Claude Sonnet 5 Agent Cost: Real Token Math, Not Sticker Price

You budgeted your Claude agent off the headline price of $2 per million input tokens, ran it for a week, and the bill came in three to five times higher than your spreadsheet said it would. You're not being overcharged. The sticker price is real, but it's the wrong number to budget an agent from, because agents don't consume tokens the way a single chat does.

Quick answer: Claude Sonnet 5 costs $2 per million input tokens and $10 per million output tokens through August 31, 2026 (then $3/$15). But the real cost of running an agent is driven by how it consumes tokens: a 10-step agent isn't 10x a single call — it can run closer to 55x, because every step re-sends a growing context, and Sonnet 5's tokenizer produces roughly 1.0–1.35x more tokens for the same text. Budget from your step count and context growth, not the per-token rate.

What does Claude Sonnet 5 actually cost per token?

Claude Sonnet 5 lists at $2 per million input tokens and $10 per million output tokens during the introductory period, which runs through August 31, 2026. After that, Anthropic's published standard rate steps up to $3 per million input and $15 per million output — a clean +50% on both, per Anthropic's launch announcement. Finout flagged the same jump in its pricing breakdown: a workload that looks cheap this week can quietly cross your old baseline within weeks, without you changing a single line of your prompt.

That step-up matters most for anyone signing off on a budget now for a system that ships in Q2 or later. If your agent barely clears its ROI math at $2/$10, it doesn't clear it at $3/$15. The honest way to plan is to run both numbers and make sure the workflow survives the standard rate, not just the promo.

Why is my agent bill so much higher than the sticker price?

Your agent bill is higher because each step in an agent loop re-sends everything that came before it. When Claude calls a tool, reads the result, and decides what to do next, that entire running conversation — system prompt, instructions, prior tool outputs, retrieved documents — gets fed back in as input on the next step. You pay for that context again on every turn.

MindStudio put a number on it that's worth internalizing: a 10-step agent isn't 10x the cost of a single call, it can be around 55x. The math is intuitive once you see it. Step one sends a small context. Step two sends step one plus its result. By step ten you're re-sending nine steps' worth of accumulated conversation as input. It's the sum of a growing series, not a flat multiple. This is the single biggest gap between what founders budget and what they get billed, and it's why "cheaper per token" and "cheaper per agent run" are two completely different claims.

There's a second, smaller multiplier stacked on top. Sonnet 5 ships with a new tokenizer that produces roughly 1.0 to 1.35x more tokens for the same English text compared to the previous generation, per MindStudio's analysis. So even before the growing-context effect, the same prompt can cost up to a third more than an identical prompt on the old model. Neither of these is a scandal — they're just facts the headline price hides, and they compound.

How do you actually budget a text-heavy Claude agent?

You budget an agent by modeling three things — steps per run, average context per step, and runs per month — then multiplying, instead of taking the per-token rate at face value. The per-token rate is an input to the model, not the answer. Here's the plug-in-your-own-numbers version you can run in a spreadsheet in five minutes.

Here's a worked example for a mid-complexity support agent, at the intro rate:

At the standard $3/$15 rate, that same agent runs roughly $240/month. Now compare that to the naive budget: if you'd multiplied 6 steps by a flat 3,000-token context, you'd have estimated around $27/month and been off by 6x. The tokenizer and retry overhead push it further. This is exactly why we tell founders the monthly API bill is rarely the expensive part of an agent — the design and integration labor is — and it's why a realistic model beats a headline every time. If you're comparing DIY-with-Claude against paying to have it built, our breakdown of custom AI agent cost for small business puts both sides of that math next to each other.

Manual estimate vs. token-model estimate

The naive way multiplies your per-step context by your step count as if the context stayed flat, and it will understate a real agent by 3x to 6x. The token-model way accounts for context growing on every step, the tokenizer overhead, and retries — the three things that actually move the bill.

The gap between these two columns is the difference between a founder who says "AI just isn't ready yet" after a surprise invoice and one who priced the workflow honestly and shipped it.

Common budgeting mistakes that blow up the bill

The most expensive mistake is treating a multi-step agent as a multiple of a single call — that alone is where the 55x-vs-10x surprise comes from. A few others show up on nearly every first bill.

Does the cheaper per-token rate actually save money on agents?

Not automatically. The intro rate is genuinely cheaper per token than the previous Sonnet generation, but for agent workloads the new tokenizer's 1.0–1.35x token inflation and the always-present growing-context effect can eat most or all of that saving. Whether you come out ahead depends entirely on your step count and how tightly you manage context, not on the sticker price.

For a single-call, text-heavy job — summarize this document, classify this email — Sonnet 5 at the intro rate is a clear win on cost. For a 10-step research or support loop, the honest answer is "model it and see," because the multipliers stack in a direction the headline doesn't advertise. This is why the same launch that made the news gets described by careful teams as "cheaper per token, more tokens per agent" — both halves are true.

FAQ

How much does it cost to run a Claude agent per month?

A mid-complexity Claude Sonnet 5 support agent handling around 1,200 runs a month costs roughly $160/month at the intro rate ($2/$10) and about $240/month at the standard rate ($3/$15), assuming 6 steps per run and a growing context. A simple 2–3 step agent at low volume can run a few dollars a month; a heavy 15-step research agent at high volume can run into the low thousands. Model your own step count and volume rather than trusting a single figure.

Why is Claude Sonnet 5 pricing $2/$10 sometimes and $3/$15 other times?

Claude Sonnet 5 is priced at $2 per million input and $10 per million output during Anthropic's introductory period, which runs through August 31, 2026, and steps up to $3 per million input and $15 per million output after that. Both figures are official — the lower one is a time-limited launch rate, and any budget for a system shipping after that date should use the higher one.

Is a 10-step agent really 55x the cost of a single call?

A 10-step agent can reach that multiple, because every step re-sends the entire growing conversation as input, making the total cost the sum of a growing series rather than ten identical calls. The exact multiple depends on how much context each step carries, but the key point holds: multi-step agents cost far more than the step count implies, so budget from accumulated context, not a flat multiple.

Does the new Sonnet 5 tokenizer make text cost more?

Yes, modestly. The Sonnet 5 tokenizer produces roughly 1.0 to 1.35x more tokens for the same English text than the previous generation, so an identical prompt can cost up to about a third more in raw token terms before any other effect. It's a real line item to include in your estimate, not a reason to avoid the model.

If you'd rather not model this yourself — or you've run the numbers and want the agent built and priced predictably instead of DIY-ing the cost surprises — 6omb builds custom AI agents with the token economics worked out up front and deployment in under 30 days. And if you want to learn to model and build these agents yourself, join the free Claude Community where builders share the templates and cost math behind real deployments.