DeepSeek V4.1-Flash-1017, Reflection Beam Weights Released, Claude Code v2.1.302

A minor update to September's V4.1-Flash. DeepSeek says it raises reasoning, coding, agent and vision scores while cutting task time and output length, which lowers cost. It has a 1M-token context and the weights are MIT-licensed. In the API, use the name deepseek-flash; the old deepseek-v4-flash name now routes to this model. The same day DeepSeek opened Search and Fetch APIs built on its own search engine, plus a developer preview of an agent harness.

By DeepSeek's own model-card numbers, Terminal-Bench 4.0 went from 31.2 to 42.3Vendor claim and HLE without tools from 36.8 to 42.6Vendor claim. Teams building agents can plug in the same vendor's search and fetch without a separate service. Old model names now route to the new model automatically, so check whether your results shift.

Reflection AI's Beam is a sparse mixture-of-experts (MoE) model with 501B parameters in total and 23B active per token. It targets coding and agent work, and the weights are Apache 2.0. The 8-bit and 4-bit versions fit on one eight-GPU server of H100, H200 or GB300 chips; the full 16-bit version needs tens of GPUs. The hosted service is free during the public beta.

Reflection says Beam matches GLM-5.2 on advanced reasoning tests while using 3 to 4 times less inference computeVendor claim. That's worth a look if you self-host agents and are pricing it out. But its Terminal-Bench 2.1 score of 80.1% trails DeepSeek V4.1 Flash and Kimi K3Vendor claim, and fewer tokens doesn't automatically mean a cheaper task.

Fork subagents are now on by default in claude -p, the Agent SDK, the desktop app and VS Code. To turn them off, set this environment variable:

CLAUDE_CODE_FORK_SUBAGENT=0

You can choose the model auto mode uses to classify risky commands with CLAUDE_CODE_AUTO_MODE_MODEL. On third-party providers, sessions with no permission mode set now start in default.

If you run claude -p in CI or scripts, subagent behavior and token use may change after the update. On a third-party provider, check whether your permission mode shifts to default. A prompt cache miss after ToolSearch loads a tool is also fixed.