Claude Sonnet 5 Got Cheaper — So Why Does It Cost More?

Transcript

Sam: Here's the trick that should annoy you. A company ships a new AI model, tells you it's cheaper — and your bill goes up.

Alex: Cheaper per token. More expensive per job. And there's exactly one number that tells you which one is happening to you. That's today. Welcome back to Dan's AI Intel — the show that tries to make sense of the fastest, strangest shift most of us will ever live through, one week at a time. I'm Alex, and as always, I'm here with Sam.

Sam: Hey. And Alex, I want to stay on that opening for a second, because it sounds almost like a conspiracy theory. "They made it cheaper, but you pay more." That can't just be — true.

Alex: It's completely true. And by the end of this, you'll see exactly how it's done — and honestly, once you see it, you can't unsee it on any model launch ever again. Which is really what this whole show is for. As Dan puts it — the AI world moves so fast that genuinely keeping pace feels impossible, so he takes the rough curiosity he can't shake and turns it into something durable you can actually hold onto. And this week, the thing he couldn't shake was a receipt.

Sam: So set it up. What actually happened?

Alex: On June 30th, Anthropic shipped Claude Sonnet 5 — their mid-tier model, the workhorse most people actually reach for day to day. Lower sticker price, and performance that suddenly crowds their own flagship. On paper, a clean win.

Sam: And the thing that hooked you was...?

Alex: The receipt underneath that win. Because when independent people actually measured what a real job costs on this thing, the story flipped. But the reason I couldn't put it down isn't really Sonnet 5 at all — it's the question it forces open: in 2026, what does the price of an AI model even mean anymore? We'll walk the whole thing — the genuinely impressive upgrade, the quiet change to the meter you're billed on, the rival across town that's actually winning the cheap race, and how you should really buy. And there's one move almost everyone makes when a "cheaper" model drops that turns out to be exactly backwards.

Sam: Okay — "exactly backwards" — and you're not going to tell me what it is yet, are you.

Alex: Not yet.

Sam: Of course not. Before we dig in — one quick personal note, and this one's straight from Dan.

Alex: Yeah. He makes this show himself, and the only reason it grows is people who like it pass it on. So if that's been you — genuinely, thank you. Right. Let's get into it. So start with the good news, because it's real and it's substantial. Sonnet 5 is the most capable mid-tier model Anthropic has ever shipped, and on a whole set of axes it gets shockingly close to Opus 4.8 — that's the top-of-the-line flagship, the expensive one.

Sam: Give me "shockingly close" in numbers, because a mid-tier model catching the flagship is usually just marketing.

Alex: Fair. On general knowledge work — the reasoning and writing most of us actually use these things for — there's a benchmark in the GDPval family, and Sonnet 5 scores 1618, to Opus's 1615.

Sam: Wait — it beat the flagship?

Alex: It's a dead heat. Three points is statistical noise. But yes — on that measure, the mid-tier model is level with the premium one. On autonomous computer use — literally driving a computer, an eval called OSWorld — Sonnet 5 hits 81.2%, Opus is 83.4%. On long terminal tasks, Sonnet 5's around 80%, Opus in the low eighties. Even on the nastiest coding benchmark, SWE-bench Pro, it's 63.2% to Opus's 69.2%.

Sam: So it loses on the really hard coding, but everywhere else it's basically breathing down the flagship's neck. And think about what that does to the buying decision — if the cheaper model ties the flagship on the everyday reasoning, the flagship's only real pitch left is "but I'm better on the genuinely hard stuff." That's a much narrower reason to pay up.

Alex: That's the whole strategic tension in one sentence, and hold that thought, because it comes back to bite in a way you will not see coming. Right — and here's what makes it more than a spec bump — it's the people who've used it. If your honest experience of older Sonnet models was "capable, but frustrating" — hesitant, a bit shallow when things got hard, not quite trustworthy on a real task —

Sam: — yes, that's exactly the complaint. It'd start strong and then kind of lose the thread.

