DeepSeek V4.1 Flash 1017 Cuts Cost While Doubling Token Window

DeepSeek rolled out V4.1 Flash 1017, a 1‑million‑token LLM that promises faster completions at a fraction of the previous price. The model appears on DeepSeek’s website, app and API as of the latest update, and it ships with a new Reinforcement Learning with self‑feedback loop that the company says improves end‑to‑end project development.

The announcement arrived alongside a pricing sheet that lists $0.15 per 1 M input tokens and $0.60 per 1 M output tokens on RunInfra. The same sheet notes a context window of 1,048,576 tokens, enough to hold an entire codebase or a long design document in a single prompt. Billing for the free Search & Fetch calls starts on Oct 31 2026 at 16:00 UTC, a date that will push developers toward the paid tier.

Performance claims and cost trade‑offs

DeepSeek describes the new version as a “Pareto improvement” over its predecessor. Benchmarks show shorter outputs, higher throughput and lower latency on the same hardware. In practice, the model finishes a typical coding task in half the time it took V4.0, according to internal tests shared on Hacker News.

The cost side is equally stark. At $0.15 per 1 M input tokens, a 10 K‑token prompt costs less than a cent. Output at $0.60 per 1 M tokens means a 2 K‑token response runs under $0.002. Those numbers undercut many OpenAI‑compatible services that charge $0.02 per 1 K tokens for similar output quality. The pricing sheet makes clear that the lower rates apply only while the free Search & Fetch calls remain active; after Oct 31 2026 the free tier disappears.

Integration ecosystem and developer impact

DeepSeek has bundled the model with several of its own tools. The release notes advise upgrading to DeepSeek Harness v0.2.1‑rc.1 or later to unlock the full feature set. Harness, CodeBuddy and WorkBuddy now route their server‑side Search & Fetch calls through the new model without extra cost, as long as the free window holds.

For developers who already use Claude Code or Codex, the API offers OpenAI‑compatible chat completions and Anthropic‑compatible messages on the same endpoint. That design choice lowers the barrier to swapping in Flash 1017 for existing pipelines. The model’s massive context window also lets a single request span multiple files, a convenience that previously required chunking logic.

Competitive context and market positioning

The LLM market is crowded with providers that tout larger context windows. Anthropic’s Claude 3 and OpenAI’s GPT‑4 Turbo both sit near the 100 K‑token mark, far short of DeepSeek’s 1 M tokens. That advantage matters for developers building code generators, game prototypes or 3D scene descriptions that exceed typical prompt lengths.

DeepSeek’s pricing undercuts the premium tier of most rivals. However, the free Search & Fetch period ends in late 2026, and the company has not disclosed any SLA guarantees. Competitors may respond with longer free windows or volume discounts, but the immediate price gap is enough to draw budget‑conscious startups.

The model’s reinforcement‑learning self‑feedback loop also differentiates it technically. While OpenAI and Anthropic rely on human‑annotated RLHF, DeepSeek claims the loop runs autonomously on each request, tightening the feedback loop between generation and evaluation. Independent verification of that claim is still pending.

What to watch

The next quarter will reveal whether DeepSeek can sustain its pricing advantage after the free Search & Fetch window closes. Watch the billing metrics announced for Oct 31 2026 and any follow‑up pricing adjustments. Also keep an eye on community benchmarks that compare Flash 1017’s latency and output quality against GPT‑4 Turbo and Claude 3 on real‑world coding tasks. If DeepSeek maintains its cost edge while delivering on the promised 1 M token context, it could force a broader shift toward ultra‑long‑context models in the developer tooling space.