Two Price Rises, One Invoice: The 1 September Cost Cliff Nobody Modelled Properly

Claude Sonnet 5 launched on 30 June 2026 with introductory API pricing of 2 dollars per million input tokens and 10 dollars per million output tokens, and that introductory rate runs through 31 August 2026. From 1 September the standard rate applies: 3 dollars per million input and 15 dollars per million output. Taken alone that is a 50 percent increase on both sides of the meter, which is unpleasant but easy to model. The part most teams have not put in the spreadsheet is that the tokenizer shipped with Sonnet 5 produces roughly 30 percent more tokens for the same English text than the tokenizer it replaced. Your bill is a price multiplied by a token count, and both terms moved.

Work the arithmetic once and the shape becomes obvious. A workload that consumed a notional 1 million tokens under the old tokenizer consumes about 1.3 million under the new one. At the introductory 2 dollar rate that is 2.60 rather than 2.00, which is the number you have been quietly paying since July without necessarily noticing, because a single line item on an API invoice does not tell you how many tokens a paragraph of your own text turned into. From 1 September the same work costs 3.90. Against a pre Sonnet 5 baseline that is not a 50 percent rise, it is closer to 95 percent. If your finance model says the September API line goes up by half, it is wrong by roughly the same amount again.

The reason this is worth an hour of attention rather than a shrug is that 1 September is a compound date. GitHub moved Copilot to token metered AI Credits on 1 June 2026 and cushioned the transition with promotional credits that run through June, July and August, so Business and Enterprise teams drop back to a base allowance roughly a third the size on the same morning the Sonnet 5 introductory window closes. A team that runs Claude Code or the Claude API for agentic work and Copilot for in editor completion is looking at two independent step changes hitting one monthly invoice, in a month that also happens to be the first full month back from the northern summer, when usage climbs anyway. Three variables moving at once is exactly the situation where a variance gets explained as noise instead of investigated.

The measurement you actually need is not complicated, and it has to be taken before 31 August or it cannot be taken at all. Pull your last 30 days of token consumption by model and by workload, not just the dollar total, because the dollar total is the thing that is about to become misleading. Split it into the categories that matter operationally: interactive coding assistance, batch or scheduled jobs, agent loops that run unattended, and anything customer facing. Then reprice each line at 3 and 15 and compare. What usually falls out of that exercise is that a small number of unattended agent loops account for a disproportionate share of the tokens, because a human gets bored and stops while a loop does not, and those are the lines where a 95 percent effective increase turns into real money.

Once you can see the split, the levers are the ordinary ones and they are more effective than they sound. Prompt caching is the first, because cached input reads at a fraction of the standard input rate and most agentic workloads re read the same system prompt, the same repository context and the same instruction files on every turn. Tiered routing is the second: the question is not which model is best, it is which portion of your traffic genuinely needs frontier reasoning and which can be served by something smaller, with a clear escalation rule for the rest. We made that argument when Claude Opus 5 arrived at frontier capability with mid tier pricing, and the September increase makes the same discipline pay roughly twice as well. Trimming context is the third and least glamorous, and with a tokenizer that is 30 percent more expensive per unit of prose, the boilerplate you never pruned is now costing measurably more than it did in May.

There is a governance point here that founders tend to hear as bureaucracy and later wish they had heard as protection. Swapping a default model, changing a routing rule or turning caching on across a production workload is a change to a system your customers depend on, and if you hold SOC 2 or you are working toward it, that change has a control sitting on top of it. The auditor question is not whether you were allowed to change models. It is whether the change was authorised, whether anyone assessed the effect on output quality before it went live, and whether you can show the assessment. A cost driven model swap made in a hurry on 2 September, under pressure from an invoice, is precisely the change that gets made without a record and then has to be explained a year later.

The same applies with more force if you have implemented ISO 42001 or are heading toward it, because an AI management system expects each AI system in use to have an inventory entry, an accountable owner and monitoring that would catch degraded performance. Routing a share of your traffic to a smaller model to control cost is a legitimate decision, and it is also a change to the capability of a system your customers touch. Record it as one. The practical artefacts are short: which models serve which workloads, who approved the routing rule, what quality check was run before and after, and what the rollback looks like. Vanta, Drata, Secureframe and Sprinto will all hold that against a control and enforce a review cadence, which is a great deal easier than reconstructing the reasoning from memory during an audit.

It is worth being fair about what Anthropic has actually done here, because the introductory window was disclosed at launch and the standard rate was published from day one. Nothing is being sprung on anyone. The tokenizer change is a different matter, not because it was hidden, but because token counts are the kind of technical detail that never reaches the person who owns the budget. That is the real lesson and it outlives this particular date. In a token metered market the price per unit is only half the story, and vendors compete on both halves. A model that is cheaper per token and more verbose in its output, or hungrier in its tokenization, can be more expensive in practice than the headline suggests, and the only way to know is to measure your own workload rather than compare price lists.

So the action for the next four weeks is narrow and worth doing properly. Baseline your token consumption now, while the introductory rate still makes the comparison clean. Reprice it at the September rates and look at the total alongside the Copilot allowance change so you see one combined number rather than two surprises. Turn on caching wherever the context is stable, put a routing rule in front of the workloads that do not need frontier reasoning, and run both changes through whatever change process you already have rather than around it. Then set a calendar reminder for the first week of October to check the real invoice against the forecast, because the value of this exercise is not the forecast itself, it is finding out how good your model of your own AI spend actually is while the stakes are still small.