Alex: This is the release most likely to change your mind. The floor moved up. The catch — and there's always a catch — is what it costs to stand on that floor.

Sam: Before we get to the catch — you said it ties the flagship on some things. How much better is it than the Sonnet it's actually replacing? Because that's the upgrade most people will feel.

Alex: Honestly, a bigger jump. The old one, Sonnet 4.6, came out in February. On that hard coding test it went from 58.1% to 63.2%. On computer use, 78.5 up to 81.2. On terminal tasks, 67 up to about 80.

Sam: Those are real steps, not rounding.

Alex: And the one I'd point at is a brutal reasoning exam — it's literally called Humanity's Last Exam, designed to be nearly impossible. Sonnet 4.6 got 34.6% on its own. Sonnet 5 gets 43.2 — but hand it tools, let it search and run code and check its own work, and it jumps to 57.4%.

Sam: Okay, so the "with tools" number is way higher than the raw one. Why do you keep flagging that?

Alex: Because that's the whole design thesis in a single data point. This model isn't really built to sit there and answer you in one shot. It's built to be dropped inside an agent — something that plans, calls tools, reads the result, and decides the next move on its own. How much it jumps the moment you give it tools — that's the entire point of the thing.

Sam: So they built it to do stuff, not just to say stuff.

Alex: Right — and hold onto that phrase, "built to do stuff," because it's about to explain the entire cost mystery.

Sam: So we've banked the good news — near-flagship quality, a real jump over the last one. But before the catch, let me zoom out, because a nagging voice is asking why we're spending a whole episode on the mid-tier model. That's usually the boring one.

Alex: That's exactly why it's the story. The frontier race has quietly split into two races. One is the old one — who's got the single biggest, smartest model on Earth. The other, and it's the louder one in 2026, is the cost race: who can hand you near-frontier quality at a price you can afford to run millions of times a day, inside an agent. Sonnet 5 is Anthropic's entire move in that second race — so the way they priced it tells you how the whole industry works now.

Sam: And the reason the pricing is the interesting part is — we've all been trained wrong on it, haven't we. Cloud computing taught us the advertised unit price is the price.

Alex: And with these models, that instinct completely fails. Two numbers move independently: what you pay per token, and how many tokens a task actually eats. Hold the first one flat — even cut it — and let the second one climb, and your bill goes up while the price list looks generous. That's the whole game. Sonnet 5 is just the cleanest case study we've ever had of it.

Sam: Right. So — now the catch. Hit me.

Alex: Here's the sleight of hand, and it's worth watching slowly, because it's the template for the whole industry now. Sonnet 5's headline price: two dollars per million input tokens, ten per million output — through August 31st. After that, three and fifteen.

Sam: And three and fifteen is...?

Alex: Exactly what the old Sonnet 4.6 cost. Same price list, much better model. Sounds like a straight win.

Sam: I feel a "but" coming.

Alex: The number's measuring a different thing than it used to. So — a token is the unit you get billed in; text gets chopped into these little pieces called tokens. Sonnet 5 ships with a brand-new tokenizer — a new way of doing that chopping — and by Anthropic's own documentation, the same text now counts as anywhere from one, to one-point-three-five times as many tokens as before.

Sam: Hang on. So the price per token didn't move, but each token is now a smaller bite — so the same paragraph is more tokens — so I pay more for the identical text.

Alex: You just described the entire trick. They raised the real price without touching the price list, because they changed what a token is, not what a token costs. Picture a shop that keeps the price per slice the same, but quietly cuts the loaf into thinner slices. Same sign in the window. More slices per sandwich. Bigger bill.

Sam: That is infuriatingly clever.

Alex: And it's not a one-off. When they rolled a new tokenizer into the Opus line earlier, independent developers measured token counts climbing 1.325 to 1.47 times. And one look at 483 real requests found tokens per request jumped 37.4% — for identical work.

Sam: Wait — 37.4% on real requests. So this isn't a theoretical "up to 1.35 times" in the docs. Somebody sent the same prompts before and after, and it billed a third more.

