Anthropic released Claude Sonnet 5 on June 30, 2026, and buried a deadline in the launch: the model costs $2 per million input tokens and $10 per million output tokens only through August 31. On September 1 the price moves to $3 and $15 — a 50% increase on both sides. If you are building anything on Sonnet 5 right now, your cost model has an expiry date, and it is closer than it feels.
Here is what shipped, what the deadline actually changes, and how to plan for it.
What Sonnet 5 is
Sonnet 5 is the new version of Anthropic’s midsize model, and the pitch is explicitly about agents: it plans, uses tools like browsers and terminals, and runs autonomously at a level that previously required larger, more expensive models. Anthropic made it the default model for free and Pro plans, which tells you how confident they are in it as an everyday workhorse.
The positioning is aggressive on price. At the launch rate, Sonnet 5 undercuts Anthropic’s own Opus 4.8, OpenAI’s GPT-5.5, and Google’s Gemini 3.1 Pro, while coverage from The New Stack describes its quality as closing much of the gap with Opus 4.8. Even at the September rate of $3/$15, it stays in midsize-model territory — the launch window just makes the first two months unusually cheap.
The part most teams will get wrong
A launch discount on an agent model is different from a launch discount on a chat model, because agents multiply tokens in ways that are easy to underestimate.
A chat request is one round trip. An agent task is a loop: read files, call tools, read the results, decide, call again. A single “fix this bug” run can burn dozens of model calls, and each call re-sends accumulated context as input tokens. That means the input price — the side that jumps from $2 to $3 — is usually the side that dominates agent bills. Output tokens get the headlines; input tokens get your invoice.
So the honest way to read the deadline: if you prototype an agent workflow in July and August, your September bill will be roughly 50% higher per token than your prototype data suggests — before any growth in usage. Budgets built on July numbers will quietly drift.
What to do before September 1
Three concrete moves, in order of value:
-
Measure your input/output split now. Log tokens per task while the discount runs. If input is 80% of your spend (common for agent loops), the September increase costs you close to the full 50%. If your workload is output-heavy, the practical hit may land differently. You cannot plan around a split you have not measured.
-
Decide which steps actually need Sonnet 5. Agent pipelines usually contain a few steps that need real reasoning and many steps that are mechanical. Routing the mechanical steps to a cheaper model — or to no model at all, where a script does the job — is the one change that keeps working after every future price change, from every vendor.
-
Re-check the numbers before you commit. Prices in this market change on weeks’ notice, and this article is a snapshot: rates verified as of 2026-07-17. Before budgeting, confirm the current figures on the official Anthropic pricing documentation and the Sonnet 5 announcement.
The bigger signal
The two-month discount is not generosity; it is an on-ramp. Anthropic is racing to get agent workloads settled onto Sonnet 5 while the price is low, betting that switching costs will hold those workloads when the price normalizes — VentureBeat ties the push to the company’s IPO ambitions. That is a fair trade as long as you walk in with your eyes open: take the cheap months, use them to gather real cost data, and make your September decision from measurements instead of launch-week enthusiasm.
Who needs to revisit the budget
If you use Sonnet 5 mainly as the default model in a personal chat plan, the API deadline is not by itself a forecast of your subscription bill. The article’s pricing window concerns the listed per-token rates. Your task is to keep plan access and API consumption separate when you decide whether the change affects you.
If you are a solo developer calling the API, the decision is more direct. A prototype that looks affordable at $2 per million input tokens and $10 per million output tokens has not yet shown you its September economics. You need the token record for each repeated task and a second view calculated at $3 and $15. The model may still be the right choice; the launch rate simply cannot be the only case in your budget.
For a team running agents, the repeated context in tool loops makes measurement the priority. Do not average every workflow into one account-wide number. Separate the agent jobs that read files and tool results repeatedly from shorter interactions, then identify which stages require Sonnet 5’s reasoning and which are mechanical. That gives you an actual routing decision instead of a blanket keep-or-replace debate.
Run this check before August 31
- Choose a representative task. Use work your agent will repeat, rather than a one-off launch experiment.
- Record input and output tokens for the complete task. Include the calls made as the agent reads files, uses tools, reads results, and decides what to do next.
- Price the same record twice. Apply the $2/$10 launch rates, then the stated September rates of $3/$15. Keep usage constant so the comparison isolates the price change.
- Mark every stage that could use a cheaper model or a script. Do not move a reasoning-heavy step merely to make the spreadsheet look better; start with work that is already mechanical.
- Set the decision date before the rate changes. Use the July and August window to gather evidence, then re-check Anthropic’s current figures before committing.
This check gives you two useful outputs: a September cost at the same workload and a list of routing changes you can test. It does not require you to predict growth before you understand the current loop.
How to frame the tradeoff
Keep Sonnet 5 across the pipeline when measured task quality and agent capability justify the September cost. The fact that it remains in midsize-model price territory may make that an acceptable result, but the decision should come from your own task record.
Use selective routing when only a few steps need the model’s planning and tool-use strengths. This is the most durable response to the price window because it reduces expensive tokens without assuming that every step is interchangeable.
Use the launch period for evaluation, not as proof of the permanent economics. The 50% per-token increase is already stated; your input/output split and the number of calls in the loop determine how that change appears in the total workflow. If Anthropic changes the published rates, or if your measured token mix changes after you alter the pipeline, redo the comparison rather than carrying the earlier answer forward.