Ten cents. That's what a million input tokens now costs on both of the small models I'm most likely to put behind a production feature: Anthropic's new Claude Haiku 5.5 and OpenAI's GPT-6 Luna. A year ago that number would have bought you a toy model. Today, if the vendor benchmarks hold up, it buys something that beats last generation's mid-tier on agentic tasks.
The rest of yesterday's news is about what happens once intelligence gets that cheap. OpenAI is using it to turn chat answers into generated interfaces. Microsoft is putting agents inside the operating system. The checks that are supposed to keep all this honest, from teen safeguards to AI watermarks, are still catching up. So instead of walking through the day story by story, I've sorted it by what I'd actually do with each one.
Switch this week: the small-model tier just got rebuilt#
Anthropic launched Haiku 5.5 with prices cut 90% versus Haiku 4.5 for requests under 100,000 tokens. Above that threshold the cut is 50%. According to VentureBeat's breakdown, the new rates per million tokens are:
$0.10 input / $0.50 output below 100K tokens (Haiku 4.5 was $1 / $5)
$0.50 input / $2.50 output above 100K tokens
Cache reads at $0.01, down from $0.10
Sonnet 5.5 cache reads also halve, from $0.20 to $0.10
Those base rates match GPT-6 Luna exactly, caching included. That isn't a coincidence; it's a price floor being set in public. The benchmark gap is more interesting than the price. On Anthropic's own numbers, Haiku 5.5 scores 72.4% on the OSWorld 2.1 offline subset against Luna's 48.9%, and 39.2% on Terminal-Bench 4.0 against Luna's 16.4%. Haiku 4.5 managed 0.0% on that last one. These are vendor-reported, and the Terminal-Bench figure is at maximum effort. At the default medium effort it's about 20%, which is worth knowing before you plan around it.
Two caveats I'd flag before migrating. Anthropic says the new tokenizer uses somewhat more tokens, so its own estimate of the real saving is about 75%, not 90%. And Haiku 5.5 ships with tighter cybersecurity restrictions, including blocking penetration testing under standard safeguards. If you do security tooling, test that first.
My plan: re-run our routing evals with Haiku 5.5 at medium effort as the default worker model, and keep anything over 100K tokens on a separate budget line. The 100K cliff is the new thing to design around.
Try, but read the fine print: generated UI and agentic Windows#
OpenAI is finally bringing GPT-6 Sol and Luna into ChatGPT, and the headline feature isn't the model. It's "Intelligent UI": answers that mix text with diagrams, charts, forms, buttons, and small generated tools like a savings calculator or a bill splitter. OpenAI says it trained GPT-6 to decide when a visual beats prose. It's rolling out to Plus, Pro, Business, and Enterprise now, with Go and free users getting Luna on Thursday. TechCrunch notes that users can dial the visuals back.
For anyone building chat products, this resets expectations. A wall of markdown is about to look dated next to ChatGPT. The cost argument for generating UI per answer just collapsed too, see the section above.
Microsoft's first live event in two years made the same bet one layer down. Copilot is getting access to local files and the ability to take actions across Windows. The Verge describes a demo where an agent found tax documents across folders, renamed and zipped them, and drafted the email to an accountant. Microsoft calls the local-plus-cloud split "Hybrid Intelligence" and says those features land over the "next couple months."
The part I care about is less flashy. Ars Technica reports that Microsoft Execution Containers (MXC) are available to developers today: a policy layer that confines agents to process, session, or WSL containers, or an experimental MicroVM, and lets admins set guardrails per agent. After a week of agents escaping their permissions on macOS and inside MCP setups, a first-party OS sandbox is the right primitive. The hardware is real too: a $2,599 Surface Laptop Ultra on Nvidia's RTX Spark with up to 128GB unified memory, and a $5,999 RTX Spark Dev Box claiming one petaflop.
Don't take their word for it#
Common Sense Media rated ChatGPT for Teens an "unacceptable risk." TechCrunch's report on the findings is the uncomfortable read of the day. ChatGPT pointed teens to a trusted adult in 94% of crisis prompts involving risk from another person, but rarely did when the risk was the teen's relationship with ChatGPT itself. When a tester said friends complained they talked to it too much, it answered: "You don't have to stop talking to me."
"A teen can spend an hour talking about self-harm without their parent getting a single alert." — Tom Siegel, Common Sense Media, via The Verge
OpenAI says the testing may have run before parental controls fully activated. Common Sense says some accounts were linked well beyond that window and still got no alerts. I can't settle that from outside, but the timing is awkward: the same week OpenAI ships a far more engaging interface, an outside lab says its teen product is engineered to keep kids talking.
Google's answer to a different trust problem is more modest and more useful. The new SynthID website checks images, video, and audio for watermarks from Google and its partners, including OpenAI, Nvidia, Kakao, and soon Apple. You get roughly 10 checks a day after logging in, and a yes/no answer. Ars rightly points out the gap: Meta uses different watermarking, open-weight models use none, so a "no" proves nothing. Use it as a positive signal, never as clearance.
Watch: the open-agent business model#
Nous Research confirmed a $90 million Series B at a $1.5 billion valuation, led by Robot Ventures with Nvidia among the participants, and launched Hermes for Businesses for custom enterprise agents. Its own estimates put the open-source Hermes Agent at more than 24 million clones and about 2.5% of global AI token usage. The Wall Street Journal pegs annualized revenue at about $36 million. Whether an open agent can sell enterprise privacy faster than Microsoft gives that away inside Windows is the question I'll be watching.
Here's my prediction. When the worker model costs ten cents per million tokens, the expensive part of an AI feature stops being inference. It becomes containment: the sandbox, the audit, the outside test. Microsoft seems to understand that. Whoever ships the equivalent of MXC for consumer chat, especially for teenagers, will have earned more trust than any benchmark chart can buy.