Frontier AI is still getting better. It's just getting increasingly expensive to prove it.
At the end of June, the developer Theo Browne posted a bill and a skull emoji. The bill was Artificial Analysis's own ledger — they run every frontier model through the full evaluation battery behind their Intelligence Index and publish what it cost them to do it. Running Claude Fable 5, Anthropic's flagship priced at $10 per million input tokens and $50 per million out, had cost $5,631. Running Claude Sonnet 5, the mid-tier model priced at a fifth of those rates: $6,015.
The cheap model sent the bigger invoice.
And Sonnet 5 really is cheap. Its per-token rate is a third lower than the model it replaced — $2 in and $10 out against Sonnet 4.6's $3 and $15, a launch discount that was scheduled to lapse in September and instead quietly became the permanent price. Anthropic cut the price, and the bills went up anyway, because the per-token price of a frontier model has stopped telling you what the model costs to use. The models have started spending tokens the way contractors spend billable hours.
The Price Fell. The Bill Rose.
If you've pointed one of this year's models at a real task and watched it work, you've seen the shape of the thing. The model doesn't answer any more — it deliberates. It reads the file, writes a plan, runs the test, reads the output, revises the plan, runs the test again. Every one of those steps is tokens, and every token is billed — even the ones you don't see.
Artificial Analysis measured it. Sonnet 5 scores six points higher than Sonnet 4.6 on their Intelligence Index. To get those six points it generates roughly 40% more output tokens per task, and on their knowledge-work evaluations it takes about three times as many agentic turns — three times the round trips of tool calls, file reads and self-checks before it commits to an answer. At like-for-like token prices that works out to about double the cost per benchmark task, more than they measured for Opus, the model tier above it. Even at the discounted price Sonnet 5 actually charges, it still costs about a third more per task than its predecessor did. A 33% price cut, followed by a 33% rise in the cost of using the thing.
Underneath this sits a detail almost nobody prices in: Anthropic's newer models use a new tokenizer that produces roughly 30% more tokens for the same text. The token itself shrank. Cheaper per token, more tokens per sentence — the unit on the price tag and the unit of work have come apart, and the price tag is the one that gets screenshotted.
Which is the real story here, and it's bigger than one model: the per-token price of a frontier model now matters less than the amount of inference the model needs in order to look smarter.
Everyone Is Doing This
Anthropic's own flagship followed the same curve. Claude Fable 5.1 took the top of Artificial Analysis's index at launch, four points above Fable 5 — and costs about 20% more per benchmark task, because it generates 1.7 times the output tokens. It would have been worse: Anthropic cut cache-read pricing by 75% in the same release, which Artificial Analysis reckon saves about $1.40 per task. A discount, in other words, on the model's new habit of re-reading everything it has ever said.
Then Grok 4.7 arrived this week and ran the experiment a third time. xAI held per-token prices exactly where Grok 4.6's were. The model responded by using 81,000 output tokens per benchmark task where its predecessor used 36,000 — 125% more. A hundred and twenty-five per cent more tokens at the same price per token is a price rise; it's just one that never appears on the pricing page.
Three releases, two labs, one shape: modest headline gains, enormous growth in the inference burned to get them. The old scaling story was a better engine every year. The new one is the same class of engine, driven much, much further — think longer, branch more, call more tools, take more turns, check the work, redo the part that failed. The capability arrives less in the engine and more in the mileage.
The Sticker Fails in Both Directions
OpenAI makes the same point from the other side of the shop. This one isn't a successor story — GPT-6 Astra and GPT-5.6 Sol are different models, sold side by side — and that's exactly why it's useful: the sticker fails even when all you're doing is choosing between two products on the same shelf. Astra is priced at $10 in and $50 out per million, 2.5 times Sol's rates on both lines — though Sol's rate is itself a promotion with an end date; in this market even the sticker prices are temporary. Developers did the reasonable thing and winced. Then The New Stack ran a real task through both: Astra finished the implementation in 80 requests where Sol needed 238, read 11.1 million input tokens where Sol read 37.8 million, and came out at $25.67 to Sol's $31.79, in two-thirds of the wall-clock time. The model that costs two and a half times as much per token is the cheaper employee.
OpenAI's own argument is that cost per successful task is now the number that matters, and they're right, which is the uncomfortable part. Within a model line, a 33% price cut can make the bills bigger. Across the shelf, a 2.5× price premium can make them smaller. In neither direction does the per-token price tell you what you'll pay — it has stopped being a fact about the model and become a pricing strategy draped over an unknown quantity of thinking.
What the Money Actually Buys
Some of this spend is pure waste. Artificial Analysis describe Grok 4.7 as "notably slow and very verbose," the way you'd describe a colleague. But a lot of the spend buys things a composite benchmark score is bad at showing. Sonnet 5's six index points came packaged with nine points on Terminal-Bench, ten on Humanity's Last Exam, seven on SciCode — and, more to the point, with the capacity to run for hours as an agent, notice its own mistakes and recover from them. I spend most days with these models running as agents in my own tooling, and the difference between a model that checks its work and one that doesn't is not five points. It's whether the job is finished when I come back to the room. The meter is also real. Both of these are true at once.
What's actually changed is the unit of progress. We used to talk about the capability of a model — one forward pass, one answer, one price. Frontier performance is now a property of model × reasoning budget × harness × trajectory length. "Effort" is a dial on the side of the machine — Sonnet 5 at maximum effort takes about six times the agentic turns it takes on low — and every notch on the dial has an invoice attached. The model is no longer the product. The model plus an undisclosed quantity of its own deliberation is the product, and only the model has a price sticker.
Why the Meter Runs
Three explanations circulate, and they stack rather than compete.
