Clara Bennett: You texted me just 'June 30th' with three dots. I assumed that meant you'd seen it.
Finn Brooks: Dude, the three dots were me trying to figure out how to say — like, Anthropic just made Claude Sonnet 5 the default for every single user tier simultaneously, priced it at ten dollars per million output tokens through August 31st, and announced it on the literal last day of Q2 before what everybody is calling a blockbuster IPO. The dots were because I didn't know where to start.
Clara Bennett: The IPO framing is where I'd start. Michael Nuñez at VentureBeat used 'blockbuster' — that's a specific bet that private-market AI valuations survive public scrutiny. A splashy launch the morning Q2 closes is not incidental to that.
Finn Brooks: No, totally — and then it gets weirder when you realize this replaces Sonnet 4.6, not Opus 4.8. They're not replacing the flagship, they're replacing the everyday model, and claiming it nearly matches the flagship. That framing is doing a lot.
Clara Bennett: That framing is the whole episode.
Finn Brooks: Yeah. Yeah okay — let's go.
Clara Bennett: Let me give you the plain version of what Anthropic actually did with the pricing. Imagine a grocery store switches to a smaller measuring cup and charges you by the cup. The price per cup drops. But you need more cups to fill the same bag. That's the tokenizer story. Anthropic changed how Sonnet 5 counts tokens — now every piece of text inflates by somewhere between one-point-zero and one-point-three-five times what Sonnet 4.6 would have counted.
Finn Brooks: Wait — so the denominator got smaller.
Clara Bennett: Exactly. And Anthropic says that inflation is, quote, accounted for in introductory pricing. Which means — and this is the part worth sitting with — the offset goes away August 31st. The prices jump fifty percent, to three dollars input and fifteen dollars output, and the smaller cup does not come back.
Finn Brooks: Okay hang on, because the early access partners also said Sonnet 5 finishes tasks where previous Sonnet models stopped short — like it self-checks without being asked. That means more tokens per run than a comparable Sonnet 4.6 job anyway, right? So you've got inflation from the tokenizer AND inflation from the model just doing more.
Clara Bennett: That's — yeah, that's the compounding part. In practice a developer building a daily finance agent, ingesting contracts, flagging payment terms — they switch to Sonnet 5 today thinking they're saving fifty percent. But the task runs longer because the model self-checks. The token count is already inflated by up to one-point-three-five. And then September arrives.
Finn Brooks: Their bill could actually be higher than if they'd stayed on Sonnet 4.6. Like, genuinely higher. Not just 'less savings than expected' — higher.
Clara Bennett: That's the realistic scenario. The headline discount is real — now, through August. After that, the math depends on your content type, your task complexity, and whether Anthropic quietly adjusts anything. None of which they've committed to.
Finn Brooks: Okay but here's what's bugging me — every tech outlet right now is writing 'Anthropic just democratized agentic AI.' That framing does not survive like thirty seconds of competitive context.
Clara Bennett: Name the context.
Finn Brooks: GPT-5.6 Sol. OpenAI launched it in preview the week before Sonnet 5, and the pitch was — autonomous subagents, split work across tasks, most agentic model yet. Word. For. Word. And then Google had Gemini 3.5 Flash in May with nearly identical mid-tier agentic positioning. So Anthropic drops Sonnet 5 and everyone acts like they invented a category that two other companies were already standing in.
Clara Bennett: That's fair — but the FrontierCode number is real. More than double Sonnet 4.6's score. That's not marketing language, that's the largest single capability gap between two consecutive Sonnet models. Does that not count?
Finn Brooks: It counts, I just — wait, actually, agentic benchmarks are still immature. We don't have agreed standards for what autonomous multi-step work even means. So doubling your score on FrontierCode is real, but real on what scale? OpenAI and Google are citing their own numbers on their own evals.
Clara Bennett: So the differentiator might not be the benchmark at all. Snowflake had Sonnet 5 available the same day — on Snowflake Cortex AI, inside the Snowflake security perimeter. That's locked-in distribution before the press cycle even starts.
Finn Brooks: That doesn't happen by accident.
Clara Bennett: Anthropic launched Sonnet 5 on June 30th, the last day of Q2, ahead of an IPO that VentureBeat is calling a blockbuster test of whether private-market AI valuations hold up in public markets. And they set pricing that, I mean, expires August 31st. So the adoption story they get to report going into that IPO is built on a price that doesn't exist by the time anyone scrutinizes the S-1. I'm not sure what to do with that.
Finn Brooks: And we genuinely don't know if that $3 and $15 pricing holds after August, or if competitive pressure from OpenAI and Google just... forces Anthropic to extend it, or revise it entirely. Like that answer doesn't exist yet.
Clara Bennett: No. It doesn't.
Finn Brooks: Good talk. Genuinely unsettled, but good talk.
