50 | Third-party host |
| Requesty, discounted | $0.10 | $0.29 | Third-party host |
| Requesty, list price | $0.20 | $0.60 | Third-party host |
DeepSeek's official endpoint doubles its rates during peak hours, so the cheapest official call depends on when your traffic runs. Batch jobs that can move to off-peak hours would cost half as much per token.
Third-party hosts price the same weights differently. OpenRouter lists $0.08 input and $0.50 output per million tokens, below DeepSeek's own off-peak rate. Requesty shows a discounted endpoint at $0.10 input and $0.29 output, against a list price of $0.20 and $0.60. These are third-party rates, not DeepSeek's, and host prices can change.
DeepSeek also opened Search and Fetch APIs at $0.75 per 1,000 calls, according to AICoder. Server-side calls from DeepSeek Harness, Claude Code and Codex are free until October 31, 2026, at 16:00 UTC. That window is temporary, so budget as if those calls will cost money once it closes.
How it stacks up against Claude Opus 5
The benchmark figures come from DeepSeek and the outlets that repeat its numbers, not from independent labs. On DeepSWE v1.1, KIE.ai reports 77.0 for the 1017 checkpoint, 74.2 for September's version and 74.0 for Claude Opus 5, as cited by DeepSeek. On Terminal-Bench 4.0, the same reporting gives 42.3, 31.2 and 51.8 for the same three entries.
That result splits the picture. DeepSeek holds a narrow lead on DeepSWE, but Claude Opus 5 leads on Terminal-Bench 4.0 by about 9.5 points. For agents that work through terminal tasks, that gap matters more than the DeepSWE margin.
The price gap is wider. KIE.ai lists Claude Opus 5 at $0.50 per million cached input tokens and $25 per million output tokens. That output rate is more than 20 times DeepSeek's $1.20 peak output price. The published comparison covers cached input and output only, so it does not produce a full bill. A team that mostly sends fresh prompts will need its own math.
For buyers, the lesson is practical. A model that trails by several points on one benchmark can still win on cost per completed task, but only if it completes the task reliably. Test it on real work first.
Open weights and the self-hosting math
Open weights change what buyers can do. A company that downloads the model can run it on its own hardware, so prompts and outputs never need to leave its infrastructure. That option separates this release from a simple price cut.
Self-hosting a model of this size is expensive. One cost estimate from BenchLM puts self-hosting at roughly $27,962 a month, against about $1,125 a month through the API for the same usage assumptions. Those assumptions are not fully laid out in the reporting, so treat the figures as a rough benchmark rather than a quote. On that estimate, the API wins on cost by a wide margin at the usage level assumed.
Markets have reacted to DeepSeek before. In January 2025, a cheap DeepSeek model hit Nvidia's stock hard. This week's decline looks broader. According to a CNBC report summarized by AI Weekly, twenty of the most valuable global AI and semiconductor stocks have shed more than $1 trillion. The reporting does not tie that drop to this release, so readers should not assume a causal link.
Security and data questions to answer first
Enterprise security and compliance teams should start with one question: where does the endpoint run? DeepSeek's official API is operated by a Chinese company. Third-party endpoints hosted in the United States may handle data differently, and each one has its own terms. The data-handling terms for DeepSeek's official API were not confirmed in the reporting, so read them directly before sending customer records through any endpoint.
The reporting also did not identify a US government response, such as limits on federal use or procurement. Buyers in regulated or government-adjacent work should watch for one.
Agent permissions deserve the same caution. India's CERT-In has warned that AI agents can read live business records without anyone noticing, as USA 24 reported. Any agent wired into a business system should have its read access reviewed as carefully as a new employee's.
Practical steps before you switch a default model
- Test on your own workloads. Run real coding or document tasks through V4.1-Flash-1017 and your current model, then compare accuracy, latency and cost. Replit's move to make OpenAI the default model for its agent shows how fast a default can change, so set up the test before any switch.
- Compare peak and off-peak rates on the official endpoint, and schedule batch work for off-peak hours where you can.
- Check where each endpoint runs and what its data terms say before you send sensitive information.
- Mark October 31, 2026, at 16:00 UTC if you use Search or Fetch through Claude Code, Codex or DeepSeek Harness, since the free server-side window closes then.
- Read the MIT license before you self-host or redistribute the weights.
What to watch next
Independent labs need to reproduce the DeepSWE, Terminal-Bench and Humanity's Last Exam results before the vendor figures carry real weight. Watch whether third-party hosts hold their rates, whether DeepSeek extends the free server-side window and whether US agencies issue guidance on its use.
Competition is moving too. Google made Gemini 4 Argon globally available on October 13 through its API in AI Studio and the Gemini platform. Forbes reports that Argon's introductory price of $2 and $10 doubles after November 30. A developer who locks in a default model this month may face a different bill in December.
