DeepSeek says its V4.1 Flash update scores 77% on DeepSWE at about $0.20 a task, and opened a search API at $0.75 per 1,000 calls

The number to watch in this release is not tokens per dollar. It is cost per finished task.

What happened

DeepSeek released V4.1 Flash 1017 on October 17, live on its website, app and API, with open weights on Hugging Face and ModelScope under the MIT license. DeepSeek calls it a minor update to V4.1 Flash with better benchmark scores, faster task completion and shorter outputs.

Per a DeepSeek cost chart reported by The Decoder, the model scores 77.0% on the DeepSWE benchmark at an average cost of just under $0.20 per task, and about 72% for roughly $0.10 in its low setting. The previous version scored 74.2% at around $0.50. Rival models in that chart land between $2 and over $10 per task. These are DeepSeek's own numbers, and The Decoder notes independent benchmarks have not come out yet.

DeepSeek also opened its Search and Fetch APIs to all developers, powered by its in house search engine. Each endpoint costs $0.75 per 1,000 calls. Calls through server side tools in DeepSeek Harness, Claude Code and Codex are free for DeepSeek API users until billing starts on October 31, 2026 at 16:00 UTC. The Decoder also notes that China reportedly plans to tighten rules on publishing open weight models.

My take

Shorter outputs are the quiet win here. A model that finishes the same job in fewer tokens is cheaper even at the same price per token, and faster in a workflow that waits on it.

How I would treat this for client work:

1. Measure cost per task on your own workflows, not per million tokens. Run a set of real past jobs through it and compare finished results.
2. The search API is priced for volume. Enriching 1,000 leads with one search and one fetch each is about $1.50 in API fees, before any model calls.
3. Check data rules before routing client CRM data to any new provider. Open weights mean you can also host the model yourself or through a provider your client already approves.
4. Keep the model a setting, not a rebuild. Prices and rankings change monthly.

Treat the benchmark claims as a reason to test, not a reason to switch.