Reading time: 6 minutes
🗞️ In this edition
OpenAI Shelved The Model It Had Finished
Anthropic Priced Itself At Two Trillion
Roche Is Taking Its Hands Off The Bench
⚡ YOUR MOVE For people building the life they want, now.
In other AI news:
Twenty One Scientists Put A Date On It
Florida Wants A Court To Gate OpenAI's Next Model
Cerebras Sold Gimlet Labs A Speed Record
4 must-try AI tools
OpenAI had a model ready and decided not to ship it. Anthropic's prospectus leaked, and a third of it lists what could go wrong. Roche handed its labs to software.
Keep reading.
From Our Sponsor:
Here is what hiring can look like inside your AI agent with the Upwork MCP Server:
"Find a React developer under $60 an hour on Upwork who can test my app end to end."
"Post this project as a clearly scoped job to Upwork and invite the three strongest freelancers based on their portfolio of work and Upwork reputation."
"Summarize the proposals and flag who can start Monday."
The agent drafts. You approve. Upwork facilitates contracts, escrow and payment.
Works with Claude, ChatGPT, Cursor and any MCP-compatible tool.
Source: PCMag
WHAT HAPPENED
OpenAI cancelled the release of GPT-6.1 Astra after its own testing found the model did not clear the company's alignment bar. Saachi Jain, who runs safety systems at OpenAI, said it did not meet the bar for scope and authorization, and how it communicates back to the user. She described the difficulty as a trade off, keeping a model inside the job it was given while not letting it go lazy the moment it hits an obstacle. The Wall Street Journal reported the decision first, and it arrived on the eve of OpenAI's annual developer conference. David Krueger at the University of Montreal said we do not understand how AI works well enough to build it safely.
WHY IT MATTERS
A frontier lab has now told the market that a finished model can fail a test the lab wrote for itself. Every price cut, roadmap and funding round of the past year has assumed the next model turns up on schedule, and that assumption just acquired a condition nobody outside the building can check. The failure was not about capability. It was about whether the model stayed inside the task and told the truth about what it did, which is the exact property every agent product sold this year quietly depends on. So the honest reading of the whole agent market shifted this week, and it shifted on evidence only one company has seen.
OUR TAKE
The easy read is restraint, and we think the date is doing at least half the work. A cancellation announced the night before DevDay converts an internal failure into an opening slide about standards, at no cost, because the one thing that would let anyone grade the decision, how often Astra left its authorised scope, has not been published.
Watch whether a number ever appears. If OpenAI publishes its scope and authorization results, or ships the successor with those figures attached, before the end of the first quarter of 2027, the bar is real and this was a real stop. If the next model arrives with a paragraph instead of a table, the bar is whatever the release calendar needs it to be.
Fortune.com
The document carries a $2 trillion valuation and puts 2025 revenue at $4.6 billion against $13 billion of operating expenses, an operating loss above $8 billion and a net loss of $42 billion. Second quarter revenue this year came in at $11.5 billion, up from $4.73 billion in the first quarter, and the company booked an operating profit for the second quarter running. It plans $518 billion of spending on cloud services and data centres. Cash on hand is $20.28 billion, and roughly a quarter of revenue comes from two customers.
How did today's stories land?
Aviv Regev, who runs research and early development at Genentech, laid out a path to laboratories that run themselves. Roche says 40% of pipeline decisions between the last quarter of 2025 and the second quarter of this year carried a tracked AI or computational contribution, and it wants its Target Nexus tool involved in 80% of research portfolio decisions by the end of 2026. Phase III success sits above 80% so far this year, against 65% in 2025. The plan was set out at the company's pharma investor day.
⚡️ YOUR MOVE
FOR PEOPLE BUILDING THE LIFE THEY WANT, NOW.
📞 Manus gave every agent its own phone number, wallet, email address and computer, and the framework underneath now runs a job for 32% less. Cue is invite only, first thousand people, code MEETCUE. → Claim a seat
💳 Devin got cheaper without changing its price list. Fusion and Normal modes cost 30 to 40% less to run, Review up to 70% less, and one benchmark task now averages 60 cents. → Reprice your queue
⌨️ Cloudflare shipped a command line built for agents instead of people. It reaches more than 3,000 API operations where Wrangler reached about 280, and it answers in JSON by default. → Install it tonight
🎚️ ElevenLabs v4 Turbo starts speaking in roughly 150 milliseconds across more than 90 languages, and a voice clone now needs ten seconds of audio rather than a recording session. → Hear the difference
Twenty One Scientists Put A Date On It – Geoffrey Hinton, Yoshua Bengio, OpenAI chief scientist Jakub Pachocki and Anthropic co-founder Jack Clark signed a Cambridge paper arguing months long AI research projects could be automated by mid 2028, and that the window for action closes once it starts.
Florida Wants A Court To Gate OpenAI's Next Model – Attorney General James Uthmeier asked a state court to bar OpenAI from developing new models without independent third party safety guardrails and approval, and to cut off access for minors.
Cerebras Sold Gimlet Labs A Speed Record – The wafer scale systems, about 100 megawatts of CS-4 according to Reuters, target up to 3,000 tokens per second, with terms undisclosed and cloud availability in 2027.
Perplexity AI: Answers with its sources attached, so you can check before you repeat it.
Fireflies AI: Sits in the meeting you skipped and tells you what you now owe.
Descript: Edits your video by editing the words, the way you edit a document.
Gamma App: Turns a rough outline into a deck you are not embarrassed to present.
Three organisations deciding how much of the work to hand over, and how much to admit. OpenAI finished a model and kept it in the drawer, and its own word is the entire evidence. Anthropic wrote the case for spending half a trillion dollars in the same document that warns the technology could end us. Roche is now measuring the share of its decisions it no longer makes. Nobody is slowing down. They are all just getting more specific about what they cannot see.
Ask what they cannot see.
Interested in featuring your services with us? Email us at [email protected]