The first is that the scaling axis moved. For a decade, capability gains were dominated by pretraining — a bigger model, more data, more compute in the oven. Test-time compute opened a second axis: train the model with reinforcement learning on problems whose answers can be checked — unit tests, proof checkers, completed tasks — and it learns to spend thinking time productively. The last five points live there, which is why every frontier release now ships with an effort dial instead of just a bigger brain.
The second is a mechanism, and it explains the verbosity better than any strategy does. The reinforcement learning that produces these models rewards the trajectory, not the token: unless efficiency is explicitly part of the reward, a correct answer can earn the same whether it took eight thousand tokens or eighty thousand. Naively penalising length sounds like the obvious fix, but done crudely it trades accuracy away, because the reward arrives at the end of the journey and the cost accrues along it — so the labs err on the side of letting the model ramble. Nothing in training charges the model for its time. The meter was installed afterwards, on you.
The third is the slack, and it's the one I find most telling. If the verbosity were the true price of the capability, you couldn't remove it. You can. A growing pile of 2026 research — overthinking studies among them — cuts reasoning tokens by a third to two-thirds with no loss of accuracy, sometimes with a gain, essentially by teaching models to notice when they're finished. The overthinking is not physics. It's what deliberation looks like when nobody inside the training loop is paying for it — an optimisation deferred because, for the lab, the leaderboard score is the deliverable and the token bill is someone else's.
The Leaderboard Doesn't Show the Meter
This quietly wrecks the leaderboard as an institution. "Model A scored 66 and Model B scored 62" reads like a fair fight. It conceals that A may have consumed several times the inference to get there — a ranking of boxers with the weight classes abolished. The benchmarkers have noticed: Artificial Analysis now publishes cost-per-task and token-consumption charts alongside every score — the invoice printed next to the trophy — which is the only reason any of the numbers in this piece exist. The labs have noticed too, which is why launch week now comes with introductory token prices and well-timed cache discounts — the retail arts, arriving in a market that used to publish one number and stand behind it.
Thirteen Cents
Then there's the foil, and it isn't one model any more — it's a shelf. Qwen, DeepSeek, GLM, and now Xiaomi's MiMo-V2.6-Pro: open-weights models, cheap, increasingly common, increasingly available. MiMo currently tops the shelf — a trillion total parameters, 42 billion active, scoring 46 on Artificial Analysis's index at 13 cents per task. Grok 4.7 at its highest effort setting posts the same 46. It pays $3.74 per task to do it — twenty-nine times as much, for the same score. On Artificial Analysis's current intelligence-versus-cost frontier, MiMo is the only model whose weights you can download — and the shelf it came from restocks every few months.
Which sharpens the question this whole autumn of releases keeps asking: are we paying for better intelligence, or for increasingly elaborate ways of extracting intelligence from the model? I genuinely don't know where the line sits, and I don't think the labs do either — both are being sold in the same SKU, and nobody itemises the receipt.
Which Model Do You Use?
A colleague messaged me at work this week: which AI model do you use? My answer came out, without my planning it, as a staffing plan. Luna for mock-ups, because it genuinely makes the best UI and it's cheap to iterate on. Sol or Astra to think through heavy data analysis and number crunching. Fable 5.1 when the problem is structural — how information actually flows through a system — or when I need to think through a complex situation. Opus 4.8 as a good general mid-model — and 4.8 over Opus 5 is a deliberate choice. I've tried Opus 5, I really have. It is sometimes great and sometimes the worst model I have ever used, and you can't staff a job with a coin toss. Given the choice, it's also Fable 5 over Fable 5.1 — four points up the index, 20% up the bill, and I keep picking the old one. Grok I do use, and it's even available at work — I just can't quite bring myself to open it there yet. It would feel like watching porn at work. And when the real job arrived — pull the data out of chats, emails, transcripts and documents, then build the deck — the answer wasn't a model at all. It was two: I'd get Opus 4.8 to do it, and Fable 5.1 to check its work.
Then the roster moved under me. Opus 5 has been poor for my work since the day it arrived. Opus 4.8, my steady mid-model, kept dropping connections halfway through jobs. And the week before Opus 5.5 launched, Fable 5.1 went soft — not on a leaderboard, in my actual work. Workflows I rely on broke, and I moved them to Codex at work, which I could only do because it was already on the roster. A friend who uses AI every day — no specialist, by her own account — told me there are daily automations she can no longer run that worked fine before. The standard reply is that we're imagining it. We weren't in April, at least: Anthropic's own postmortem admitted weeks of degraded Claude Code — including quietly lowering the default reasoning effort to make it faster. Fewer thoughts, worse answers, and nobody was told at the time. If you know, you know.
Then Opus 5.5 arrived, and it's a genuine leap. At high effort, Artificial Analysis has it six points above Opus 5 at the same setting for half the cost per task, on fewer tokens rather than more. At max — the setting behind the headline number — it tells the old story: 63% more output tokens per task than Opus 5, and a 20% price cut that almost exactly cancels it out. Same model, two stories, depending on which dial you read.
That's the other thing the staffing plan is for. Not just cost — cover. Any one supplier can have a bad week, and nobody announces it.
I'm paid to use these models, and I pay to use them — there is a budget every month, professional and personal, and it gets spent against exactly these economics. The working answer, after enough invoices: nobody who pays the bill buys "the best model" any more. You spend the expensive thoughts where thinking is the job, and the cheap ones everywhere else. Sometimes the flagship doesn't do the work; it checks the work. Deliberation gets deployed the way any dear resource does — sparingly, and where it pays.
We used to buy intelligence by the model. Now we hire it by the thought. I've stopped asking which one is smartest. I ask who is worth it for this job — and who should check their work.