DeepSeek-V4.1-Flash-1017 Lands
DeepSWE 77.0, Price Unchanged
An in-place replacement for September's V4.1-Flash: same API name, same price, DeepSWE v1.1 up from 74.2 to 77.0 and now above GPT-6 Astra and Claude Opus 5.5.Oct 17, 2026
On October 17, 2026 DeepSeek shipped DeepSeek-V4.1-Flash-1017, replacing the V4.1-Flash from September 10 while leaving both the API name (deepseek-flash) and the pricing untouched. DeepSWE v1.1 rose from 74.2 to 77.0 — above GPT-6 Astra (74.1) and Claude Opus 5.5 (74.2). Claims of a "crushing win" or an inevitable market crash, however, do not all survive contact with the primary sources.
- Verified Confirmed in an official blog, repository or primary document
- Vendor-reported The developer's or vendor's own claim, not independently verified
- Third-party tested Independently checked by an outside developer, outlet or measurer
- AIfrontier view Our own interpretation
1. What happenedVerified
DeepSeek announced the new version in its API docs and changelog on October 17. Source: DeepSeek API Docs (Oct 17)
- The model name stays
deepseek-flash. Calling convention and pricing are unchanged — the same name now serves the new weights. - The default model on the web and app was updated too.
- DeepSeek's own framing is a "performance-cost Pareto improvement": better benchmarks plus shorter task runtime and shorter outputs.
- Optimised for DeepSeek Harness, Tencent CodeBuddy and WorkBuddy; Harness
v0.2.1-rc.1or later is required. - Weights are open, published on Hugging Face and ModelScope.
- The same day, the in-house Search & Fetch API opened to everyone at $0.75 per 1,000 calls. Server-side tool calls through DeepSeek Harness, Claude Code and Codex are free for API users until billing starts on October 31, 2026 at 16:00 UTC.
The release replaces the contents of an existing endpoint rather than adding a model. Application code does not change, but behaviour does — so shifting benchmarks and output length can reach production unnoticed. If you hold regression tests, now is the moment to run them.
2. What improvedVendor-reported
DeepSeek's published figures, against the September 10 V4.1-Flash. Differences in parentheses.
| Benchmark | V4.1 (Sep 10) | V4.1-1017 (Oct 17) | Change |
|---|---|---|---|
| DeepSWE v1.1 | 74.2 | 77.0 | +2.8 |
| HLE (no tools) | 36.8 | 42.6 | +5.8 |
| Terminal-Bench 4.0 | 31.2 | 42.3 | +11.1 |
| ProgramBench | 20.3 | 29.1 | +8.8 |
| NL2Repo-Bench | 65.4 | 68.2 | +2.8 |
| Automation-Bench v1.0.6 | 54.8 | 60.5 | +5.7 |
| Agents’ Last Exam | 31.8 | 35.4 | +3.6 |
| Chartography (with tools) | 78.9 | 83.0 | +4.1 |
| BabyVision (with tools) | 89.6 | 92.5 | +2.9 |
| ZeroBench-main (with tools, Pass@5) | 49.0 | 51.0 | +2.0 |
The largest gains are on Terminal-Bench 4.0 (+11.1) and agentic work. The harnesses are stated explicitly: Claude Code for Terminal-Bench 4.0, Mini-SWE-V2 for DeepSWE. HLE is measured on the pure-text subset (42.6, asterisked 44.0). Source: DeepSeek changelog
3. Drawing the line around "crushing win"Third-party tested
On DeepSWE v1.1 alone, the field looks like this. Argon is the figure this site recorded for the September 30 launch.
| Model | DeepSWE v1.1 | Input / output (per 1M tokens) |
|---|---|---|
| DeepSeek-V4.1-Flash-1017 | 77.0 | $0.30 / $1.20 (off-peak $0.15 / $0.60) |
| Gemini 4 Argon | 77.9 | $2 / $10 ($4 / $20 from Dec 1) |
| Claude Opus 5.5 | 74.2 | $4 / $20 |
| GPT-6 Astra | 74.1 | $10 / $50 |
| Claude Fable 5.1 | 67.4 | not confirmed on this site |
AIfrontier view Precisely: (1) on DeepSWE v1.1 it beats GPT-6 Astra, Claude Opus 5.5 and Fable 5.1; (2) it does not quite reach Gemini 4 Argon (77.9); (3) the cost gap is an order of magnitude. On input pricing it is roughly 1/33 of Astra and 1/13 of Opus 5.5, or about 1/67 off-peak. "Crushing win" is fair as a statement about cost; as a statement about capability it rests on one benchmark, so it does not mean best on every axis.
