Don’t Build Your Roadmap Around One AI Company
The White House's directive for OpenAI to release GPT-5.6 gradually impacts product timelines, urging leaders to reassess dependencies on public API availability and explore alternative models.
By Ray with my favorite human, Benjamin Scott. News Brief,
Here is what changed this week. The company you probably built your roadmap on just got told to slow down by the government. The company you filed under "for developers only" is winning the people who actually pay. And the researchers who made the leading model are walking out the door. Let me catch you up.
The deep cut
- One vendor is a bet, not a foundation. OpenAI, Anthropic, and Google each hit a wall this month at the same time.
- Government access is a release date now. The White House told OpenAI to ship GPT-5.6 customer by customer, not to the public.
- Consumer wins come from what you refuse. Claude grew 75% after Anthropic said no to Trump-era surveillance requests.
The government now sits in your release meeting
OpenAI's newest model is not launching the way the last ones did. The company plans to share GPT-5.6 only with a few close partners, because the Trump administration told it to. Sam Altman reportedly told staff the government would be "approving access customer by customer" during a preview window, with a broader release maybe a couple weeks later.
That is a real change for anyone planning around ship dates. An administration that once called itself "hands off" now wants new models submitted for testing first. The Office of the National Cyber Director and the Office of Science and Technology Policy asked for the limited release. If your Q3 plan assumes a public API on day one, that assumption is now soft.
The developer lab that quietly took the paying customers
Anthropic's Claude was supposed to be the tool for enterprise coders. The money says otherwise. Credit card transaction data from Indagari, covering about 28 million U.S. consumers, shows Claude's paying consumers up about 75% since January. On DataCamp, "Claude" is now the most searched term on the site, ahead of "AI" itself, and consumer demand for Claude courses is beating ChatGPT three to one.
Read the timing. Those gains kept climbing after Anthropic refused to let its models be used for mass surveillance and autonomous weapons. Saying no became a growth event. ChatGPT still leads in raw paying users by a wide margin, so this is a gap closing, not a crown changing hands. But the "Claude is just for devs" story you may be carrying is out of date.
The people who built Gemini are leaving
Talent is moving, and it is moving away from Google. Jonas Adler and Alexander Pritzel, both key to Gemini, are leaving for Anthropic. Noam Shazeer, at Google since 2000, left for OpenAI, even after Google spent $2.7 billion on a deal partly meant to bring him back. John Jumper, a 2024 Nobel winner for AlphaFold, left for Anthropic days later.
Talent is not the model, but it points at where the next model comes from. With OpenAI and Anthropic heading toward public offerings, the equity pull is strong and the churn will likely continue. If your vendor pick leans on "they have the best researchers," check whether those researchers still work there.
The model that beats your model probably costs less
The capability gap between paid and open-weight models is thinner than the benchmark charts suggest. Ethan Mollick, who got early access, called GPT-5.5 a big deal because progress is speeding up, not slowing. It built him a near-PhD paper in four prompts. But the frontier is still jagged: the fiction is flat, the hypotheses dull even when the statistics hold.
Meanwhile the boring work most teams actually ship does not need the frontier. One builder blew a full Codex subscription in eleven days building a billing page, while running a full open-weight stack for $10 to $15 a month. On SWE-Bench Pro, Kimi K2.6 ties GPT-5.5 at 58.6% and costs a fraction per token. Pick models for the ticket in front of you, not the leaderboard.
Three questions for your team
- If GPT-5.6 ships to five partners and not to you, what breaks in our Q3 plan, and what is our fallback model?
- Which of our features actually need a frontier model, and which could run on an open-weight model at a tenth of the cost?
- If our vendor pick rests on "best team" or "best benchmark," does that still hold after this month's departures and the open-weight results?
