DeepSeek V4.1 Flash: near-frontier open weights, MIT, ~40x cheaper

552B-param MoE, ~1M-token context, weights on Hugging Face, off-peak input at $0.15/M. Plus Codex CLI hardens sandbox trust and OpenAI's Agents API opens up.

Nowline SEP 13 7:00 PM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • The spec: a 552B MoE that runs light

    DeepSeek V4.1 Flash is a sparse mixture-of-experts model — 552B total parameters but only ~8B active per input token (16B on output), so it serves fast. MIT-licensed weights are on Hugging Face, with a ~1M-token context and up to 384K output tokens.

  • ~30-40x cheaper than the frontier

    API pricing is $0.30/$0.15 per 1M input tokens (peak/off-peak) and $1.20/$0.60 output, with cache hits at $0.006/1M — roughly 30x cheaper on input and 40x on output than GPT-6 Astra or Claude Fable 5.1, both at $10/$50. Same task, an order of magnitude off your bill.

  • Where it wins, where it doesn't

    It posts 90.6 on Terminal-Bench 2.1 (ahead of GPT-5.6 Terra's 87.4) and 90.9 on GPQA Diamond, but only 30.0 on the harder Terminal-Bench 3.0 and 65.6 on MathArena Apex. Strong for everyday agentic coding and reasoning; test the hard-agent tasks before you commit.

  • Build this weekend: your own private coder

    MIT weights let you self-host, fine-tune, and ship commercially with no revenue thresholds. The sparse-MoE design plus ~1M context makes a local repo-wide coding agent, or a cheap million-token RAG pipeline, genuinely affordable to run yourself.

  • Elsewhere: Codex CLI 0.154 locks down sandbox trust

    Codex CLI 0.154.0 stops running workspace-controlled helpers before trust is established, and its macOS sandbox now blocks terminal input injection. It also adds ExternalMessage support, so external content can start or join a turn with tool-level authority. Upgrade with pip install --upgrade openai-codex.

  • Elsewhere: OpenAI's Agents API opens to everyone

    OpenAI's Agents API is in public beta: it runs the agent loop on OpenAI's infra — session orchestration, context compaction, sub-agent coordination, and crash recovery — with no fees beyond tokens and tool use. It plugs into E2B, Modal, Cloudflare, Vercel, and Oracle sandboxes.