Alex: For the same work, yeah. That's the part I want people to internalize — the meter reading changed underneath a number that didn't. And nobody has to lie for it to happen. The price per token on the page is completely honest. It just isn't the price of anything you care about.

Sam: Okay, so that's one lever — the ruler got shorter. But you said "the second, larger effect." There's more?

Alex: There's a bigger one. The model does more per task. Remember "built to do stuff"? Sonnet 5 is trained to be more agentic — to run those autonomous loops. And one analysis found that at maximum effort, it burns roughly 40% more output tokens per task than Sonnet 4.6 — and runs about three times as many of those agent loops.

Sam: Three times the loops. And every loop is more tokens I'm paying for.

Alex: Every single loop. More loops means more capability — it's checking its own work, trying again, driving the browser one more step. But more loops also means more billed tokens, on every loop.

Sam: So stack it up. Each token counts for more, and it's running way more tokens. Those don't add — they multiply.

Alex: They compound. And notice these two levers are doing very different jobs. The tokenizer one is invisible — it's a change to the accounting, you'd never feel it in the product. The loops one is the opposite: you feel it as the model being better, more independent, more capable. One's a bookkeeping change, one's a genuine upgrade — but they land on the exact same line of your bill.

Sam: And that's what makes it so slippery to argue about. Half of it you'd happily pay for. And this is a quick aside worth flagging — because if this "you're really paying for tokens, and tokens hide the true cost" idea grabs you, we did a whole episode on it: "AI's Hidden 70x Subsidy," number 20, just last week. That's the deeper economics sitting under all of this. But for today, hold the compounding — shorter ruler, times harder-working model. Let's see what that does to an actual bill. Right — so we've got the shorter ruler and the harder-working model, compounding. This is the bit I've been waiting for: forget the mechanism for a second, what does all that actually do to what I pay?

Alex: So — real numbers. That same analysis estimated the all-in cost to finish one representative task. On the old Sonnet 4.6: about a dollar-twenty. On Sonnet 5: about two dollars twenty-nine.

Sam: Nearly double. For the model they're selling as the cheaper option.

Alex: Now here's the line I want you to actually sit with. On that same task, they put Opus 4.8 — the premium flagship, the expensive one — at about a dollar ninety-seven.

Sam: ...Say that again. The cheap mid-tier model costs more to finish the job than the premium flagship it's supposed to undercut?

Alex: On a genuinely hard job, at high effort — yes. The thing you'd reach for to save money can cost you more than the thing you were trying to avoid paying for.

Sam: That's the "oh, come on" moment. Is that solid? Or is that one angry developer with a spreadsheet?

Alex: Good instinct — and here's the honest boundary. The exact dollars are one firm's measurement, not a number Anthropic published. So treat two-twenty-nine and one-ninety-seven as estimates, not gospel. But the machinery underneath — the tokenizer change, the loop counts — that part is documented, and corroborated. So: the precise figure, wobbly. The direction, not in doubt.

Sam: So I shouldn't quote the exact cents, but the shape is real — cheaper sticker, higher bill, and sometimes higher than the flagship.

Alex: That's exactly the right way to hold it.

Sam: And actually, the fact that it's an estimate is kind of the point, isn't it. There's no published "cost per task" number from anyone. The whole industry reports price per token — the one number we've just spent ten minutes proving is a decoy.

Alex: And that's the quiet tell in the whole thing. The reason a random firm has to reverse-engineer this with a spreadsheet is that the number that matters isn't on any pricing page. Which is exactly why the only reliable version of it is the one you measure yourself, on your own actual work — your prompts, your effort level, your task. Anybody else's dollar figure is a hint. Your own is the truth.

Sam: So we've got: the bill can beat Opus, and you were careful about the numbers. But I want to push on the framing — is this Anthropic being sneaky? Because it's starting to sound like you're accusing them of a scam.

Alex: This is the part I find genuinely interesting — and no, I don't think it's a trick bolted onto a good model. The capability and the cost are the same thing, seen from two sides.

