Nvidia open-sources Flora, a low-cost engine for labeling robot video

Nvidia released Flora, an open robot video labeling engine that undercuts GPT-6 Astra on cost. Moonshot plans open weights for Kimi K3.1 Pro on November 4, and Anthropic is urging teams to fold subagents into a single agent. Reports also point to OpenAI using deep discounts to win Anthropic customers.

1. Nvidia open-sources Flora, a robot video labeling engine

   Flora labels an hour of robot video for 72 cents against $88 with GPT-6 Astra, at slightly higher quality, though it sometimes falls short on descriptive accuracy.

   Models: GPT-6 Astra

2. Kimi K3.1 Pro tech blog says open weights arrive November 4

   Moonshot's tech blog says open weights for Kimi K3.1 Pro, a post-trained version of K3, are planned for November 4; day 0 platform support comes from a Reddit post.

   Models: Kimi K3.1 Pro

3. Anthropic says a single agent can replace multiple subagents

   Anthropic cites a comparison where one agent loading context on demand ran about twice as fast as multiple subagents, and it has published a migration kit on GitHub.

4. OpenAI offers discounts of up to 90 percent to win Anthropic's enterprise customers

   A report says Rippling, Replit and DataCamp cut AI costs by up to 90 percent after switching to OpenAI, suggesting OpenAI is trading margins for market share.

5. OpenMobile-2 trains open smartphone agents across more than 110 apps

   An open training environment with over 110 apps, including 35 replicas of popular ones, gives open smartphone agents what the authors call the top score on real Android devices.

6. Meta researchers test small models that read raw bytes instead of tokens

   Byte-level models near 1B parameters improved faster than token models, but token models stayed ahead in actual runs; claims of beating Llama 3.2 1B and Gemma 3 1B are extrapolated.

7. Jack Clark moves his forecast for self-improving AI to 2027

   Fast progress on the P-Zero Research benchmark for autonomous experiment design has led Anthropic co-founder Jack Clark to expect self-improving AI by 2027, according to reports.

8. Qwen3.8-Flash-Next 125B runs at about 20 tok/s on 16GB VRAM via NVMe offloading

   A community project offloads experts to NVMe SSD so the 125B model reaches about 20 tokens per second on an RTX 4090 with 8GB system RAM; the benchmarks are self-reported.

   Models: Qwen3.8-Flash-Next

Written by Claude (claude-opus-5-5) from the stories below.