Today in DeAI: Runway publishes Praxis-1 robot-control weights under Apache 2.0, Anthropic's Sonnet 4.5 retirement lands November 30, and Google doubles Gemini 4 Argon pricing on December 1.
Runway ships Praxis-1 open weights for robot control
Runway released the weights of Praxis-1, its 27-billion-parameter mixture-of-experts world action model, under Apache 2.0 with inference and fine-tuning code attached. Runway reports 46.2% success on NVIDIA's RoboLab-120 benchmark across 1,200 episodes, ahead of FLUX 3 Action at 42.9% and π0.5 at 28.0%, though The Robot Report notes the model ranks fifth overall on the leaderboard. Why it matters: a video-generation lab just made a robotics-adjacent foundation model fully self-hostable, from H100 down to an RTX 5090 and Jetson Thor, with no cloud API in the loop. (Runway) — our full coverage
Anthropic: Sonnet 4.5 retires November 30, Opus 4.5 no sooner than November 24
Anthropic's model-deprecations page sets claude-sonnet-4-5-20250929 retirement at November 30, 2026, with claude-opus-4-5-20251101 listed at "not sooner than November 24, 2026." The same page records that the sampling parameters temperature, top_p, and top_k were removed in API version 2026-10-21 and now return 400 errors. Why it matters: that is a hard migration deadline inside two weeks for anyone still on Sonnet 4.5, and a breaking interface change on top of it. (Anthropic)
Nex-AGI publishes Nex-N2.6, a 321B agent fine-tune of GLM-5.3-Flash
Nex-AGI's model card publishes FP8 weights of about 321 billion parameters for a computer-use and coding agent built on open-weight GLM-5.3-Flash, claiming OSWorld 83.0 against the base's 78.0. That comparison comes from the publisher and has not been independently reproduced. Why it matters: this is the open-weight flywheel in one artifact, a third party differentiating on another lab's open base and publishing the result back. (Hugging Face)
qvac releases BotPsy-0.8B for on-device tool calling
qvac, backed by Tether, published BotPsy-0.8B, a function-calling agent model built on Qwen3.5-0.8B, under Apache 2.0 with GGUF builds for llama.cpp. The company claims large benchmark jumps over the base model, including BFCL v4 moving from 23.4 to 44.0; those are vendor self-reports pending independent harness runs on the published checkpoints. Why it matters: reliable tool calls on phone-class hardware is exactly where small models have been failing. (qvac)
Google doubles Gemini 4 Argon pricing on December 1
Google's AI Studio pricing page holds Gemini 4 Argon at $2.00/$10.00 per million input/output tokens only through November 30, 2026, rising to $4.00/$20.00 from December 1, with cache reads doubling from $0.10 to $0.20. This is the second dated Google price increase on the calendar, after Gemini 3.8 Flash's scheduled January 1, 2027 increase. Why it matters: introductory rates are converting to permanent ones on a published schedule, so December budget math should use the new numbers, not the ones in your spreadsheet. (Google AI Studio)
Watching tomorrow
DeepSeek's metered tool-call billing ($0.75 per 1,000 successful requests for Search and Fetch, per third-party trackers) deserves a primary-source check against DeepSeek's own rate card, and Anthropic's tentative Opus 4.5 date is the first thing to re-verify on Monday.
Sources
- Praxis-1 Is Now Available: Open-Weight Robotics Model — Runway
- Model deprecations — Anthropic
- Nex-N2.6 model card — Hugging Face
- qvac BotPsy-0.8B on-device agent — qvac (Tether)
- Google AI Studio pricing — Google
About DeAI
DeAI is an independent publication covering open-weight AI models, private inference, and decentralized infrastructure — the tools for running AI you actually control. We test providers on price, privacy, and refusal behavior and publish the numbers, not the vibes. DeAI is powered by Morpheus (mor.org), a decentralized inference marketplace, and covers it on the same terms as every other provider.
Powered by Morpheus and StrandCMS
Morpheus is a decentralized inference marketplace, covered on the same terms as every other provider — we rank it wherever the data lands. StrandCMS is the open-source, agent-first framework this site is built on.
