DeepSeek ships V4.1-Flash-1017 under MIT license, claims Opus 5-beating agentic scores at a tenth of the cost

DeepSeek's October 17 point release updates its 552B-parameter September model. The company says it activates only 16 billion parameters per token. Its headline claim is that it beats Anthropic's Claude Opus 5 on several agentic benchmarks at under one-tenth the per-task cost: 77.0 on DeepSWE v1.1 against Opus 5's 74.0, and 60.5 on Automation-Bench V1.0.6, up from 54.8. DeepSeek also reports faster task completion and shorter outputs, which matter as much as raw scores for agent economics.

The model adds self-feedback reinforcement learning aimed at end-to-end coding and 3D-modeling workflows. Weights are published on Hugging Face under the permissive MIT license. API pricing is $0.30 input and $1.20 output per million tokens, with a 1M-token context window. In parallel, DeepSeek opened its proprietary Search and Fetch APIs to third-party developers at $0.75 per 1,000 calls. That positions it as a low-cost grounding layer for agents, not just a model vendor.

The release lands in an unusually dense week for open weights: Meta's Muse Spark Open, Reflection's Beam, NVIDIA's Nemotron 3.5 Super VL and Ant's Ling 3.1 Flash all shipped within days. DeepSeek's pitch is the cost frontier rather than the absolute capability frontier, and its MIT license is looser than Meta's Muse Model License.

There are caveats. The benchmark numbers come from the vendor, and developers are openly debating whether the Opus 5 comparison will survive independent agentic evals. Reliability is another question: Perplexity's sources note five DeepSeek outages in October alone. The silent push into the live API, with no announcement, also means production users got a model swap without notice. Watch for third-party DeepSWE reproductions and uptime data over the next week.