A fortnight ago a CFO I advise was celebrating. Her team had shifted a large customer-service workload onto Anthropic's new Claude Sonnet 5, priced at the same $3 and $15 per million tokens as the model it replaced, with an introductory rate set below even that. She assumed the migration was, in effect, free. Three weeks on, the invoice arrived heavier than before. The sticker price had not budged. The monthly bill had.

Here is the number that refuses to add up. Sonnet 5's list price matches its predecessor, and to Anthropic's credit the company was upfront that its new tokenizer would count more tokens for the same text, and it set the introductory period so the switch would land roughly cost-neutral. The trouble surfaces once you push real work through it. Independent analysis via Artificial Analysis found the new tokenizer counts up to 1.35 times more tokens for identical text, the model emits roughly 40% more output tokens, and it takes about three times as many agentic turns to finish a task. Stack those together and the effective cost lands near 15% above Opus 4.8, the pricier flagship it was meant to save you money against. Same job, same advertised price, more money to complete it. Everyone this month is reading a price war. I am reading a pricing illusion, and if you stop here you will keep optimising the one number, price-per-token, that no longer maps to what you pay.

A price war with theatre up top and fake savings underneath

The last two weeks looked like a rout in the buyer's favour. xAI shipped Grok 4.5 at $2 and $6 per million tokens, landing on par with GPT-5.5 on Terminal Bench, the kind of agentic coding test that used to separate the flagships. Meta undercut everyone with Muse Spark 1.1 at $1.25 and $4.25, well below GPT-5.5 and Opus. OpenAI, with its own chief openly rattled by rising costs, has been planning sharp price cuts and steering buyers toward cheaper tiers for lighter work. On a spreadsheet of headline rates, costs are collapsing.

Then look at what actually drives spend. I call it tokenmaxxing, and it works through three quiet levers sitting underneath the advertised rate. One is tokenizer inflation: a model can charge the same per token while its tokenizer chops your identical prompt into more tokens, as Sonnet 5's does at up to 1.35x. Another is output verbosity, the roughly 40% of extra tokens a model spends saying the same thing. The last is agentic loop length, the number of turns a model burns to reach a finished answer, which for Sonnet 5 ran close to triple. A model that looks cheaper per token can be dearer per finished job, and no line in the pricing table warns you.

Grok 4.5 makes the same point from the other direction. xAI's real pitch has little to do with the $2 rate. What xAI is selling is roughly 4.2 times fewer output tokens than Opus 4.8 on comparable work, an efficiency argument dressed as a price argument, and the efficiency is what moves the invoice. The people closest to the meter already sense this. Sam Altman has called escalating token costs a "huge issue," and one Uber executive described burning through the company's entire 2026 AI budget on agentic use cases before the year had properly got going. Agentic systems multiply turns, and turns are where tokenmaxxing feasts. A 30% cut to a per-token rate means nothing if your agent now loops five times where it once looped twice.

The frontier just became a commodity input

So the savings are staged. The commoditisation, though, is real, and that is the part worth your attention. Capability at or near the frontier is now available from many vendors at prices sliding toward the floor, and a growing share of that supply is open-weight and Chinese. On OpenRouter, Chinese open models have been reported to account for more than 30% of the token traffic coming from US companies, and the draw is straightforwardly price against the American incumbents.

The specifics should make a CTO sit up. Zhipu's GLM-5.2 shipped open-source under an MIT licence with a one-million-token context window, at roughly a tenth of the cost of Anthropic's competing tier. DeepSeek, for its part, is building its own inference chip to cut its reliance on Nvidia and Huawei, and has open-sourced a framework to speed up serving. When a top-tier context window can be had for a tenth of the price, and the supplier is driving its own serving cost down at the silicon level, the question "which model is best" stops being a strategy question. It turns into a procurement detail that changes every quarter.

