The Price Fell. The Bill Didn’t.

In June 2026 a new AI model cost more to run than the one it replaced. OpenAI launched GPT-5.5 in April at roughly double its predecessor’s per-token price, and I wrote about rising costs on that basis.²

This is not the case right now. Anthropic cancelled a planned 50% price rise on its mid-tier Sonnet model in mid-August 2026 and locked the old rate in.^corrected On 1 September it cut the price of cached reads on its top model, Fable, by 75%. Cached reads are what agents do constantly, so on that model the cut passes straight into agentic workloads.^corrected On 22 September it released Opus 5.5 at US$4 in, US$20 out per million tokens, cheaper than its own predecessor.¹ OpenAI answered a week later with GPT-6.1 Sol at $2 and $10. Google priced Gemini 4 Argon at $2 and $10 the next day.²

Three labs landed in the same US$10–20 output band inside ten days, one of them at an introductory rate.

There’s also a new pricing floor. Budget models now run at US$0.50 per million output tokens and below, and OpenAI’s new Decisions service prices bulk classification at $0.10 per million in, with no charge for output.¹² For sorting and triage work, the metered cost of intelligence is heading towards insignificance.

So why didn’t anyone’s bill go down?

Price per token was always only half of it. The other half is tokens per task.

A non-agentic contract summary still costs cents. An agent is different. It plans, re-reads the documents, calls tools, checks its own output, and re-sends the conversation history at every step. The study I cited in June put agentic tasks at roughly 1,000 times the tokens of a single chat exchange, with up to 30-fold variance between runs of the same task.³

The legal AI platforms have shown this to be true, most visibly Harvey. Harvey upgraded its agents in March. Usage rose about twentyfold in 2026, a figure the company itself has used. By June, Bloomberg reported in September, citing people familiar with the figures, Harvey’s gross margin had gone from around plus 50% to around minus 50%. It was reported to be paying its model suppliers roughly $1.50 for every dollar its customers paid it. Harvey declined to comment on the figures at the time and has not confirmed them since.⁴

The legal AI platforms have the best consumption data in the market, and they may have misjudged the curve badly enough to spend a quarter underwater. A law firm or in-house legal team forecasting its own usage may have even less to go on.

By September the margin was reportedly positive again.⁴ Seat prices didn’t move in either direction. The trade press credited Harvey’s new in-house model, but that model was a research preview at the time, not live in the product.⁵ Bloomberg’s sentence, as quoted by the few outlets that kept it, was that margins recovered after the launch “and other changes to how it uses AI”: routing work between cheaper and dearer models, caching what gets re-read, cutting the number of steps the agents take.⁴ The other half of the fix came from its suppliers. The cancelled August price rise lands directly on an agent-heavy bill, and so does the September cache cut for any work running on Fable.^corrected The headline collapse came later; by the time Opus 5.5 and Sol launched, the margin had already reportedly turned.¹

Your cost is price per token times tokens per task. You don’t control the first number, and this quarter it fell without your help. The second number is set by choices: how the agent plans, how often it re-reads the file, which model handles which step, when it stops (controls the platform sets if you run its out-of-the-box agents, and your design choice for anything you engineer yourself). On the cases that have surfaced so far, vendors that post-trained their own models this year were buying fewer steps, not cheaper rates. One published experiment, Fireworks AI’s Ember-1 results from late September, got the same tasks done with about 40% fewer tokens after post-training.⁶

Step limits, scoped tools, a defined point where the agent stops and hands back. Those controls exist, and they are cost controls. On a flat seat the vendor feels them; on consumption pricing you do. And the platforms are moving that way. When Legora moved its flagship tier to consumption in June, its own announcement said customers pay “a recurring Platform Fee plus Credits,” with the details “set out in the customer’s individual agreement.”¹³ The meter is arriving; a rate card hasn’t arrived with it.

Control of the agents and the systems around them has always been the key.

Now something becomes clear about the price you pay, at least while you are on a fixed seat rather than a meter.

