Finn Brooks: Dude, I woke up Monday and my feed was just — Sonnet 5, Sonnet 5, Sonnet 5. And I had this very specific reaction which was like, wait, didn't Sonnet 4.6 just come out?
Clara Bennett: February 17th. Four months ago.
Finn Brooks: Four months! And it's already the old default. Sonnet 5 launched June 30th and immediately it's what free-tier users get. Pro, Max, Team, Enterprise — all of them. Plus GitHub Copilot rolled out generally available, plus Amazon Bedrock for enterprise workloads. That is a lot of surface area on day one.
Clara Bennett: Mm, and the performance claim is the part that actually deserves scrutiny. Anthropic is saying Sonnet 5 approaches or surpasses Opus 4.8 on knowledge work benchmarks — and Opus 4.8 sits at five dollars per million input tokens.
Finn Brooks: Which means — hang on — if that claim holds, why does Opus exist as a product anymore?
Clara Bennett: That's exactly where I want to go. Because Anthropic made a deliberate choice here — and I don't think it's accidental that the answer is complicated.
Clara Bennett: Think of it like a contractor. Same blueprints, same renovation outcome — half the invoice. That's the claim. And the thing that's actually new isn't the cheaper price, it's what the cheaper price now buys you. Planning, browser use, terminal access, autonomous multi-step execution. That capability used to live at Opus.
Finn Brooks: Wait, like — those aren't just features they unlocked, they physically moved them down a tier?
Clara Bennett: That's the claim Anthropic made on June 30th, yes. Now — here's where I pump the brakes a little, because OpenAI said almost exactly this about GPT-5. Google said almost exactly this in the same window. Rebecca Bellan at TechCrunch flagged it explicitly — the agentic framing is convergent. It's the same language in slightly different fonts across three companies simultaneously.
Finn Brooks: So is it a real capability jump or are we just — yeah, okay I love that framing actually, same press release different fonts.
Clara Bennett: The marketing language isn't new. What is concrete — what I think survives the skepticism — is the financial decision it forces. A developer running, say, a hundred-million-token job is now choosing between paying roughly five hundred dollars for Opus 4.8 or significantly less for Sonnet 5. That's not a benchmark. That's a real line item someone has to sign off on.
Finn Brooks: And that decision didn't exist four months ago because Sonnet 4.6 wasn't even in that conversation performance-wise.
Clara Bennett: And that's actually the take I want to push back on — the one circulating right now. The story people are accepting is that Anthropic made a principled safety decision. They pulled back Sonnet 5's cybersecurity capabilities, they kept Mythos and Fable restricted, responsible adults in the room. That's the press release. And Madison Mills at Axios flagged the tension, but I don't think it's gotten enough traction.
Finn Brooks: Wait, what's the tension exactly?
Clara Bennett: iOS 26.5.2. Apple shipped it June 29th — one day before Sonnet 5 launched — patching a vulnerability that was found with Claude AI's help. So the model they're describing as having 'much lower ability' on dangerous cyber activities... just helped close a real security hole in Apple's operating system.
Finn Brooks: Okay no — wait. Does that complicate the safety framing or does it actually prove it? Like, finding a vulnerability to patch it is — I mean, that's defensive. That's the good version of the thing.
Clara Bennett: That's the generous read. And it might be right! But — here's where I get stuck — Anthropic's disclosure says Sonnet 5 was not deliberately trained on cybersecurity tasks. So either the capability emerged anyway, or 'much lower ability' is doing a lot of work in that sentence.
Finn Brooks: Oh that's — yeah, that's the crack in it.
Clara Bennett: The real question is whether practitioners — people actually stress-testing Sonnet 5 in production — will read Anthropic's risk tiers as honest design or as a mechanism that keeps Mythos and Fable scarce while pushing Sonnet volume. Those are different things. One is safety architecture, the other is a commercial pressure valve.
Finn Brooks: And that's the thing I keep — okay, I can't figure out where to land on this. Because Sonnet 5 is now the default for free-tier users. Free. The people least likely to ever spin up a multi-step autonomous workflow. Like, the person who gets Sonnet 5 as their default is probably asking it to rewrite a cover letter.
Clara Bennett: And the practitioners who would actually know — the developers running production agentic workloads — they're still on Opus 4.8. Still evaluating. We don't have independent verification yet. What we have is Anthropic's benchmarks and, as I said, three companies using identical framing in the same quarter.
Finn Brooks: Which means if Sonnet 5 actually is what they say it is, Anthropic just collapsed the argument for their own premium tier. Deliberately. And if it isn't — we're just in another benchmark cycle and nobody's equipped to call it yet.
Clara Bennett: Yeah. That's genuinely the open question. I don't have an answer.
Finn Brooks: Me neither. Which is kind of a weird place to end up — but I think it's the honest one. Thanks for thinking through it with me.
