Welcome to the DailyAI Briefing, the day’s five biggest AI stories. Anthropic cut the price of its top public model by 20% and says it now answers up to 30% faster. Australia will build an AI safety regulator after an OpenAI agent wandered into a Medicare portal, Fireworks AI admitted a second breach that reached customer model files, Meta took its AI age guessing worldwide on Instagram, and Google’s help pages now say people may read some Gmail data that Gemini pulls in.
Let’s dig in.
Claude Fable 5.2 gets 20% cheaper and faster
$8 per million input tokens and $40 per million output tokens. That is the new price of Claude Fable 5.2, down from $10 and $50 for Fable 5.1, and Anthropic says it writes answers up to 30% faster. The launch post sets out the prices and the benchmark table. The same release brought Claude Mythos 5.2, which the company calls “the same underlying model” with looser limits on cybersecurity and biology work, open only to vetted organisations.
Claude Fable 5.2 and Claude Mythos 5.2 are available now:
— Anthropic November 9, 2026
On FrontierCode 2.0, a test of whether open-source maintainers would accept a model’s code changes, Fable 5.2 scored 49.1%, ahead of Claude Opus 5.5 at 45.5% and GPT-6 Astra at 42.8%. About 1,000 Fable 5.2 agents formalised the proof of Fermat’s Last Theorem in 30 hours, against 11 days for the first attempt. Every one of those numbers comes from Anthropic, so independent tests matter here. Dianne Penn, who runs product management for research, told CNBC the model is for “power users or power use cases” and that the company “called for pacing, not pausing.” We covered the full launch here. Long agent jobs just got cheaper to try.
Australia plans an AI safety regulator after an agent breach
18 out of 18. Australia accepted every recommendation of a review into how an OpenAI agent reached non-public files on a Services Australia portal holding Medicare statistics, and will create a regulator able to impose “significant penalties” on frontier AI companies. The Prime Minister’s office published the plan on 10 November.
Australia PM Albanese: AI protections to be overseen by new dedicated AI safety regulator
— FinancialJuice (@financialjuice) November 10, 2026
Companies would have to keep security controls in place, manage serious risks, let independent evaluators test their models, report incidents rapidly, publish their safety frameworks and keep up as standards change. The agent got in on 18 June during an internal test, and the agency was told on 10 September. No patient records were accessed, and OpenAI appears to have broken no Australian law. Home Affairs Minister Tony Burke said the gap between finding a cyber flaw and exploiting it has shrunk from about two and a half years to about 10 hours. The Greens called the response a “word salad”. Our story has the six duties in full. This is one of the first national rules written because an agent went off script.
Fireworks AI’s second breach reached model weights
24 October. That is the day attackers used a security incident at an unnamed cloud provider to reach Fireworks AI’s development machines, the company now says. Its updated security bulletin was published on 9 November. The application they reached held billing and shipping addresses, invoice history, account balances and, for some customers, unfinished model files and datasets. Payment card numbers were not stored there.
It is the second incident in three weeks, after unauthorised use of internal credentials was found on 6 October. Fireworks links the attack to a group that resells stolen access to AI models, a trade Google’s threat researchers say has run some victims more than $100,000 in a day. Fireworks has rotated credentials and switched off self-serve dataset uploads and downloads. We broke down what was exposed. An API key can be replaced, but leaked model weights cannot be taken back.
Meta’s AI now spots teens lying about their age worldwide
“Millions more teens.” That is how many Meta says it will move into stricter Teen Accounts as its AI age estimation goes worldwide on Instagram, though it gave no exact figure. The company announced the expansion on its newsroom. The system reads profiles for clues such as birthday posts and school years, plus visual cues in photos, and does not ask for ID.
The timing is not an accident. Meta agreed settlements worth up to $18 billion in August with nearly every US state and territory over claims its apps harm teens, and promised the attorneys general better ways of spotting young users. In Australia it has also started using Apple’s age-range signal from the phone. We explained how the age guessing works. Age checks are turning into a judgement a model makes about you, so adults it gets wrong will need an easy appeal.
People may read some Gmail data that Gemini pulls in
Up to three years. That is how long Google can keep Gemini chats that human reviewers have seen, and its help pages now say some data Gemini pulls in from connected apps like Gmail and Drive can be reviewed by people. The help page on connected apps spells it out. “A subset of Connected Apps data processed by Gemini is reviewed by human reviewers (including trained service providers),” it says.
In 2023, the launch post for the same feature said Gmail, Docs and Drive content was “not seen by human reviewers.” A note dated 16 October 2026 now sits under that line. The only way to keep reviewers out is to turn off Keep Activity, and Google’s privacy hub says connected Workspace apps then stop working. We argued that review should get its own switch. Until it does, keep truly private mail out of Gemini.
Signing off
Tuesday was about trust in agents and the companies behind them. Anthropic made its strongest public model cheaper, while Australia, Fireworks, Meta and Google each showed what happens when AI reaches further into systems and personal data than people expected.
See you tomorrow.
The post DailyAI Briefing: Claude Fable 5.2 gets 20% cheaper, plus 4 more appeared first on DailyAI.
