Token prices held. Salesforce's bill still had a say in its margin guidance.

What the article says

Lindsay Clark's report covers Spencer's remarks at the Deutsche Bank Technology Conference. Roughly six months ago Salesforce let Claude loose across its R&D cycle, and that is part of the reason full-year margin guidance stayed where it was. The figures as printed: 20.5 percent operating margin for the quarter ending 31 July, 20.1 percent full-year guidance, and Benioff's $300 million Anthropic figure from May.

Next, Spencer said, comes refinement: the model gets chosen for the task, a second- or third-generation model is often fine, and the optimisation runs across vendors. Internally the company has OpenAI, Cursor and Claude, and is experimenting with X's Grok.

Two things worth spelling out, because the article leaves them implicit. It does not explain why the spend is as large as it is. And it is about Salesforce's own engineers' coding tokens: nothing about customer data, per-seat pricing or Salesforce's product prices. We did not open the conference transcript; Spencer's remarks reach us through The Register.

We agree with the diagnosis: a line item large enough for a finance chief to mention alongside margin at an investor conference is a planning problem. Our conclusion differs, and below we say where it is weak.

The price list rules out the easy explanation

Sonnet 5 was scheduled to go from $2 to $3 per million input tokens and from $10 to $15 per million output tokens on 1 September. The increase did not happen; the introductory price is now the standard price. On Fable 5.1, a cache read costs $0.25 per million tokens against $1 on Fable 5, a quarter of the rate, while input and output are $10 and $50 on both. Anyone wanting to explain Salesforce's bill with a price rise has no evidence on the price list.

A bill is unit price multiplied by quantity. If the unit price held or fell and the bill still had a say in margin guidance, the quantity grew. The Register does not say that; it is our inference. What we can list is what makes quantity grow.

The first is adoption: let a tool loose on an engineering team and in month six more people use it for more things than in month one. That is the point. The second is steps per task: newer models run more reasoning steps for the same job, as we wrote on 4 September about Gemini Flash 3.8. The third is tokenisation: Anthropic's own pricing page notes that the tokenizer used from Claude 4.7 onwards produces roughly 30 percent more tokens for the same text, and that the exact increase depends on the content. None of the three appears on a price list. All three appear on the invoice.

You know the unit price. You do not set the quantity.

This is the defensible claim, and it is narrower than it first looks. Per-token, per-seat or per-call pricing tells you exactly one thing: the unit price. Total spend depends on quantity, shaped partly by the vendor (model behaviour, tokenizer, defaults) and partly by your own success (more usage means the thing works). Neither side gives the buyer a control that fixes the annual total in advance.

Salesforce's answer is the right one at its scale: model per task, older generation for simpler work, load spread across vendors. That takes hold of the quantity side. Usage stays; cost per unit of work falls. At $300 million a year, that work pays for its own team.

For a company of 20 to 200 people the lesson is simpler. If your finance lead cannot give you next year's AI line even as a range, you do not have a budget. You have a unit price.

What we think, including the parts that cut against us

First, what this is not. Salesforce is spending tokens on its own engineers' code: a cost story, not a sovereignty story. No regulated customer data appears anywhere in the piece, and turning it into a data-protection argument would be dishonest.

Second, Salesforce's answer is smarter model choice, not self-hosting. Nothing suggests its own hardware would be cheaper at that scale; it very probably would not be. Salesforce is not our customer. We work with companies of 20 to 200 people in regulated or data-sensitive sectors, where the question is different: does the whole workflow fit on known capacity at a known monthly fee?

Third, guidance was not cut. It was not raised. The 20.5 and 20.1 percent are as printed, and anyone who writes that AI pushed Salesforce's margin down is misquoting the source.

Fourth, this is our fourth cost piece in one week. It earns its slot only because the mechanism is different, quantity in a week when list prices stood still, and because the counter-evidence is in the text: Anthropic dropped the increase it had planned for 1 September and cut one price to a quarter. Without that, this article would be advertising.

Fifth, a flat monthly fee is not the absence of risk. Our own hardware plus a known operating cost is a different risk profile: the variance is ours, the ceiling is yours, and the ceiling is real. A given capacity holds a given quantity; outgrow it and you make a hardware or tier decision in advance, with a price tag on it. The quantity question does not go away with us. It gets asked earlier.

Sixth, our own house. Where the data allows it, we use public coding models in development too, so Salesforce's cost problem touches us at a smaller scale. The difference is who pays: we do, and the customer's monthly fee does not move.

Where we stand

If you want the three routes side by side, public API, self-hosting and an operated private platform, our own private AI guide sets out the cost comparison with the drawbacks of each. What we actually operate is on the platform page; to walk through your own workflow, tell us what you are working on.

Consumption pricing is not wrong. For experiments, low-stakes work and fast-changing needs it is exactly right, and Salesforce is right to shop across vendors. The trouble starts when a dependency billed by quantity sits under something the company uses every day and nobody wants to switch off. From then on the bill follows the vendor's release calendar and your own success, and neither moves on your schedule.

If an AI cost already shows in your margin, knowing the unit price is not enough. You need to know where the top is.

Guides: Private AI for business · Private ChatGPT · Private AI hosting · Private AI platform · Your own private AI · GDPR-compliant AI hosting