Sam: Unpack that.

Alex: They haven't published the exact recipe — no frontier lab does. But the shape is known. These are transformer networks, pretrained to predict text, then heavily post-trained with reinforcement learning — from human feedback, from AI feedback, and Anthropic's own Constitutional AI, where you train the model against a written set of principles instead of hand-labeling every case. And layered on top, more and more, is agentic reinforcement learning — training it not just to answer, but to run in a loop, call a tool, read the result, decide the next move.

Sam: Okay, translate the layers for me, because "Constitutional AI" and "reinforcement learning from AI feedback" is a lot of jargon in one breath. What's actually going on there?

Alex: Think of it in three passes. First pass, pretraining: you feed it a huge chunk of text and it just learns to predict what comes next — that's the raw, book-smart layer. Second pass, you teach it manners and judgment with reinforcement learning — humans, and other AIs, rating answers so it learns which ones are actually good. Constitutional AI is a clever twist on that: instead of rating a million cases by hand, you hand it a written set of principles — a constitution — and train it to critique itself against those rules.

Sam: So the first pass is the reading, the second pass is the raising.

Alex: The reading, then the raising — that's it. And then there's a newest pass, the one that actually matters here: agentic reinforcement learning.

Sam: And that agentic training is the engine.

Alex: It's the entire engine of this release. People poking at it noticed Sonnet 5 even seems aware of its own token budget — it fires off parallel tool calls aggressively early in a task, then gets more cautious as it nears its limit. It plans more, it checks more. That's precisely why it can drive a terminal to a finished result with less hand-holding —

Sam: — and precisely why it runs three times the loops and burns the tokens. You literally cannot have the autonomy without the loops.

Alex: And you can't have the loops without the bill. Sit with the incentive for one second: the company earns more exactly when the model works harder — and they trained it to work harder.

Sam: So that's not a bug in the pricing. That's the business model wearing a lab coat. Okay. So if Sonnet 5 isn't really the "capable and genuinely cheap" answer — who is? Because that was the whole promise.

Alex: The honest answer points across town, to Google. They shipped Gemini 3.5 Flash back at their I/O conference on May 19th, and on the two things people actually mean by "cheap" — price and speed — it isn't close.

Sam: Numbers.

Alex: Flash is a dollar-fifty per million input, nine dollars output — and cached input, if you're reusing context, is fifteen cents per million. That undercuts Sonnet 5's standard three-and-fifteen by roughly half, and that cached rate is on another planet. Then, speed. Flash streams around 289 tokens a second. Google markets it as about four times faster than rival frontier models.

Sam: And Sonnet 5's speed?

Alex: Between about 56 tokens a second at low effort, and 78 at maximum.

Sam: So Flash is — what, four or five times faster, and half the price. For anything I'm actually sitting there watching load, that's the whole ballgame.

Alex: That's the difference between a snappy product and a spinner. And both now carry a one-million-token context window, so the old "Claude just holds more" advantage has narrowed there too.

Sam: But surely Flash is dumber. That's the trade, right — you pay for cheap and fast with quality.

Alex: Less than you'd think. Flash posts 76.2% on that terminal-task benchmark, and on an aggregate "intelligence index," it scores 55 — ahead of Sonnet 4.6's 52. A Flash-tier model outscoring a recent Claude Sonnet on overall intelligence.

Sam: Huh. So it's not the budget bin. It's genuinely good, and cheap, and fast. That's the combination that's supposed to be impossible.

Alex: And that cached-input rate is the tell for where this is going. Fifteen cents a million for reused context — that's a price built for something that hammers the same instructions over and over, thousands of times an hour. That's not a price for a person chatting. That's a price for an army of agents.

Sam: Let me make sure I've actually got the shape, and I'm not just cheerleading for Google here. Where does Claude still win?

Alex: Careful, real-world code quality, and bug-fixing on genuinely hard problems — that's where Claude still tends to lead. Gemini leads on raw throughput, on price, and on stringing together lots of tool steps. So it's not "one is better." It's two labs betting on different futures.

