Sonnet 5 Costs $2 Today and $3 in September. Model for September.

Anthropic's new midsize model runs agents at near-flagship quality, but the launch price is a promotion that expires on August 31.

If your cost projections use the intro numbers, every one of them is 50 percent too low for the second half of the year.

The bigger shift is architectural: the midsize model can now carry the agent loop that used to require the flagship.

Anthropic shipped Claude Sonnet 5 on June 30. The headline framing from most coverage was "cheaper way to run agents," which is true and also the least interesting part. The part worth planning around is buried in the pricing table.

The number that changes on September 1

Sonnet 5 launched at $2 per million input tokens and $10 per million output tokens. Those prices hold through August 31. On September 1 they become $3 input and $15 output. That is a 50 percent increase on both sides, scheduled, announced, and easy to miss because it lives in a footnote rather than a headline.

For a chat product where a user sends a few thousand tokens and reads a few hundred back, nobody notices. For an agent, it lands differently. Agents are token furnaces. A single autonomous coding run reads files, reasons over them, calls tools, reads the tool output, reasons again, and repeats until the task is done. The input side balloons because every step re-sends accumulated context. Output stays high because the model is writing plans, code, and verification notes the whole way through.

So the class of workload Anthropic is selling Sonnet 5 for, long-running autonomous agents, is exactly the class where a 50 percent price change compounds hardest. If you ran a two-week pilot in July on the intro price and used those numbers to forecast your Q4 agent bill, your forecast is wrong by half. Not by a rounding error. By half.

This is not a trick. Introductory pricing is normal, and Anthropic told everyone the end date up front. The mistake would be ours: building a cost model on a promotional rate and treating it as the steady state. Model your unit economics on the $3 and $15 numbers. If the workload only works at $2 and $10, it does not actually work.

Why the intro price exists in the first place

Anthropic is not being generous for its own sake. The intro price is a wedge. They want teams to move their agent loops onto Sonnet 5 during the summer, get the integration done, wire it into production, and build habits around it. By September, switching away over a price increase costs more in engineering time than the increase itself. That is the whole play, and it is a reasonable one. The right response is not cynicism, it is to make the decision on the real price so you are not surprised into a migration you did not budget for.

The capability story underneath the price

The reason any of this matters is that Sonnet 5 is genuinely good enough to run the loop. On agentic coding it scores 63.2 percent, against Opus 4.8 at 69.2 percent and the previous Sonnet 4.6 at 58.1 percent. On knowledge work it edges out Opus 4.8. And it verifies its own outputs without being told to, which is the behavior that actually separates a model you can leave alone from one you have to babysit.

Six points behind the flagship on coding, ahead on knowledge work, at a fraction of the cost. For most agent work that is the trade you want. The flagship is not the model that runs your loop anymore. It is the model you escalate to when the midsize one gets stuck.

That inverts the default a lot of teams built last year. The old pattern was Opus everywhere, because the cheaper models dropped tasks halfway and you spent more cleaning up than you saved. The new pattern is Sonnet 5 as the workhorse, with a thin escalation path to Opus for the genuinely hard steps. Route the planning and the ambiguous judgment calls to the flagship. Let the midsize model do the ninety percent that is reading, editing, running, and checking.

What to actually do

Three concrete moves.

First, rewrite your cost model on the post-August numbers before you commit to anything. If a project only pencils out at the intro price, it is not a project, it is a demo with a countdown timer.

Second, instrument token usage per task now, while you are on the cheap price, so you have real per-run numbers instead of guesses. You want to know that your average coding task burns, say, 400K input and 60K output tokens, because that turns the September price change from an abstract worry into a line item you can multiply out. Most teams do not measure this and then act shocked when the invoice arrives.

Third, build the escalation path deliberately instead of defaulting to one model for everything. A cheap workhorse plus an expensive specialist, with a clear rule for when to promote a task, beats running the flagship on tasks that never needed it. The verification behavior in Sonnet 5 makes this safer than it used to be, because the workhorse will more often catch its own mistakes before they reach the point where you would have wanted the flagship anyway.

The pattern to remember

Every capable-and-cheap model launch comes with a version of this. The launch price is a marketing instrument, not a fact about the world. The capability is real and worth adopting. The two things are separate, and the discipline is refusing to let the temporary price talk you into economics that only hold until a date on a calendar.

Sonnet 5 is a good model at a good price for exactly the kind of autonomous agent work it was built for. Adopt it. Just adopt it at the September price, because that is the one you will actually be paying while your agents are still running.