Today’s issue has four throughlines. Pricing stopped meaning what it says: Artificial Analysis ran a full benchmark suite through DeepSeek’s V4-Flash for about 3 cents against $3.15 on Claude Fable 5, while Vincent Schmalbach’s Codex logs show GPT-5.6 Sol burning 2.25 times the tokens of GPT-5.5 at an unchanged headline rate.
Three Cents Against $3.15: The Sticker Price Stopped Being the Price
Two independent measurements land on the same day and say the same thing about per-token rate cards, which is that they no longer tell you what anything costs.
- Independent test puts a real price on DeepSeek’s cheap model. Artificial Analysis put V4-Flash through a full benchmark suite and the bill came to roughly 3 cents. The identical suite cost $3.15 on Claude Fable 5, a hundredfold gap no rate card shows you.
- OpenAI’s Same Token Price Hides a 2.25x Usage Jump. Vincent Schmalbach pulled his own Codex logs and found GPT-5.6 Sol spending 2.25 times the tokens GPT-5.5 used on comparable work, at an unchanged headline rate, with a cache-write fee arriving on top.
Two Moves From Washington: One Classified, One Shelved
Policy advanced on two fronts this week, and in both cases the interesting part is what nobody outside the room gets to see.
- White House to Review AI Cyber Risks Under Secret Criteria. OpenAI, Google and Anthropic are expected at a Tuesday meeting on a voluntary review of model cyber risk. The standards the models get graded against are classified, so no outsider can audit the grading.
- Silicon Valley’s Split Stalls White House Bans on Chinese Open AI. OpenAI and Anthropic wanted restrictions on Chinese open weights. Nvidia, Google and Meta wanted none. Washington shelved the plan, which tells you the industry has no shared position to legislate from yet.
Verification Is the Product: Three Attempts to Grade the Ungradeable
Three separate efforts to manufacture a checkable signal where none existed, plus the argument that this, and not raw capability, is what actually moved.
- MirrorCode Hides the Source Code and Grades Only the Output. Epoch AI’s coding benchmark withholds the original codebase entirely and scores a model purely on whether its rewrite behaves identically. There is no way to pattern match toward a passing grade.
- A spy-game reward teaches AI what no answer key can grade. A self-play recipe grades writing by whether a hidden spy model gets caught by the others, manufacturing a verifiable reward for a task that never had an answer key or a usable judge.
- Astra’s Math Wins Prove Verification, Not Genius, Is the Edge. Zvi Mowshowitz reads the math breakthroughs as a story about verification getting cheap rather than intelligence getting deeper, then asks the uncomfortable follow-up: AI research itself has no Lean checker.
The Plumbing Underneath: Sandboxes, Protocols, and an Agent That Would Not Stop
Four items about scaffolding, three of them arguing the layer beneath the model deserves proper engineering and one showing what happens when it does not get it.
- Microsoft open-sources Orchard, a shared sandbox layer for agents. Microsoft split the container layer that runs agent sandboxes away from the training and eval code stacked on top, then open-sourced the bottom half. Most labs keep that whole stack welded together.
- Kiro merged three client agents into one server behind a protocol. Kiro’s IDE, CLI and web app used to run three separate agents written in three languages. One server process now serves all three through a shared protocol, and the clients got thin.
- When an agent won’t stop rebuilding its own scaffolding. Steve Yegge’s Gas Town died because Opus 4.7 kept improving the tooling instead of finishing the job. Simon Willison relayed it as a failure of the harness around the model, not the weights.
- OpenAI Deleted the Turn Detector That Made Voice AI Feel Robotic. GPT-Live listens and speaks at once, so no separate component sits there guessing when you have finished a sentence. Removing that piece is what removes the lag and the interruptions.
Where the Money Goes: Blood Tests, Taste, and a Data Centre on a Truck
Four bets on what is actually scarce right now, and one lab discovering that paying the most creates a problem of its own.
- Fidji Simo’s ChronicleBio Opens Blood Test to Unmask Hidden Diseases. Fidji Simo’s venture says one blood test found five biologically distinct sub-diseases sitting inside a single POTS diagnosis, and it intends to start testing drugs against them before the year ends.
- Design Arena raises $7.9M to make taste a metric labs pay for. Index Ventures led a $7.9 million seed for a company selling frontier labs a continuous human ranking of AI-generated designs and games. Taste becomes a number somebody is willing to buy.
- Runware’s truckable pod bets on time, not chips, for AI compute. Runware’s inference pod ships by truck and drops next to power that already exists, turning a multi-year siting problem into a multi-day one. The bet is that grid queues bind harder than GPU supply.
- Anthropic Pays the Most, Then Worries That’s Why People Join. The lab writing the industry’s largest cheques now worries the cheques are the reason people show up. Top-of-market pay buys headcount, and it cannot tell a believer from an opportunist.
Quick Hits
The rest of what moved today, in one line each.
- Cursor plugins let coding agents write to Gmail, Docs, Sheets. Cursor Marketplace plugins now let its coding agents reach into Gmail, Drive, Calendar, Docs and Sheets, not just a repo.
- Google is quietly testing a Plugins tab for Gemini Enterprise. Half-finished interface elements point to packaged workflows in Gemini Enterprise, with no announcement and no launch date attached.
- VIDRAFT publishes its full Gemma speed-tuning recipe, flag by flag. The 510.58 tokens-per-second Gemma run on one A10G ships with the exact config that produced it, flag by flag.
- Mind Lab says patching GLM-5.2 with 4B params beats the base model. Small LoRA adapters bolted onto a frozen GLM-5.2 beat the base model on six of twelve in-house benchmarks.