The DeepSeek V4.1-Flash-1017 update is live, and it is mostly a coding and cost story. DeepSeek says this new checkpoint beats its September V4.1-Flash on every reported benchmark, keeps the 1M-token context, and opens its search tools to all developers. All the scores are DeepSeek’s own, so treat them as claims until independent tests land.
Here is the short version before the detail.
- API name:
deepseek-flash, live on the DeepSeek site, app and API. - Context: 1M tokens, with up to 384K tokens of output, per MarkTechPost.
- Price: $0.30 per million input tokens and $1.20 per million output tokens at peak, half that off-peak.
- Search and Fetch APIs: $0.75 per 1,000 successful calls.
- Free window: server-side calls from DeepSeek Harness, Claude Code or Codex cost nothing until November 1, 2026.
If you only build with models, the pricing and the free window are the parts that affect your bill.
What changed in DeepSeek V4.1-Flash-1017
This is a post-training update, not a new model family. DeepSeek calls it a “Pareto improvement in performance and cost,” which means better results without a higher price. The company also says it finishes tasks faster and writes shorter outputs than the September release.
The architecture stays the same. It is a multimodal mixture-of-experts model with a 552B backbone plus 196B Engram parameters, and it activates only 8B parameters per token during prefill and 16B during decode. That small active footprint is why the price can stay low.
Reasoning effort is now a 1 to 100 scale. The API’s low, high and max tiers map to 50, 75 and 100, so you can trade speed for depth per request.
How the benchmarks compare with Claude Opus 5
The headline result is DeepSWE v1.1, where DeepSeek reports 77.0 against Claude Opus 5’s 74.0. That is a real win for a model this cheap. The table shows it is not a clean sweep.
| Benchmark | DeepSeek V4.1-Flash-1017 | Claude Opus 5 |
|---|---|---|
| DeepSWE v1.1 | 77.0 | 74.0 |
| Terminal-Bench 4.0 | 42.3 | 51.8 |
| ProgramBench | 29.1 | 37.0 |
| NL2Repo-Bench | 68.2 | 75.3 |
| HLE (no tools) | 42.6 | 56.3 |
Opus 5 still leads on terminal work, repo-level tasks and raw reasoning. Also, DeepSeek ran Terminal-Bench in Claude Code and DeepSWE in Mini-SWE-V2, so setups differ between rows. One more oddity: MarkTechPost’s summary says “all 11 benchmarks” while its table lists nine rows.
DeepSeek V4.1-Flash-1017 price and the Search API
Official peak pricing is $0.30 per million cache-miss input tokens and $1.20 per million output tokens, with off-peak rates at half. AI Weekly lists $0.15 and $0.60, which would match the off-peak rate. Third-party hosts also quote their own numbers, so check DeepSeek’s pricing page before you budget.
The bigger news for builders is the new Search and Fetch endpoints. Search returns 1 to 50 ranked results, each with a rerank score and an authority tier, and Fetch returns a page as Markdown. Both cost $0.75 per 1,000 successful calls, empty results and failures are free, and the limit is 5 requests per second per account.
Because server-side calls from Harness, Claude Code and Codex are free until November 1, now is a cheap time to test agent workflows. Plan for the rate card afterward, since post-deadline pricing for those tools is unconfirmed.
Open weights: what is confirmed and what is not
The September V4.1-Flash weights sit on Hugging Face as deepseek-ai/DeepSeek-V4.1-Flash under an MIT license. AI Weekly says the 1017 weights were posted under MIT too, but MarkTechPost says the license and hardware needs for 1017 are not disclosed. Read the model card yourself before you ship anything.
That gap matters if you plan to self-host. A model this size needs serious GPU memory, and no reliable hardware guide exists yet. If you want other open models to compare, see our look at Reflection AI Beam’s open weights and Mistral Large 4’s price and weights.
Who should actually care about this update
Developers running coding agents get the most out of it. A cheaper model that scores well on repo-fixing tasks cuts the cost of every retry, and long agent loops add up fast. The 1M-token context also lets you feed in large codebases without chopping them up.
Casual chatbot users will notice less. Shorter outputs and faster replies help, but you will not see a night-and-day change in everyday questions. And if you work in a regulated field, the unclear license and data questions around open weights are reasons to wait for the model card and your legal team.
Search teams should also look at the new endpoints. Ranked results with an authority tier are useful for research agents that need to decide which sources to trust, and a flat per-call price makes costs easy to predict.
Frequently Asked Questions
Is DeepSeek V4.1-Flash-1017 free to use?
The chat site and app are live, but API use is paid. The one free piece is server-side search and fetch calls from DeepSeek Harness, Claude Code or Codex until November 1, 2026.
Does it beat Claude Opus 5?
On DeepSWE v1.1, DeepSeek reports 77.0 versus 74.0. Opus 5 leads on the other four benchmarks in the table above, and all numbers are self-reported.
Which API model name do I use?
Use deepseek-flash. It is also listed on OpenRouter as the 1017 snapshot.
Can I run it on my own hardware?
Not easily. The weights are large and DeepSeek has not published hardware guidance for this checkpoint, so wait for the model card details and community tests.
My take: if you run coding agents, spend an afternoon testing deepseek-flash against your current model before November 1, while search calls are free. Judge it on your own repos, not on DeepSeek’s chart. If it holds up at these prices, it is cheap enough to become your default for routine tasks, with a stronger model kept for the hard ones.
Leave a Reply