Sam: Spell the bets out.

Alex: Anthropic's bet: buyers will pay for capability and autonomy — for a model that just finishes the work — and they'll accept a rising cost per task as the price of that. Google's bet is the opposite: speed and true low cost win the high-volume future, the one where you're running this thing millions of times a day inside an agent.

Sam: And those two bets are almost mirror images of the cost story we just told. Anthropic's whole value is the model does more per task — which is exactly the thing that runs up your bill. Google's whole value is do it cheap and fast a million times — which only works if each call is dirt cheap.

Alex: Right — each lab's strength is the other lab's expense. And it's genuinely not obvious who's right. If the future is a handful of hard, high-stakes jobs, you want the model that finishes them. If it's oceans of small automated steps, you want the one that costs a rounding error each time.

Sam: And this is where it hooks back into the thing you keep circling — "the model you can actually use."

Alex: It does. And, quick aside — we pulled that exact thread apart in "GPT-5.5 versus Claude Mythos," episode 8, back in May: the frontier model you can actually use, versus the one you can't. This is the same fight, one lab later. Because the frontier race quietly split in two. One race is for the single smartest model. The other — louder in 2026 — is for cost: near-frontier quality at a price you can run at scale. Sonnet 5 is Anthropic's move in that second race. And Gemini Flash is running away with it.

Sam: Right — this is the part people actually need. If I'm sitting here tomorrow deciding what to use — what's the rule?

Alex: And to be clear, none of this makes Sonnet 5 a bad buy. It makes it a specific buy. The sweet spot is low and medium effort. Down there, the tokenizer-and-loops tax is small, and you're getting near-Opus quality at a mid-tier rate — quality that old Sonnet pricing simply couldn't buy. For the bulk of everyday work — drafting, summarizing, straightforward code, routine agent steps — Sonnet 5 at low or medium effort is one of the best-value things on the market. And Free and Pro users get it as the default now, so most people will feel the upgrade without ever touching a setting.

Sam: And that "effort" setting — just so everyone's with us — that's a dial you actually turn? Low, medium, high, max?

Alex: It is. Roughly, it's how much thinking and how many of those loops you let it spend before it answers. Low effort, it's quick and cheap and often plenty. Crank it to max, and you've told it: go away, run every loop you can, check everything, spare no tokens.

Sam: So the dial that makes it smartest is literally the dial that makes it most expensive. And the trap is the top of it.

Alex: The trap is exactly the top of the dial. That's where the loops and the token count climb until — as we just saw — you can pay more than Opus, at similar quality. And the developers reacting on launch day went straight at this. One likened it to buying the same box of chocolates every day, while the box quietly shrinks and the price holds.

Sam: That's the thinner-slices thing again, exactly.

Alex: Another said the models feel "optimized for wealth extraction" as much as for solving your problem. And the blunt practical question kept coming up — why pay Sonnet 5's high-effort rate, when Opus on low effort might cost about the same and just do the job? For high-volume or really cost-sensitive work, a lot of them pointed at cheaper Chinese models — DeepSeek, GLM-5.2 — competitive mid-range performance at a fraction of the cost.

Sam: So the rule almost writes itself.

Alex: It does. Use Sonnet 5 at low-to-medium effort as your value workhorse. When a task is genuinely hard, escalate to Opus — rather than max-effort Sonnet. And when speed and raw price dominate, reach for Gemini Flash.

Sam: And that's the "exactly backwards" move from the top, isn't it. The instinct is: hard task, so crank the cheap model to maximum. And that's the single worst-value thing you can do.

Alex: That's the one. Hard task, max out the "cheap" model — and you've quietly overpaid the flagship. Backwards.

Sam: There's one more thing in Sonnet 5's favor you mentioned offhand — it's actually safer?

