DeAI Daily Brief — 5 Oct 2026

Today in DeAI: four hyperscalers retire hosted models in the same mid-October window, Axios reports a Reflection AI open-weight model is close, Anthropic flips its API default to Sonnet 5.5, and Z.ai's GLM-5.3 weights are confirmed live on Hugging Face.

Four clouds retire hosted models in one two-week window

Google Cloud's deprecations page lists 16 open-model Model-as-a-Service endpoints on Vertex AI — DeepSeek, GLM, gpt-oss-20b, Kimi K2 Thinking, Llama 3.3 70B, MiniMax M2, E5, and five Qwen3 variants — all retiring October 21, 2026, with self-deployment on Model Garden as the documented alternative. OpenAI's deprecations page shows 12 legacy GPT snapshots and five fine-tuned lines shutting down October 23, routed to GPT-5.6 Sol, Terra, and Luna. Microsoft's Foundry schedule retires gpt-4.1-nano October 14 and shows Kimi-K2.7-Code already retired October 3; Snowflake's release notes put claude-4-sonnet and openai-gpt-4.1 at end-of-life October 14. Why it matters: hosted convenience carries a lifecycle clock as well as a token bill, and for the open-weight half of the list the documented migration path is running the weights yourself. (Sources: Google Cloud, OpenAI, Microsoft Learn, Snowflake) — our full coverage

Axios reports Reflection AI's frontier open-weight model is close

Crypto Briefing relays Axios's October 4 report, citing unnamed sources, that the Nvidia-backed startup founded by ex-DeepMind researchers Misha Laskin and Ioannis Antonoglou is preparing an open-weight model pitched as competitive with top Chinese open-weight systems while initially trailing leading US closed models, monetized through an "AI factory" offering. Reported compute deals with Nebius and SpaceX and a valuation of up to $25B remain secondary-outlet figures; Reflection declined comment, and no release date has been announced. Why it matters: a well-funded US entrant would widen a day-one checkpoint supply that has run through Kimi, GLM, DeepSeek, and Qwen — but nothing here is verifiable until a Hugging Face repository or announcement post exists. (Source: Crypto Briefing) — PULSE: the discourse, labeled

Anthropic flips the API default to Sonnet 5.5 at unchanged $2/$10 pricing

Anthropic's product page states Sonnet 5.5 is priced the same as Sonnet 5 at $2/$10 per million tokens in/out but "typically needs far fewer tokens to do the same work" — an efficiency-based cost claim, not a sticker-price cut, and one builders should measure on their own traces. Anthropic has not announced a breaking parameter change for the flip. Why it matters: a bill can move without any published price moving, which is a reason to pin model IDs and watch your own token counts through a default transition. (Source: Anthropic)

GLM-5.3's full weights are live — and went up weeks before the "two weeks" framing

Z.ai's GLM-5.3 repository on Hugging Face shows the full 753B-parameter FP8 weights, public and not gated, with 1.42M downloads. The repo was created August 25 and last modified September 4, so the practical state is that the most-discussed frontier coder is self-hostable today; a custom "glm-5.3" license applies, worth reading before depending on it. Why it matters: a 753B model is multi-node territory, so the hosted-API question collapses to hardware economics and license terms, not availability. (Source: Hugging Face)

Google ships externally verifiable differential privacy in Gboard federated learning

Google Research's blog describes a federated learning system that runs the privacy-relevant computation inside trusted execution environments so anonymization guarantees are externally verifiable and auditable, with Gboard already adopted at production scale. The "provably private" framing is Google's own, and the paper itself discusses that TEE attestation covers the code path, not side channels. Why it matters: verifiability is becoming the bar privacy claims get measured against, which sets the standard for TEE-hosted inference venues. (Source: Google Research)

Watching tomorrow

StepFun's Step 5 Preview, a 600B mixture-of-experts model with an API and open weights, is slated to land October 15 — the next checkpoint release-watch.