What Claude actually costs a UK business, and the date that changes it

Update, 12 August 2026: the September date has been cancelled

This briefing was published on 9 July 2026 and its headline premise no longer holds. On 10 August 2026 Anthropic made Claude Sonnet 5's introductory API pricing of 2 USD per million input tokens and 10 USD per million output tokens permanent, and stated that the increase to 3 and 15 USD previously scheduled for 1 September 2026 will not occur. There is no price step-up to prepare for. We are leaving the briefing in place rather than deleting it, because the analysis underneath the date is unaffected and still holds: the tokenizer change means the same text produces roughly 30 per cent more tokens on Claude 4.7 and later models, so a flat price per token is still not a flat bill. Read what follows as written on 9 July, with the pricing date now dead. For current sterling figures see our GBP worked example for August 2026.

What changed

Anthropic released Claude Sonnet 5 on 30 June 2026 with the API model ID claude-sonnet-5, and it is now the default model on Claude Free and Pro. It launched with introductory API pricing of 2 USD per million input tokens and 10 USD per million output tokens, described at the time as running through 31 August 2026, with the standard rate of 3 and 15 USD, the same rate Claude Sonnet 4.6 carries, due from 1 September 2026. That step-up was cancelled on 10 August 2026, as the update above records. Separately, Anthropic's Claude Platform documentation records a change that has had far less attention: "Claude Sonnet 5 uses a new tokenizer. The same input text produces approximately 30% more tokens than on Claude Sonnet 4.6." The launch post puts the range at roughly 1.0 to 1.35 times, depending on the content. On 2 July 2026 Anthropic also added organisation-level spend controls to Claude Enterprise.

Why it matters for UK business

The headline most firms acted on was the September date, and it was always the less important of the two. It has now gone entirely. The one that remains is this: a flat price per token is not a flat bill, because the bill is price multiplied by volume, and the volume moved. If your cost model was built on Sonnet 4.6 token counts, it understates the cost of doing the same work on Sonnet 5, and the introductory rate was quietly absorbing the difference. With that rate now permanent, the absorption is permanent too, which is genuinely good news, but it does not make the underlying volume change disappear from a per-document or per-transaction price.

It is also worth being precise about what was never happening. Anthropic was not charging more for Sonnet than it did a generation ago: the September rate would have been exactly the Sonnet 4.6 list price, so a firm comparing like for like on price alone would have seen no change at all. In the event, it is lower than the previous generation and staying there. The change that matters is in what a unit of work costs, and that is a harder thing to notice, because it does not appear on any price page. It shows up on the invoice.

The tokenizer is also not new. Anthropic scopes it to Claude 4.7 and later models and Claude Mythos Preview, so any firm already running Opus 4.7 or 4.8 has been on it for some time and has the data to prove what it did to their token counts. Firms still on Sonnet 4.6 or earlier are the ones facing the change cold.

There is a second-order effect that catches engineering teams rather than finance teams. If the same text now occupies more tokens, then a context window holds less of your document than it did, and an output ceiling tuned on the previous generation may truncate an equivalent response. Neither failure announces itself as a cost problem, but both surface as one when someone raises the limits to fix them.

On the buying side, the published plan structure is seat price plus usage: Claude Enterprise is listed at 20 USD per seat per month plus usage billed at API rates and is billed annually; Team standard seats are 20 USD on annual billing or 25 USD monthly; Pro is 17 USD per month on annual billing or 20 USD monthly. All of those are US dollar list prices. Anthropic does not publish a sterling price list, so any GBP figure in a business case is your exchange assumption, not Anthropic's, and should be labelled as such.

What to do, and what not to do

Do:

• Recount your actual production prompts against Sonnet 5, using the token counting endpoint. Measure the prompts you really run, rather than scaling last quarter's number by a rule of thumb. The 30 per cent figure is Anthropic's general guidance, not a promise about your workload.
• Rebuild the unit-cost model on measured post-tokenizer counts at the current rates of 2 and 10 USD. Those rates are now the standard price rather than a discount with an expiry date, so a business case built on them no longer needs a September contingency.
• Re-check output ceilings and any hard context assumptions in chunking or retrieval code, since the same text now occupies more of both.
• If you quote clients a per-document or per-transaction price, requote it on measured token counts rather than absorbing the difference quietly. The price per token did not rise, but the tokens per document did.
• Turn on the Claude Enterprise spend controls if you are on that plan. Since 2 July 2026 administrators can set an organisation-level spend limit with alerts at 75 and 90 per cent, and see usage and cost broken down by group and by user.

Do not:

• Tell anyone Claude Sonnet 5 gets more expensive on 1 September 2026. It does not, and it never will: Anthropic cancelled that increase on 10 August 2026 and made the launch rates permanent. If you have already put the step-up in front of a client, correct it in writing.
• Multiply the 30 per cent and the 50 per cent together. They apply to different things, and the product is a number no source supports.
• Switch model tier purely to dodge the step-up before you know where your tokens actually go. Caching, output length and model mix usually move the bill more than tier selection does.
• Assume the Enterprise model entitlements give you full role-based control. They are Enterprise-plan and in beta, and role-level model restrictions bind only to custom roles, so User, Admin and Owner roles can still use every model enabled at organisation level.
• Convert the list prices to sterling in a client document without naming the rate and the date you used.

Where The AI Consultancy fits

Building a defensible unit-cost model for a Claude deployment, measuring where the tokens actually go, and deciding which workloads justify which tier is the kind of work our Claude consulting engagements cover. If you have a Claude build running and no confident answer to what a unit of work actually costs on it, that is a scoped piece of analysis rather than a project. Our earlier briefing on why per-token prices fell while bills rose set out the general argument; this is the Claude-specific, vendor-documented case of it. The AI ROI calculator is a starting point for framing the business case.

All figures in this briefing are US dollar list prices published by Anthropic and verified on 9 July 2026 against anthropic.com, platform.claude.com and claude.com/pricing, then re-verified on 12 August 2026. Corrected 12 August 2026: the 1 September 2026 rise to 3 and 15 USD that this briefing was written about was cancelled on 10 August 2026, when Anthropic made the 2 and 10 USD rates permanent. The tokenizer analysis is unchanged. Anthropic does not publish a sterling price list. Pricing pages change; confirm current rates before committing a budget.