Alex: Worth saying, yeah. Anthropic reports lower rates of bad behavior than Sonnet 4.6, better resistance to prompt-injection hijacking, less sycophancy, less hallucination — cyber-misuse safeguards on by default, and a deliberately capped ability on offensive cyber tasks. For an agent that's touching real systems, that hardening is worth real money too — it just doesn't show up on the token meter.

Sam: Which is kind of the theme of the entire episode. The stuff that matters isn't on the meter.

Alex: That's the durable takeaway — and it's the one thing I want everyone to carry out of here. In 2026, the token rate is a decoy. The number that actually decides your bill is cost per completed task — and by that measure, the industry's "cheaper" models are quietly getting more expensive, even while the price-per-token headlines keep falling.

Sam: So — stop comparing price per token.

Alex: Stop comparing price per token. Start measuring cost per finished job — ideally on your own workload — because that's the only number that survives a new tokenizer or a jump in agent loops.

Sam: And that's more doable than it sounds, right? You don't need a firm with a spreadsheet. Take a task you actually run — a real one — run it on each model at the effort you'd actually use, and look at the total tokens billed to finish it, not the rate. That single number quietly folds in the tokenizer and the loops and everything else.

Alex: And the rate lies to you; the receipt doesn't. The beautiful part is it's future-proof — next launch, next tokenizer, next round of "it's cheaper now," you just run your own task again and read the total. You never have to trust the headline.

Sam: Measure the job, not the token. I think that's the whole show on a sticky note. Okay, let me try to land the plane. Three things I'm walking away with. One: Sonnet 5 is genuinely a big upgrade — near-flagship quality, a real jump over the last Sonnet.

Alex: True.

Sam: Two: it can cost more to finish a job than the last Sonnet — sometimes more than the premium Opus — because they changed the meter with a new tokenizer, and trained it to run way more loops. And that's not a scam, it's the business model: they get paid on tokens, and they trained it to spend more tokens.

Alex: Exactly.

Sam: And three — the buying rule. Cheap model at low-to-medium effort for the everyday stuff, Opus when it's genuinely hard, Gemini Flash when speed and price rule the day. And whatever you do — stop watching the per-token price, and start watching the cost per job.

Alex: That's the episode. And if you want the tell for whether all this is heading somewhere honest — watch three hinges. September 1st, when Sonnet 5's intro pricing ends, and the real three-and-fifteen finally meets that new tokenizer in live bills. Whether Anthropic ever starts reporting cost per task, instead of just per token. And whether the cheap-and-fast tier keeps eating the work that mid-tier models used to own.

Sam: And if it does — "capable and cheaper" stays a real option. Just not always the one wearing the Claude badge.

Alex: And that's it for today — genuinely, thank you for spending the time with us. I hope you're walking away seeing this a little more clearly, because the real story here isn't one model. It's that a whole industry has quietly learned to move the cost somewhere the price list can't see it — and that's the kind of shift that's easy to miss, and expensive to miss. That's exactly what makes this moment worth following as closely as we can.

Sam: And a quick note, for full transparency — straight from Dan: "I'm Dan. AI moves too fast to keep up with, so I built my own stack of AI tools to research, analyse, verify and illustrate the questions I can't stop thinking about — mostly to learn it myself, and I share what I find. AI-assisted, fact-checked, worth a second look."

Alex: Before we go, one genuinely useful thing you can do: follow the show. Whatever app you're listening in right now, there's a follow, or a plus button — it's one tap, it's free, and it does two things. You'll get every new episode the moment it lands, and honestly, for a small independent show like this one, a follow is the single biggest lever there is for helping it reach other people who are trying to make sense of all this. So if this was worth your time — go ahead and hit follow.

Sam: And one last thing, and this one's from Dan too: "I make this mostly to keep up with AI myself, and I publish it in case it helps you. I'd genuinely love to make it better — so if there's something you'd push back on, or a topic you want me to go deeper on next, tell me: podcast at connectiveshift dot com. I read every single one — and honestly, it's what decides what I dig into next."

Alex: It really does. Thanks for listening, everyone — we'll see you next time.