DeepSeek V4.1-Flash-1017: The Cheapest Serious Coding Model Just Got Better

Two days after a five-hour outage knocked DeepSeek's API offline, the Chinese lab shipped DeepSeek-V4.1-Flash-1017 — a faster, leaner update to its cheapest AI model. DeepSeek calls it a "minor update." The numbers tell a different story.

What's new

Flash-1017 is an updated post-training checkpoint of DeepSeek V4.1 Flash (launched September 10, 2026). Same architecture, same size, same price — new weights only. It ships under the same API name (deepseek-flash), so existing integrations were moved over automatically.

The benchmark gains, from DeepSeek's own published evaluations:

• DeepSWE v1.1 (agentic coding): 77.0, up from 74.2 — ahead of Claude Opus 5 (74.0) and GPT-5.6 Sol (73.0) on this test
• Terminal-Bench 4.0: 42.3, up from 31.2
• Humanity's Last Exam: 42.6, up from 36.8

Raw reasoning still belongs to the frontier models: Opus 5 leads 56.3 to 42.6 on HLE. This is the budget workhorse, not the frontier champion.

Why the price is the real story

DeepSeek's API costs $0.15 per million input tokens and $0.60 per million output tokens (off-peak; peak weekday hours are double). Per DeepSeek's own cost chart, a coding task completes for just under $0.20 at the top-effort setting, down from ~$0.50 — and a low-effort setting still scores above 72% at ~$0.10 per task. Rivals sit between $2 and $10+ per task.

Shorter outputs are part of the trick: the model is tuned to finish tasks with less rambling, and every ungenerated token is money saved.

Open weights change the risk math

The weights are published under the MIT licence — commercial use allowed. That's the real answer to the October 15 outage (five hours, 8:11–13:04 Beijing time, a full week's API spend refunded as credits): teams that self-host never notice DeepSeek's downtime.

DeepSeek also opened a Search API and Fetch API at $0.75 per 1,000 successful requests, free via server-side tooling until November 1, 2026 — a clear bid to wire agents into its infrastructure.

Who should care

If you run coding agents at scale, there is no cheaper way to get serious agentic-coding scores right now. If you need top-tier raw reasoning, Opus 5 still wins. And if DeepSeek's uptime worries you — self-hosting is built into the release.

---

This is a condensed explainer. The full breakdown — including the KV-cache technical story, the outage details, a 3-step how-to, and the honest caveats — is on NeurAITool: DeepSeek V4.1-Flash-1017: Faster AI After the Outage.