In June, a firm paying the going rate for a premium platform was being subsidised. The vendor was losing money on every customer dollar, and its investors were covering the difference. By September the engineering had caught up and the margin was positive again.⁴ Then the input price collapsed. Through all of this the seat price did not move, but the platforms started flirting with consumption pricing.

So the gap between what you pay and what it costs to serve you has swung twice this year, by large amounts, in both directions, and fixed-price customers may be unaware. There is no public pricing to check this against. The platform players’ pricing is opaque. Maybe that is deliberate.

Aggregators publish estimates; when one such analysis ran last year, a company spokesperson called its assumptions “wildly off” but did not provide any clear evidence to the contrary.⁷ Its chief operating officer described the model in September as “seats-based pricing, but with the option to do consumption for folks who are asking us for it.”⁸ The consumption rates are not published either.

This is not necessarily a bad thing. Plenty of markets pay an intermediary to take on price risk and sell certainty, and the clearest example is electricity: the wholesale price is so volatile that a flat retail rate is generally worth paying for. Whether AI price risk is worth the same premium is not yet known. A vendor that carried its customers at a loss for a quarter has a reasonable claim to the recovery. But the gap between the spot price and your price is the product you are buying. If the underlying model price has fallen, the question is what the platform gives you on top of the usage, and whether that is worth what the gap costs.

If a wrapper were passing tokens through, its price would be the spot price plus a margin for what sits around it: the wrapper’s engineering, data storage, additional security, support. So compare the routes. Anthropic lists enterprise access at US$20 a seat a month, with metered usage on top that scales with the model used; a metered bill moves when the rates move, and this quarter they moved down.⁹ ChatGPT Enterprise has no published price; procurement reports put it at $45 to $75 a seat.¹⁰ Microsoft meters Copilot at $30.¹¹ A premium legal platform is reported around $1,200 a seat a month, and that figure has not moved all year, in either direction.⁷

The comparison is not fair, and the unfairness is the point. The wrapper seat buys legal workflow, security the lab seat doesn’t carry, and several frontier models bundled behind one contract instead of one lab’s. Those are real things. Run the lab route raw and you pay for them a different way: engineers and integration work to make a bare model do what the platforms ship out of the box. Either way, at today’s input prices, that layer is most of what the money buys. And for one quarter this year the comparison ran the other way: on the reported June figures, Harvey was spending about $1,800 on models for every $1,200 seat, so its customers were buying tokens below spot.⁴

Run the test on your own numbers with the pass-through calculator: your seat count, discount and usage mix, priced at this month’s spot rates. Every rate in it is editable.

Where the money is visibly going is a different question, and this year it has a public answer.

Harvey became the first official legal assistant of the US Open, put its name up at Fulham’s ground in the Premier League, signed NBA and WNBA teams, and on 28 September launched its first television commercials across broadcast and streaming, with print in the New York Times and Financial Times and outdoor ads as far from home as Sydney.¹² GPT-6.1 Sol launched the next day at $2 in and $10 out.² And in Melbourne, I’ve watched Harvey’s boards, the name plus “AI Tailored for Law”, ring the MCG at AFL game after game this season, around the boundary and along the second tier. No deal was ever announced; the LED inventory there is sold as advertising through the league’s signage agency rather than announced as partnerships, so a buy like that leaves no press trail. You just see it from the couch.²⁰ Legora, Harvey’s main rival, signed Jude Law for a global campaign in April, put its logo on a PGA golfer’s sleeve and on Chelsea’s training kit, and has opened offices at a pace that took it to sixteen cities on four continents by June. Harvey opened Dublin, Paris and Singapore inside six months itself.¹²

To be fair about where that money comes from: it is raise money, not margin. On the reported June margins, customer fees weren’t funding anyone’s stadium signage. These companies have raised more than a billion US dollars between them this year, and brand is a legitimate use of it in a land grab.¹²

But it tells you what the competition is about. Nobody buys the Chelsea training kit to compete on unit cost, and nobody buys stadium signage to reach a few thousand law firm procurement teams. Advertising at this scale says these companies believe their market is much bigger than law firms. With the raw material heading downwards, what a premium seat price buys is some mix of engineering you can’t inspect, security and controls you can partly inspect, and distribution and brand you can see from a stadium seat. Which makes it fair to ask, at the next renewal, how much of the price is the product and how much is the brand.

