Every week in AI brings a launch. This week brought a price collapse, a public listing, and a decision about access that I think will matter longer than either.
For a long time the frontier story was about capability. This week the story moved to cost and to ownership. What does intelligence cost per finished task, and who gets to buy the most sensitive version of it.
The price war stopped being polite
On November 18, Google released Gemini 4 Neon. It costs 3 dollars per million input tokens and 15 dollars per million output tokens, with a context window of one million tokens. Google says it ranked first on legal research and tax benchmarks, and it scored 61 on the Artificial Analysis Intelligence Index. That price is less than half the 8 dollars per million input tokens charged for the top models from OpenAI and Anthropic.
On November 16, Alibaba released Qwen4-Max and reported a reduction of about 60 percent in average cost per task compared with its predecessor, using about 35 percent fewer tokens. The same day, Tencent launched Hy4 Flash at 7 cents per million input tokens. Tencent said satisfaction in its own blind tests of office tasks approached that of a much more expensive rival, and the company plans to release the model weights.
Read those moves together. Cost per task is now the contest. A model that uses fewer tokens can win even at the same list price, and a model that costs cents can take the routine work and leave the expensive model only the hard steps.
Builders no longer ask only which model is smartest. We ask what a finished task costs, and which steps deserve the expensive brain.
OpenAI put the cyber model on open sale
On November 17, OpenAI released GPT-6.2 Astra and GPT-6.2 Sol. Astra is the flagship for professional, scientific, and coding work, at 8 dollars per million input tokens and 40 dollars per million output tokens. OpenAI says that input price is 20 percent lower than its predecessor. Sol is the everyday model, at 2 dollars per million input tokens and 10 dollars per million output tokens.
The access decision is the bigger news. Two weeks earlier, OpenAI had limited its strongest cyber model to approved security teams. This week, Astra and Sol both carry the company Critical rating for cybersecurity, and they are on open sale, with a monitoring system called Control Room in place of a restricted list. Any defender must now assume an attacker can buy the same tools.
On performance, keep the claims in their lanes. OpenAI says Astra more than doubled the pass rate of Claude Fable 5.2 on a legal benchmark built by Harvey, at about one sixth the cost per task, and led on finance and consulting work. An independent index puts Astra at 62 and Fable 5.2 at 65. Both can be true. A model can trail on a general index and still win on cost per finished task.
Anthropic became a public company
On November 18, Anthropic priced its initial public offering at 83 dollars a share, above the planned range of 70 to 78 dollars. Trading began on Nasdaq on November 19 under the ticker ANT, and shares opened at 111.50 dollars. Coverage from CNBC and Fox Business put the raise at about 81.3 billion dollars and the fully diluted valuation at about 1.5 trillion dollars.
I read a listing like this in a practical way. A public AI lab has to show its numbers on a schedule. Pricing and compute commitments stop being rumors and become filings. Your supplier is now answerable to public markets every quarter. And when one frontier lab trades in public, every private valuation in the field gets compared against a real market price.
What this means if you build
First, design your product so the model can change. Prices moved by half, and in one case to cents, in a matter of days. Routing work between a strong model and a cheap model is no longer an optimization. It is basic defense.
Second, measure cost per finished task, not cost per token. A model that finishes in fewer steps can be the cheaper model even at a higher list price.
Third, update your security assumption. When a cyber capable frontier model moves from a vetted list to open sale, the safe assumption is that both sides have it. Defense has to be built for that world, not for the world where only the good teams held the strong tools.
I build in public, so my own numbers stay open to check. My Instagram audience at @milandadon stands at 77,152 followers, my Twitch audience stands at 15,206 followers, and my Threads posts drew 1,018,440 views across the trailing 30 days. Proof matters for a founder too, and proof should be current or it should not be quoted.
My take
This week, intelligence went on sale and one of its makers went public. The builders who win the next year will treat cost, routing, access, and trust as part of the product, not as paperwork after the product is done.
At Manifest Anything, where players describe a game and the system builds it, that is the daily work. Choose the right model for each step, keep the cost of a finished creation low enough for real people to use, and keep a record of what the system did.
If your conference or team is working through what these shifts mean in practice, I speak on building with AI in the open. Book me for a keynote here.
