DeepSeek V4.1-Flash-1017: specs, pricing and Claude Code setup
Terminal-Bench 42.3 in the Claude Code harness, DeepSWE 77.0, 1M context and MIT weights, with $0.60/1M output off-peak plus step-by-step Claude Code setup and free Search/Fetch until Nov 1.
DeepSeek just made its cheap model a lot better: Terminal-Bench 31.2 to 42.3, and output still costs $0.60 per 1M tokens this weekend. What shipped today (DeepSeek-V4.1-Flash-1017): - Terminal-Bench 4.0: 42.3, up from 31.2 (run in the Claude Code harness) - DeepSWE v1.1: 77.0, up from 74.2 - Shorter outputs and faster task runs than the last version - Open weights on Hugging Face, MIT license, 1M context Who gets it: anyone with a DeepSeek API key. If you already call deepseek-flash, you are on the new model now. Old names like deepseek-v4-flash route to it too. Try it inside Claude Code in 3 steps: 1. Get an API key at http://platform.deepseek.com 2. In your terminal: export ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic export ANTHROPIC_AUTH_TOKEN=<your DeepSeek key> export ANTHROPIC_MODEL=deepseek-flash[1m] 3. cd into a project and run: claude First task to try: "Read this repo and list the 3 riskiest functions, with reasons." Pricing per 1M tokens, off-peak: $0.15 input, $0.60 output. Peak is double: weekdays 01:00-04:00 and 06:00-10:00 UTC. Weekends are off-peak all day. Bonus: web search and fetch calls made through Claude Code, Codex or DeepSeek Harness are free until Nov 1 (Beijing time). After that, $0.75 per 1,000 calls, plus the tokens to read the results. Fine print: - These benchmarks are DeepSeek's own table. Test on your repo before you switch. - claude-opus model names map to V4 Pro, billed at the Pro price. - Check the data policy before sending client code or anything private. https://api-docs.deepseek.com/quick_start/agent_integrations/claude_code #AI #DeepSeek