The June piece noted vendors signalling a move from seats to consumption pricing. Some moved. Legora’s flagship tier went to consumption in June, GitHub moved every Copilot plan to usage billing the same month, and Microsoft metered 365 Copilot’s agent features in September.¹³ The premium legal platforms mostly kept their seat prices and went after their own costs instead.

Firms have started metering themselves regardless. At ILTACON this year a firm’s knowledge management counsel described finding, on a consumption bill, that a single PowerPoint had cost him $2. Another panellist: “most lawyers don’t see how tokens are spent.” The budget norm settling across the market is roughly 1% of firm revenue.¹⁴

And the oldest problem in legal tech procurement remains: opaque pricing. Above the Law reported in March that firms of identical size using the same products routinely pay 40 to 50% apart, because there are no published rate cards and pricing is negotiated firm by firm.¹⁵ A legal AI consultancy’s blog tells of a top firm quoted over £200 a lawyer and cut 60% after one email, with no firm or platform named.¹⁶

The June estimates still hold in US dollars: around US$1,200 per seat per month at the premium end before negotiation, seat minimums still apply, and discounts of 40 to 60% off opening prices are still widely reported.⁷¹⁹ In Australian dollars the band has drifted down with the currency. At this month’s exchange rate the premium range works out to roughly A$17,000 to $34,500 a seat a year, against the A$18,500 to $37,000 I quoted in June, and research bundles can push past A$41,000.¹⁷¹⁹ Note what moved that number: the currency, not the vendors. These products are priced in US dollars, and an Australian buyer carries the exchange rate on top of everything else in this article.

To put a firm-sized number on it: a top-tier Australian firm runs somewhere between 500 and 1,500 lawyers on firm and directory counts, with the AFR’s latest partnership survey putting top-tier partnerships at 126 to 239 partners.¹⁸ Take 750 lawyers as the representative case. Platform seats for every lawyer at the reported list rate come to about A$15.5 million a year; at the widely reported discounts, A$6 million to $9 million. The same lawyers’ usage bought direct at today’s spot rates runs from about A$360,000 for chat-style use to A$5 million for agent-heavy work, plus trivial seat fees. A flat enterprise chat seat for every lawyer, at the midpoint of reported ChatGPT Enterprise pricing, comes to about A$0.8 million.¹⁰ One top-tier firm, Corrs Chambers Westgarth, deployed Harvey firmwide in February.²¹ Its terms are not public, and the 750-lawyer scenario above is an illustration, not an estimate of that deal. No major Australian firm’s terms are public.

The shaded band on that chart is the real Australian story. This quarter’s input-price collapse changed your vendor’s costs, not your bill. The negotiation band changes your bill, and by more than the collapse changed anything: on a 750-lawyer deployment, the distance between list and a well-negotiated rate is about A$9 million a year, which is more than the entire direct-at-spot cost of agent-heavy usage for the same lawyers. Until a consumption rate card surfaces, the discount conversation is the only lever an Australian buyer actually holds.

What hasn’t surfaced anywhere public is the per-unit pricing of the new consumption tiers. If your team is on one, you currently know more about this market’s real prices than the press does.

The questions from June still work, so I’ll only add to them.

And keep the master copy of your team’s knowledge outside the tools, in formats you own. The argument of The Stack You Didn’t Choose got stronger this quarter. Control of the knowledge system has been the main game the whole time.

In June I said tokenomics was becoming a board-level line item. It’s past that now. The price of intelligence is falling toward the floor, and what you pay will be decided by how well the systems that consume it are run, and by who runs them.