The cost of AI coding just fell 50x — and the model is free to download
DeepSeek's new open-source AI model handles coding tasks for around $0.20 each, compared to $2–$10 on rival services, which changes the economics for any team building with AI.
- 77%
- score on the DeepSWE coding benchmark
- ~$0.20
- cost per AI coding task (rivals: $2–$10)
- MIT
- license — free for commercial use
- Oct 31
- last day for free search API access
On October 17, 2026, DeepSeek — the Chinese AI lab that keeps surprising the industry — released a new model called V4.1-Flash-1017. You can download it for free and use it in commercial products. Tasks that cost $2–$10 on models like Claude Opus 5 or GPT run for about $0.10–$0.20 here, meaning rivals charge 10 to 50 times more for the same work.
Part of why it can be this cheap is how it was trained. DeepSeek used a technique called reinforcement learning with self-feedback: the key savings come from removing human judges from the training loop — instead, the model rates its own outputs and learns from that. DeepSeek says the result beats its predecessor on every task it measured, scoring 77% on DeepSWE — a standard test for AI coding agents.
What makes this release different from most open-weight drops is built-in web search. DeepSeek opened its own search engine to developers as an API — $0.75 per thousand lookups, free until October 31. That means any AI agent you build can fetch live results from the web rather than rely on training data that stops at a fixed date. Live web access was the missing piece for open models, and it just got filled.
For teams building software with AI — whether for clients or for their own products — the question is no longer whether a capable model exists at low cost. It does. Test it on your actual work first — pick a real workflow, run it, and judge the output honestly. If it holds up, the savings are worth acting on.