Issue No. 1 · Page III · June 2026

Cost and trust — one downstream axis

Page I showed the saturation: nearly every serious model now clears the bar on expert legal reasoning. Page II showed where the chains broke — the failure modes, and the mechanisms behind them. That is the intellectually interesting part.

This page covers what comes after: once capability is settled as a question, a different set of questions opens. Cost and deployment governance are among them. They are real, they are important for anyone making a procurement decision, and they are not a revelation. A 54× spread among identical scores and a short list of non-negotiable governance questions — these are the shape of the analysis, and they live downstream of the mechanism work, not in competition with it.

The cost spread

Among the four models that scored a perfect 20/20, the price of producing that identical grade — the three graded essays, nothing else — spans approximately fifty-four times: from $0.18 at the bottom to $9.74 at the top. Same rubric. Same judge. Same perfect score. Wildly different bill.

The bill is the product of two independent things: how much a model writes (its token appetite; a token is the chunk of text a model reads or writes) and what each token costs (its rate). Score predicts neither. And the two perfect scorers at the ends of the price spread reach the same grade by opposite routes. claude-fable-5 works the problem hard — roughly 96,000 output tokens of reasoning — at a premium rate. deepseek-v4-pro gorges in the other direction, consuming on the order of 365,000 total tokens, and stays cheap only because its per-token rate is rock-bottom. Same 20/20 at both ends; nothing alike underneath.

Between them sits the model the field’s headline numbers tend to skip: gpt-5.5 reached the identical perfect 20/20 on the fewest tokens in the field — about 72,600 total, of which only 37,799 were output, the lowest output count of any model by a wide margin. Terse and right. That is a real axis of difference, and one vendors do not advertise: two models can earn the same flawless grade with a five- to ten-fold gap in how much they write. Token appetite is independent of both score and rate — and it is half of what you actually pay for. The frugal path is its own kind of capability.

verified (exact, labeled): deepseek-v4-pro $0.18 · gpt-5.5 $1.32 · claude-fable-5 $9.74 · all 20/20
6 approximate run-level points · shape = vendor: circle anthropic / square openai / triangle google / diamond other

How the costs were measured

Each model was priced as delivered through its own natural front door. Anthropic models ran on their native harness; others ran via OpenRouter, a widely-used model aggregator, pending direct-API confirmation. Models accessed through an aggregator are marked with a ^‡ in the table above. The cost figures represent the full API spend for the three-task battery — Property, Criminal Procedure + §1983, Evidence — and nothing else; harness overhead is not included. Where channel differences could have introduced cost bias, they were corrected to a common basis before comparison.

One consequence worth stating clearly: the open-weight model (deepseek-v4-pro) ran via an aggregator that, if anything, overstates its cost relative to a direct-API or self-hosted deployment. Its $0.18 figure is therefore a ceiling, not a floor. The directional claim — that the cheapest path to a perfect score is dramatically cheaper than the most expensive one — only strengthens on that correction.

The tier label on the box predicts nothing

Vendors tier their own lineups — “flash,” “pro,” “mini,” “max” — and the names carry an implied ordering: the “pro” model is the capable one, the “flash” model is the fast, cheap one you reach for when the job is light. On this battery, Google’s own pair inverted that ordering twice over. gemini-3.5-flash — the budget-positioned tier — scored 19 and cost roughly $1.52. Its “pro” sibling, gemini-3.1-pro, scored 18 and cost about $0.99.

So the “flash” model scored higher than the “pro” model and cost more — its larger token appetite outran its cheaper rate. A double inversion of the vendor’s own positioning, on both the axis the label promises (capability) and the one it implies (price). The procurement lesson is blunt: a tier name is a marketing artifact, not a measurement. It does not reliably predict the outcome or the bill. You have to test the specific job. (These are the approximate run-level figures this page uses elsewhere; treat them as directional, not exact.)

The sticker price is not the real price

The cheapest path to a perfect score today runs through an open-weight model. That is a striking fact. It means the best achievable result on this battery is no longer locked behind a proprietary frontier model with a premium per-token price. Capability, at least on this measure, has genuinely commoditized.

But for a lawyer, the sticker price is not the real price. The calculation that actually matters is broader, and for most of the options on this list, it does not resolve in the model’s favor.

What this means in practice: the commoditization of capability does not retire the hard questions. It raises the importance of cost and governance by removing the last excuse for not asking them. Previously you might have accepted a governance shortcut because you needed a particular model’s capability. That bargain is gone. Today you can get the same result from multiple providers. The selection criterion that remains is entirely about the deployment: where does this run, who controls it, and can I put client work there?

One caution that cuts across every point on the cost curve

Neither cost nor brand substitutes for verifying the specific task: the newest flagship drew a flawless diagram of a wrong holding — the confident, polished, wrong answer is the one most likely to pass review, wherever it sits on the cost curve. The mechanism behind that failure is page II’s subject (where the chain broke); the property exhibit is at exhibits/opus-4-8-property.html.

Read the work. Choose the deployment.

The numbers in this issue are real, and they are directional. A battery of three essays is not a full assessment of a model’s practice range, and fifty-four times on a small spend does not guarantee the same ratio on a different task mix. What does generalize: the capability ceiling is flat, the cost axis is not, and the governance questions are non-negotiable for professional use regardless of which point on the cost curve you choose.

This dispatch publishes the actual answers so you can check the lawyering yourself — not to offer a conclusion to be relied on, but to give you the evidence to form your own. A score tells you a rubric was satisfied; the work tells you how. The exhibits are at exhibits/index.html: the full prompts and the models’ verbatim answers. The methodology page will show the rubric, the judge, and the channel correction in detail.

Until then: verify the specific task, read the output before it travels, and make the deployment choice with the same rigor you would bring to any vendor that touches client confidences. The model is no longer the constraint. The deployment always was.

‡ Non-Anthropic models in this battery ran via OpenRouter pending direct-API confirmation. Cost figures represent API spend for the three-task battery only; harness overhead excluded. Aggregator costs for open-weight models likely overstate direct-API or self-hosted costs. All costs are channel-corrected to a common basis before comparison.