Cloudflare's Clef-omni: open-weight decision model for audio+video
A 30B MoE that scores audio and video with no transcription step; Cloudflare also cut Clef-flash to $0.038/M, and Devin now bills GPT to your ChatGPT plan.
Copy markdown
A decision model that hears and sees
Clef-omni is a 30B MoE (3B active) that scores text, images, audio (WAV/MP3) and video (MP4/WebM) against your allowed answers in a single pass — no transcription step. It runs ~20ms on text, under 100ms on image or audio, and ~300ms on a 21-second clip with sound, live now as @cf/cloudflare/clef-omni on Workers AI at $0.150/M input (64K context). You can gate, route or moderate on raw audio and video without wiring up a separate speech-to-text stage.
Clef-flash is now cheaper than Jev — with a catch
Hosted Clef-flash input drops to $0.038/M from $0.090, undercutting TypeSafe's Jev ($0.042/M). The trade: the hosted context window shrinks from 64K to 24K tokens (Cloudflare says only 0.24% of requests exceed it). The open weights are unchanged and still handle 256K if you self-host.
Clef itself got ~2x faster, same price
Moving Clef to SGLang roughly halves median latency — 616ms to 305ms at ~3,400 tokens, 2,721ms to 1,635ms at ~16K — on identical weights and unchanged $0.240/M pricing. The speedup rides on SGLang 0.5.22 if you run it yourself.
Devin now bills GPT usage to your ChatGPT plan
Cognition lets you link a personal ChatGPT Go, Plus or Pro subscription; every GPT call Devin makes then draws from that plan's quota instead of your Devin credits (fast/priority variants excluded; needs Devin Pro, Max or Teams). If the plan runs dry it falls back to Devin quota, so sessions keep running.
Grok Bot gets its own email inbox
xAI gave the @bot agent an @mail.grokbot.com address it can use to sign up for services, contact businesses, and schedule meetings on your behalf — team admins have to switch it on. It's a new primitive for agents that need to transact over email, not just chat.