AI stopped shipping one model, one format, and one workflow today. It started shipping menus, and the smartest builders spent the day deciding what to trust.
The Frontier Keeps Splitting: New Tiers, New Formats, New Provers
Three separate releases this week point the same direction. Frontier labs are no longer shipping one model, one format, or one proof method. They are shipping menus.
- OpenAI Cuts GPT-5.6 Into Three Priced Tiers. A closed preview shows three tiers, Sol, Terra, and Luna, priced by how much reasoning depth a task needs, with wider access tied to a pending security review in the United States.
- ByteDance Pushes AI Video Past the 3 Minute Mark. App listings point to a beta mode that stretches AI generated video from 30 seconds to 180, a jump that tests whether character identity and motion hold up over six times the runtime.
- Mistral Open Sources a Model Built to Prove Code Correct. Leanstral 1.5, a 119 billion parameter theorem prover, checks formal proofs and verifies code rather than generating it, giving builders a free way to test correctness claims instead of taking them on faith.
Who Controls the Model, Who Controls the Compute
Behind this week’s headlines sits a quieter argument. It is about who gets to run which tools, who actually needs the chips they are selling off, and who owns the reasoning a smaller model inherited.
- Alibaba Bars Staff From Using Claude Code. Alibaba has classified Anthropic’s coding tool as high risk software and will push employees toward Qoder, its own in house alternative, starting July 10.
- Spare Chip Sales Are a Demand Story, Not a Retreat. Jamin Ball argues that SpaceX and Meta renting out idle compute reflects weak model demand at two laggards, not a broader pullback in AI capital spending.
- How a Compression Trick Became an IP Fight. Sergio Paniego traces distillation’s shift from shrinking models to transferring frontier reasoning wholesale, a change that now sits at the center of AI’s biggest copying dispute.
Verification Is the New Craft: What Builders Are Learning About Agents
As agentic coding tools mature, the hard problems have moved. The bottleneck is no longer generating code, plans, or tests. It is deciding what to trust, what to check, and where the real leverage sits.
- The Model Isn’t the Moat Anymore, the Harness Is. Aparna Dhinakaran argues that as coding models commoditize, the control loop wrapping the model, not the model itself, is where durable advantage in agentic coding will live.
- Agents Write Code Faster Than Anyone Can Check It. As coding agents outpace human review speed, the gap between a claim and its proof becomes the real constraint on shipping software, not generation speed itself.
- Find What You Don’t Know Before You Build It. Thariq Shihipar, an Anthropic engineer, argues teams should use Claude to surface unknowns before building, not discover them mid project when fixes cost far more.
- Stop Telling the Model How to Test Its Own Work. A tip from a Claude Code fireside chat argues that rigid testing rules waste effort the model could instead spend using its own judgment.
Zooming Out: Mapping the Ecosystem’s Slow Layers and Its Gaps
Two analyses this week step back from the news cycle to ask what actually moves fast in AI, what moves slowly, and where the open source stack is still thin.
- Why AI Feels Both Instant and Glacial at the Same Time. Drew Breunig’s fourteen layer map of the AI ecosystem shows why builders bet on months while the industry’s real constraints move on the scale of decades.
- A Map of 24,626 Open Source AI Projects Shows Where the Stack Is Thin. The Gap Map scores 24,626 projects on openness, capability, and adoption, showing exactly where the open source AI stack still has holes.
Today’s Quick Hits
- What $2,000 and $40,000 Buy in Local LLM Hardware. One builder’s real parts list prices out running near frontier models locally instead of on hosted APIs, from entry level to serious rig.
- Google Tests a Triage Inbox Inside Gemini. Spotted builds show a three filter sorting panel for Workspace users, another sign Gemini is shifting from chatbot to email agent.