Google retires Gemini 2.5 Pro and Flash — Dec 1 silent reroute

2.5 Pro requests silently become Flash — plus Anthropic open-sources proof-of-model, Cloudflare slips billing, Appwrite prunes idle deploys.

Nowline NOV 4 2:00 PM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • Your pinned 2.5 Pro calls become Flash on Dec 1

    Google deprecated gemini-2.5-pro and gemini-2.5-flash on Nov 2; from Dec 1 requests auto-route to gemini-3.8-flash and gemini-3.5-flash-lite at those models' rates (2.5 Flash-Lite is spared). Pin an explicit current model ID now, or your 'Pro'-quality calls will quietly land on a lighter model next month.

  • Anthropic open-sources a proof of which model answered

    FlashProve ships under Apache 2.0 — a prover, a fast C++ verifier, and a Lean verifier with machine-checked security proofs — using a zkSNARK to prove a reply came from a committed model without revealing its weights. A one-time ~39-minute commitment covers ~16M requests and proving runs ~110 tok/s on a single H100 (roughly 72–94x GPU overhead, 1,024-token cap), so you can hand users a cryptographic receipt of the model behind a response.

  • Cloudflare AI Search delays billing — and quietly slows sync to 24h

    Billing now starts Dec 1 instead of Nov 1, so usage until then is free, but new AI Search instances now default to a 24-hour sync interval (up from 6), meaning changed content can sit a day stale. Drop it toward the 15-minute floor via the instance's Sync interval setting or the sync_interval API field if your index has to stay fresh.

  • Appwrite starts deleting idle deployments Nov 5

    From Nov 5, Appwrite Cloud deletes function and site deployments that have been inactive for more than 30 days by default. If you depend on an old build you haven't redeployed in a month, redeploy it before it's pruned.

  • CrowdStrike ships a kill switch for rogue AI agents

    Falcon Guardian's new AI kill switch lets admins halt a misbehaving agent, block its relaunch, and cut its network access — available now, with endpoint-level termination still rolling out. If you ship autonomous agents into managed environments, assume security tooling may stop a run mid-flight and design for clean interruption and resume.