Cloudflare released Clef and Clef-flash, two decision models available on Workers AI with Apache 2.0 weights for local experimentation. They take an input and typed questions, then return probabilities for allowed answers. That makes them relevant to bounded jobs such as routing support requests or deciding when a workflow needs human review. Cloudflare also announced reinforcement-learning fine-tuning for Clef. For builders, the useful experiment is to compare structured decisions against their existing classifier or general-purpose model on real cases, measuring wrong decisions alongside latency. The release provides another option for a specific workflow step; its vendor benchmark claims do not establish performance on every application.
AI news
Cloudflare opens Clef decision models as Google’s TPU satellite reaches orbit
Google said its Project Suncatcher prototype satellite, developed with Planet, launched aboard SpaceX’s Transporter-18 mission and successfully made contact. The experiment will gather evidence on how TPUs cope with launch stress, radiation and temperature extremes. This is an infrastructure research milestone, with important engineering questions still being tested. Learners can use it to connect AI’s growing compute demands with hardware reliability and energy constraints. Builders should read it as an experiment in possible future infrastructure: Google has announced an operating prototype and a measurement campaign, rather than a production service developers can deploy to today.
OpenAI’s dated API notice deprecated GPT-5.3-Codex, GPT-5.1 and GPT-5.4-Nano, with removal scheduled for April 1, 2027. The recommended replacements are GPT-6-Sol for the first two and GPT-6-Luna for Nano. Teams using these identifiers have a concrete maintenance task: inventory dependencies, rerun representative evaluations and prepare migration before the shutdown. Deprecation starts the transition window; it does not mean the endpoints have already stopped working.