Comparisons circulating on social media put the cost at about $0.20 per task versus roughly $7.5 for GPT-6 and $22 for Fable 5, a 40×-100× gap. DeepSeek's own October 17 post does include a chart of DeepSWE v1.1 Pass@1 against cost per task, and the $0.20 figure plausibly comes from that series — but we could not read the plotted values, and Fable 5's pricing is not confirmed in a primary source here. If you plan to quote numbers, check the chart yourself or measure on your own tasks.
4. Architecture and priceVendor-reported
The model card describes a 552B backbone plus 196B of Engram parameters, activating only 8B per token during prefill and 16B during decode — an asymmetry that pays off on input-heavy workloads. Context is 1M tokens and the licence is MIT. Source: Hugging Face
So the active-parameter figures of 8B and 16B are correct. But the total is 552B plus 196B, which is not a laptop model. "An 8B model caught the frontier" conflates activation with total size.
5. "DeepSeek shock again" is a forecast, not a factAIfrontier view
On January 27, 2025, the arrival of DeepSeek R1 triggered a selloff in US tech: Nvidia fell about 17% in a day, losing roughly $600 billion in market value. That was the "DeepSeek shock." Nvidia rebounded 8.8-8.9% the next session, with commentators calling the selloff an overreaction even then.
Predictions that the same playbook will repeat are circulating. They are predictions, not observations. As of this writing (October 18) there is no primary evidence of how Monday October 19 will trade. Three things can be said objectively.
- This time is different in kind. It is not a new model launch but an in-place swap on an existing API, at unchanged prices. It does not show unit prices collapsing — if anything, "same price, better model" is tailwind for AI demand.
- Cheaper inference cuts both ways. Lower cost per task widens the addressable use case. The investor worry is over-investment in capacity, and falling unit prices alone do not establish that.
- The warning was already on the table. On October 13, 2026 the IMF flagged high valuations and leverage as risks to AI-linked equities. The caution predates this release.
We therefore do not join the "Monday will definitely crash" framing. A directional forecast and its confidence are separate things. If you want to prepare, the useful move is not to bet on the forecast but to check how exposed your holdings are to a scenario in which AI capital spending slows.
6. What it adds up toAIfrontier view
What is established: an open-weight model now matches frontier-priced models on coding and agentic benchmarks at an order-of-magnitude lower unit cost. DeepSWE 77.0 falls just short of Argon's 77.9, but input pricing is about 1/33 of Astra. "Nearly identical performance, two orders of magnitude apart in price" is the same pattern that played out in the decision-model layer.
Three practical implications. (1) There is still a lot of headroom for moving high-volume agent runs down to cheaper tiers. If DeepSWE-class performance is available at $0.30 per million input tokens, much of what goes to Argon or Kimi K3 for routine work can be substituted. (2) Watch in-place endpoint swaps. Without regression tests you cannot detect the behaviour change. (3) Search & Fetch billing starts October 31. Workflows that assume free server-side search should measure usage before the cost lands.
Sources
- DeepSeek API Docs, "DeepSeek-V4.1-Flash-1017 is here" (Oct 17, 2026) https://api-docs.deepseek.com/news/news261017/
- DeepSeek API Docs changelog (Oct 17, 2026 and Sep 10, 2026) https://api-docs.deepseek.com/updates/
- Hugging Face, "deepseek-ai/DeepSeek-V4.1-Flash" (552B + 196B Engram, 8B prefill / 16B decode, 1M context, MIT) https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash
- Hugging Face, "deepseek-ai/DeepSeek-V4.1-Flash-1017" https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017
- OpenRouter, "DeepSeek V4.1 Flash" https://openrouter.ai/deepseek/deepseek-v4.1-flash
- Reuters, "DeepSeek sparks AI stock selloff; Nvidia posts record market-cap loss" (Jan 27, 2025) https://www.reuters.com/technology/chinas-deepseek-sets-off-ai-market-rout-2025-01-27/
- Reuters, "Nvidia shares jump, rebounding after DeepSeek-linked drop" (Jan 29, 2025) https://www.reuters.com/technology/nvidia-shares-jump-rebounding-after-deepseek-linked-drop-2025-01-29/
- The Guardian, "Threat of AI share crash puts world financial stability at risk, IMF warns" (Oct 13, 2026) https://www.theguardian.com/business/2026/oct/13/ai-share-crash-risk-global-imf-debt-corrowing-cyber-attacks
- This site, the Gemini 4 Argon article (DeepSWE comparison values: Argon 77.9 / Astra 74.1 / Opus 5.5 74.2 / Fable 5.1 67.4)