Welcome back! Anthropic and OpenAI both dropped new models yesterday, 90 minutes apart. Anthropic's is as good as its best model for less than half the price, and OpenAI's are half the price they were. 20 countries signed a letter asking to keep AI under human control. The US, China and the UK didn't. And people stopped visiting news sites because AI already told them what's in it.
In today's Generative AI Newsletter:
Anthropic: Is Opus 5.5 really as good as Fable?
OpenAI: What did it ship 90 minutes later?
UN: Why didn't the US, China and UK sign?
News: Where did a third of the readers go?
Claude Opus 5.5 is as good as Fable for less than half the price
Anthropic released Opus 5.5 and says it performs at Fable 5.1's level on most work.
It costs $4 in and $20 out per million tokens. Fable is $10 and $50.
Artificial Analysis, the independent benchmark site, gave it 58. That's the highest score it has ever measured. Fable 5.1 and GPT-6 Astra sit at 53.
Beats Fable on every test in Anthropic's own table
Terminal-Bench, the long coding task test, 66% vs Fable's 56%
Anthropic says it's 30% faster than Opus 5 and 40% cheaper to run
Said to be way better at writing and communicating
PRO users won’t feel as left out since this model is Fable-level
OpenAI launched new models 90 minutes later
Anthropic itself says the real gap to Fable is smaller than the numbers suggest, and it still points you to Fable for the hardest work.
If you pay for Claude, it's already your default model. Limits went up 20% and there's a one-time reset waiting in Settings until October 22.
This is the first model Anthropic shipped since its CEO asked the whole industry to slow down. Looks like the slowdown is just talk.
Free AI courses
GenAI Academy put its courses in one place for free.
Start with "What Can AI Actually Do For You?" if you're new.
30 minutes, no setup, no account to configure.
The Claude Starter Course gets you running in 35 minutes.
After that, pick by the problem you have:
Burning through usage limits has its own course
Building no-code agents
Rolling AI out across a team in 30 days, from pilot to adoption data to full deployment
One idea shouldn't take six rewrites to post.
Posting everywhere means rewriting one idea six times, so you post to one, or none. SureThing turns one idea into native posts for every platform.
OpenAI answered 90 minutes later with a price cut
OpenAI shipped GPT-6 Sol and Luna an hour and a half after Anthropic's post. Sol is for coding and agents. Luna is for cheap high-volume jobs like summarizing documents.
Both cost half of what the 5.6 versions did. Sol is $2 in and $10 out per million tokens. Luna is $0.10 and $0.50.
Artificial Analysis gave Sol a 48. Same score as the model it replaces. Opus 5.5 got 58.
So Anthropic went for the score and OpenAI went for the price. Altman says price per task is what should matter and nothing on the market comes close.
In ChatGPT Work and Codex for paid plans, not in regular chat yet
Free users get Luna in the desktop app
Paid users got a banked usage reset, same as Claude did the same day
Nobody has run Sol against Opus 5.5 yet. OpenAI's own comparisons were against Opus 5, which was already old by the time the post went up.
DevDay is September 29, and Altman says they still have too much to launch by then. We’ll keep you updated.
The US, China and UK skipped the AI control pact
20 countries and the EU signed a call to keep AI under human control at the UN this week. Finland and Norway wrote it. Canada, Germany, Australia and the UAE signed.
The three that didn't are the US, China and the UK.
It asks for three things.
Companies test models before release and let outside evaluators in
Governments share reports when something goes wrong
The UN looks into building a watchdog for AI
None of it is binding. Nobody has to do anything.
The next day Trump told the UN the US rejects any globalist scheme to control AI, and that government documents will now say "super intelligence" instead of "artificial intelligence."
The two countries that build the models you use are the two that stayed out. We'll see who signs next.
People stopped clicking on news sites
Traffic to the biggest US news sites is down about 30% from its peak in late 2024, according to Similarweb data reported by Axios. AI answers replaced the click.
A separate count of the top 50 sites found:
CNN's traffic from Google search fell 61% in two years
USA Today went from 203 million visits a month to 90 million
Search traffic across all 50 is down 42%
Google rolled out AI Overviews to everyone in May 2024, and the slide started within a few months. Nobody can prove that's the cause, but the timing is hard to ignore.
The AI companies aren't sending readers back either. Links from ChatGPT and the rest make up 0.4% of news site traffic.
People click a result 8% of the time when there's an AI summary on the page, and 15% when there isn't.
Not everyone is losing. Substack is up 179% in three years for some reason. Why do you think that’s the case?
Tool of the Day: Pexo
Pexo makes a whole video from a chat message. Describe it or paste a URL, and it writes the script, picks a different AI model for each shot and sends back a finished video with voiceover and captions. Free to start, no card, Google sign-in. Paid plans start at $30 a month.
Try this yourself:
Sign in at pexo.ai with Google and pick "Create a launch video."
Paste a product page URL and add "keep it under 30 seconds."
Check the script and shot plan it shows before rendering, then give notes in chat.
Download it, or tell it to try a different tone.
Who it's for: anyone who needs a launch or explainer video this week and doesn't want to learn five video tools.
Everything else you shouldn't miss
Six big banks say AI shopping agents will bring more scams: Bank of America and five others published a joint paper warning agents could push whatever pays them the highest commission instead of what's best for you.
Discord starts sorting every user into teen or adult today: a model guesses your age from your account history, over 90% of users won't have to prove anything, and adults it gets wrong can appeal with a credit card instead of an ID.
OpenAI's human graders reportedly got fired for using AI to grade the AI: contractors rating ChatGPT answers were caught using chatbots, and one admitted picking the worst answers on purpose.
China is reportedly investigating DeepSeek and Moonshot: The Information says regulators questioned executives about the millions of user prompts Anthropic said both companies routed through Claude.
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.
