DeepSeek V4.1 Flash 1017: free Harness/Claude Code/Codex calls until Nov 1
Long post breaks down the launch: DeepSWE 77.0 over Opus 5, Agents' Last Exam 35.4 vs 28.6 but 14 points behind on HLE, $0.003 per million cached input tokens, plus open Search & Fetch APIs.
DeepSeek quietly started giving away inference today. Server-side calls routed through DeepSeek Harness, Claude Code, or Codex don't get billed until November 1. That detail is buried under the benchmark chart in V4.1-Flash-1017's launch post, and it says more about the strategy than the scores do. The scores are still worth sitting with. DeepSWE v1.1 went from 74.2 to 77.0, just ahead of Claude Opus 5's 74.0. Agents' Last Exam climbed to 35.4 against Opus 5's 28.6. Automation-Bench hit 60.5 versus Opus 5's 50.3. Those are benchmarks built around multi-step agent work, not single-turn reasoning, and that is exactly where DeepSeek chose to win. On HLE, pure reasoning with no tools, it still trails Opus 5 by 14 points. This is a model tuned for the agent loop, not the exam. Pricing is where the gap turns absurd. DeepSeek's V4.1-Flash family prices off-peak cache-hit input at $0.003 per million tokens. Claude Opus 5 charges $0.50 for the same cache hit, GPT-5.6 Sol charges $0.40, Kimi K3 charges $0.30. Run 50 million cached tokens through each and you get $0.15 on V4.1-Flash, $15 on Kimi K3, $20 on GPT-5.6 Sol, $25 on Opus 5. For a coding agent that rereads the same repo context every turn, that is not a discount. It is a different cost structure. They also just opened Search and Fetch APIs to every developer at $0.75 per 1,000 successful calls, nothing charged on empty results. Pair that with a harness built to be pluggable, swap any runtime component out like a part, and the actual play comes into focus. DeepSeek isn't chasing best-model-of-the-week. It's trying to become the default cheap call inside every harness willing to route to it. I'd watch what happens on November 1, when the free window closes and the real pricing test starts.