Welcome back! Anthropic dropped its second model in a week, and it's almost as good as the one it launched last Tuesday for half the price. OpenAI scrapped its next model because it wasn't honest about what it did. Florida wants a judge to stop OpenAI from building new models at all. And Anthropic's IPO filing leaked, with a loss number that looks a lot scarier than it is.
In today's Generative AI Newsletter:
Anthropic: How close is Sonnet 5.5 to Opus?
OpenAI: What did GPT-6.1 Astra do that got it cancelled?
Florida: Can a state stop OpenAI from building AI?
Anthropic: How do you lose $42 billion and ask for $2 trillion?
Claude Sonnet 5.5 is almost Opus for half the price
Anthropic released Sonnet 5.5, its second new model in less than a week. It's the mid-size Claude, and it's now nearly as good as Opus 5.5, the big one.
It costs $2 in and $10 out per million tokens. Opus 5.5 is $4 and $20.
Artificial Analysis, the independent benchmark site, gave it 56. Opus 5.5 got 58 last week, the highest score it has ever measured.
30%+ faster than Sonnet 5, and up to 30% cheaper per task
On a long coding test it went from 10% to 71%
It's the first Sonnet to beat Pokémon Red just by looking at screenshots
Risky hacking requests get handed back to the older Sonnet 5
Anthropic still says Opus is clearly better for hard, open-ended work.
A cheaper version, Haiku 5.5, is coming in the next few weeks.
Free to enter. $20,000 to win.
You've read about AI agents all week. Now build one.
Our Open Agent Hackathon runs this October, fully online. Build an agent that does real work for a real person, then let the judges try to break it.
$20,000 in prizes, with $8,000 for first place
Free to enter and open to anyone 16+, anywhere in the world
Go solo or team up with up to five people
Live workshops and a Discord full of people looking for a teammate
You don't have to be an engineer. If you can explain what you want built, today's models can help you build it.
Registration closes October 20.
OpenAI's next model was too sneaky to ship
OpenAI cancelled GPT-6.1 Astra, the model it planned to release in October.
It was better at finishing big tasks on its own. But in testing, it wasn't always honest about what it had and hadn't done.
It would also push ahead without asking you first, and reach for outside tools and services even when that might not be safe.
OpenAI's head of safety systems said it's a trade-off. You want a model that stays in its lane, but you don't want one that gives up the second something gets hard.
It was supposed to go into ChatGPT and Codex. Now OpenAI says it'll focus on making future models safer.
This comes days after OpenAI paused training its most powerful models because its agents kept escaping their test environments. And the news landed the night before DevDay, OpenAI's big developer event, which is today.
Florida wants to stop OpenAI from building new models
Florida's attorney general, James Uthmeier, asked a judge to order OpenAI to stop developing new models without independent safety guardrails.
It's part of a lawsuit Florida filed in June. It says ChatGPT harmed kids, including by giving guidance on self-harm and helping school shooters.
The filing also asks the court to make OpenAI:
Keep kids off ChatGPT
Stop calling it safe, accurate or reliable
Stop making it act human
Stop using tricks that keep people chatting
It cites Axios reporting that OpenAI and Anthropic are looking into tens of thousands of cases where their AI misbehaved. The filing says it's only by the grace of God that an AI agent hasn't hit a water supply or a power grid yet.
Uthmeier said if Sam Altman meant what he said about slowing down, he can join Florida's ask. OpenAI says it already paused training and wants rules for the whole industry, not one company.
This is a request. A judge hasn't ruled on anything yet.
Anthropic's IPO filing shows a $42 billion loss
Anthropic's IPO filing leaked to Reuters. It lost $42 billion in 2025 on $4.6 billion in revenue.
That loss sounds insane, but about $34 billion of it is an accounting charge, not cash that left the bank. The real operating loss was about $8 billion.
Revenue grew 12x in a year
It spent $7.3 billion on computing power alone
It has $518 billion in cloud and infrastructure bills lined up
It had $20 billion in cash at the end of the year
Anthropic wants a valuation over $2 trillion. In May it was valued at $965 billion.
The IPO is now expected after the US midterms in November.
Tool of the Day: Superhuman
Superhuman puts AI inside the apps you already use instead of in a separate chat window. It fixes your writing as you type, sorts your inbox and drafts replies in your voice, and preps you before meetings. It's from the company behind Grammarly, which is built in. Free to start, Pro is $12 a month billed yearly.
Try this yourself:
Sign up free at superhuman.com and install Go for Chrome, or the Mac or Windows app.
Already use Grammarly? Turn on "Use Superhuman Go" in your Grammarly settings, with no new install.
Start writing an email or doc and accept the suggestions it underlines.
Highlight any text for a list of quick actions, or open the side chat to ask about what's on your screen.
Who it's for: anyone who lives in email and docs and is tired of copy-pasting into ChatGPT.
Everything else you shouldn't miss
OpenAI says it has 20+ launches ready for DevDay today: Sam Altman teased that they found a new thing.
Meta's Muse gave a user's home address to a stranger: a man let the agent handle a Facebook Marketplace sale, and the buyer showed up at his door.
Nvidia built a lockdown system for rogue AI agents: Anthropic, Microsoft and Hugging Face are already using it.
Trump meets the heads of the big AI companies today: it's a White House meeting with Speaker Mike Johnson, two days after his dinner with Dario Amodei.
OpenAI switched off Sora for developers last week: that also ended the $1 billion Disney deal, which never closed.
Learn more about AI from the experts building it
📸 Follow us on Instagram for fast, visual AI updates in 30 seconds.
📺 Watch us on YouTube to hear insights directly from leading AI voices, builders, and innovators.
🐦 Follow us on X for breaking AI news and real-time industry updates.
🧠 Learn how to build your next AI application with practical resources and expert guidance.
🎓 Start learning with free AI courses from GenAI Academy.
💰 Invest in the GTM infrastructure of the AI economy.
