StepFun Step 5 Preview: a 600B agent model at $1/M, weights Oct 15

A cheap, long-horizon coding model you can call now and self-host in three weeks. Plus: GitHub axes six Copilot models Oct 19, and MiMo ships a 20x speed tier.

Nowline SEP 24 6:00 PM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • Step 5 Preview: callable now, weights in three weeks

    StepFun's 600B sparse-MoE (27B active per token, 1M context, text and images in) is live on the API and Studio at $1/M input — just $0.05 on cache hits — and $2.70/M output, with BF16 weights promised Oct 15, so you can prototype now and move to self-host later. Vendor benchmarks look strong (DeepSWE v1.1 67.7%) but independent Artificial Analysis puts it mid-pack (Terminal-Bench 4.0 33.3%, 24th of ~200 models), so pressure-test it on your own agent loop before you commit.

  • GitHub retires six Copilot models on Oct 19

    Copilot is dropping Gemini 3.7 Flash, GPT-5.5, GPT-5.4, GPT-5.4 mini, GPT-5 mini, and Grok 4.5 across every surface on October 19 — anything pinned to those in scripts, extensions, or CI needs to move to the recommended replacements (GPT-5.6 Sol/Luna, Gemini 3.8 Flash, Grok 4.6). Enterprise and Business orgs get the swaps auto-enabled unless an admin has turned them off.

  • Update: MiMo-V2.6 adds a 20x UltraSpeed tier and a half-price batch API

    Xiaomi filled out this week's MiMo-V2.6 launch with a Pro UltraSpeed tier that advertises up to 20x faster generation (live in Desktop, via the API, and on OpenRouter) and a Batch API that runs async jobs at 50% of real-time rates in OpenAI- and Anthropic-compatible formats — the cheap, fast path for evals, labeling, and bulk generation.