I want to be plain about what this means, because a lot of leadership energy is still going to the wrong place. Chasing the single winning model is fighting the last war. The frontier model has become an input, like bandwidth or compute, where you assume plenty of adequate suppliers and design so you can swap between them cheaply. Bandwidth stopped being a differentiator the moment everyone could buy it. A model checkpoint is heading the same way.

Own the plumbing

If the model is a commodity input, the durable asset is the plumbing you build around it, and none of it shows up on a vendor's pricing page.

Start with a model-agnostic routing layer, so any given request goes to the cheapest model that clears your quality bar for that request, and so moving a whole workload from one vendor to another is a config change rather than a rebuild. OpenAI steering buyers toward its own cheaper internal tiers is a quiet admission that routing is where the value sits; they are handing you the lever inside their walls because they know you want it. My advice is to keep that lever outside any single vendor's walls instead.

Alongside it, run a private evaluation harness on your own workloads rather than on public leaderboards. Serious buyers already decide this way. A Databricks benchmark reported by Forbes found open models, GLM-5.2 among them, matched premium closed models on quality while costing about two-thirds as much per completed task. Leaderboards have stopped deciding enterprise buying, and rightly so, because a model that tops a reasoning chart can still lose on your data once tokenmaxxing is priced in.

And then there is the one metric that ties the whole thing together: cost-per-completed-task. Price-per-million-tokens can go in the bin. Everything I have described folds into that single figure, and it is the only one that matches your invoice. Measure it per workload and your own surprises will surface, the way that CFO's did.

The prescription follows from there. Route high-volume, low-difficulty work to the economy tiers and to open weights, where the per-task cost is lowest even when the per-token rate is not. Reserve frontier models for the hard reasoning that warrants them, and prove that case on your data before you commit. Benchmark continuously, because the winner keeps rotating. Steer clear of single-vendor lock-in, because the one certainty in this market is that today's best deal is next quarter's overpriced default.

India got here first, out of necessity

There is a reason this reads as familiar to me. Indian enterprises have been running this playbook for a while, out of plain cost discipline rather than any grand strategy (though it amounts to the same thing). When budgets are tight, you learn fast to right-size the model to the task. Enterprises here are picking smaller, right-sized models over frontier LLMs on cost grounds, treating inference efficiency, the business of minimising tokenmaxxing, as the competitive lever rather than access to the biggest checkpoint. That is the posture the rest of the market is now backing into.

Data residency sharpens the same instinct. Regulated buyers in banking, insurance and government cannot always ship data to an export-restricted API sitting offshore. Self-hostable open weights such as GLM-5.2 and Qwen sidestep the problem, which is a big part of why they are spreading through Indian BFSI and public-sector work. The hyperscalers are responding to the pull: Google is now hosting Gemini models domestically in India for lower latency and residency compliance. An enterprise that must keep data in-country and cannot lean on any one foreign API has little choice but to build the model-agnostic, cost-per-outcome stack I have been describing. Constraint forced the architecture that the unconstrained are only now discovering they need.

The contrarian punch is uncomfortable for the Western buyer drunk on cheaper-token headlines. The organisations that looked behind, the ones that could never treat the frontier model as a status purchase, turn out to have built the right muscle. They optimise for outcomes because prestige was never on the menu.

My take

The price-war coverage is a distraction, and an expensive one, because it trains buyers to defend a number that has quietly stopped meaning anything. Most enterprises are not going to be sunk by picking the wrong model. They will be sunk by building the whole stack around one model at all, a model that will be a commodity by the next quarterly launch and might well be beaten by an open-weight release before your procurement cycle even closes.

What survives all of that is the plumbing. Own your routing and your evaluation harness, and above all own your cost-per-completed-task data, and you own the one thing no vendor can undercut with a press release. I would rather my team spend a month building a clean way to swap models than a month arguing about which model to marry. Bet on portability. Logos are the thing you should be able to throw away. And if you want to see what the finished version of that bet looks like, study the buyers who never had the luxury of the alternative.