Claude Sonnet 5 Is Cheaper Than Opus
— Until August 31
Read the Footnote.

Anthropic just handed your AI agent a 2.5x discount. Read that again — not a smarter model, a cheaper one. And they're quietly hoping you don't read the footnote that comes with it.

The footnote says: through August 31, 2026. After that, the discount is smaller, the model still burns roughly 3x more agentic turns than the last one, and the "cheaper than Opus" headline you built your budget on stops being true. I've been running agents on a content pipeline every single day, and if there's one thing this year taught me, it's this: the price per token lies. The price per task tells the truth.

So let's do the thing nobody in the launch coverage did. Let's read the footnote out loud, then run the real math — the one that decides whether migrating your agent stack to Sonnet 5 saves you money or hands you a surprise cloud bill in September.

1. What Happened

On June 30, 2026, Anthropic shipped Claude Sonnet 5 — the most agentic model in the Sonnet line, with the API ID claude-sonnet-5. It carries a 1M-token context window and became the default model on Claude's Free and Pro plans the same day. That last detail matters more than the launch post admits: millions of people started talking to Sonnet 5 immediately, most of them not knowing the model changed under them.

The headline number is the price. Anthropic set introductory pricing at $2 per million input tokens and $10 per million output tokens. Against Opus 4.8's $5/$25, that's exactly 2.5x cheaper on both sides. On benchmarks, Sonnet 5 scores 63.2% on agentic coding — versus Opus 4.8's 69.2% and Sonnet 4.6's 58.1% (numbers via TechCrunch). Its Intelligence Index is 53; Opus 4.8 sits at 56.

A second story landed in the same 24 hours. Per Reuters, the US export-control directive that had suspended access to Claude Fable 5 and Claude Mythos 5 was lifted, with access restoring from July 1. Anthropic's newsroom confirmed the directive existed; the exact removal date is single-sourced, so treat it as "per Reuters." Two events, one day: agent economics and the geopolitics of model access hitting the same point. That's rare, and it's why this is worth writing about today rather than next week.

2. Why This Is a Paradigm Shift

The shift isn't "the model got smarter." It didn't, much — 53 vs 56 on the Intelligence Index is a rounding error to most real workloads. The shift is that Anthropic deliberately moved the war out of the "whose intelligence is higher" arena and into a colder one: who can do agentic work cheaply and reliably without a human standing over it.

TechCrunch put it exactly right: "the differentiator isn't going to be who can do agentic work best, but how cheaply they can do it and how reliably without human oversight." Read that as a strategy statement, not a review. The frontier stopped being a leaderboard flex. It became a unit-economics fight.

And the introductory price is the tell. You don't slash your best agentic model to 2.5x below your flagship because you're feeling generous. You do it to seat everyone on an agentic workflow — to make "just let the model do the whole task end to end" the default habit before the meter goes back up. Habits form in weeks. Prices reset on a date. Anthropic knows which one moves faster. The paradigm shift is that the model is now a loss-leader for the behavior, and the behavior is what gets monetized once the promo ends.

3. The New Architecture in Plain English

An agentic model isn't a chatbot that answers once. It's a model that takes a goal, then loops: think, call a tool, read the result, think again, call another tool — until the job is done or it gives up. Each loop is an "agentic turn." Sonnet 5 runs roughly 3x more of these turns than Sonnet 4.6, and produces ~40% more output tokens per task (Artificial Analysis). More turns means more autonomy — and more tokens billed.

A model that's "cheaper per token" and "does 3x the work per task" is not automatically cheaper per task. Sometimes it's more expensive — Artificial Analysis pegs one benchmark task at about $2.29 on Sonnet 5, and states outright that before the promotional price, Sonnet 5 can cost more per task than Opus 4.8.

The tools are the other half of the picture. An agent with no hands is useless — it can reason all day but can't touch your Salesforce, your files, your inbox. Tools are what give it hands, and the Model Context Protocol (MCP) is how those tools plug in. Zapier's Daniel Shepard, a senior engineer, handed Sonnet 5 a two-part job — update Salesforce account tiers, then send a launch announcement to enterprise contacts — and it finished end to end (via TechCrunch). The model didn't do that because it's smart. It did it because it had connected tools. Cheaper model, same requirement: without the plumbing, autonomy is a demo, not a result.

4. My Content Factory Case (Real Numbers)

I run Content Factory — a content pipeline stitched together in n8n, feeding a Telegram bot, Google Sheets, and a rotating cast of models. Every agentic step in it has a cost meter, not just a quality check. That habit came from getting burned.

Here's the concrete failure that taught me. A research agent that used to cost me one fixed amount per run suddenly, on the same task, racked up multiple times the calls — because a newer model "tried harder." Beautiful in the demo, painful on the invoice. It didn't get a wrong answer; it got the right answer by taking a scenic route through my token budget. That's exactly the Sonnet 5 profile: more turns, more output, more autonomy — and a bill that doesn't move the way the per-token headline promised.

So this week I'm not swapping Opus out for Sonnet 5 wholesale. I'm placing Sonnet 5 on specific links in the chain — the steps where speed matters and a mistake is cheap to recover: first-draft generation, competitor scraping summaries, routing. On those, I measure two numbers per run, not one: quality and cost-per-run. The intro window through August 31 isn't a gift to me. It's a free trial to build my own unit-economics table while errors cost nothing. By September I want a spreadsheet that tells me, per pipeline link, whether Sonnet 5 stays or goes — as arithmetic, not vibes.

5. The Cost Math That Wakes Up CFOs

Let's put the two pricing regimes side by side, because the gap between them is the entire trap.

Now layer in behavior. Sonnet 5 runs ~3x the agentic turns and ~40% more output tokens per task than Sonnet 4.6. Artificial Analysis benchmarks one task at ~$2.29 on Sonnet 5 and says flatly that, before the promotional price, Sonnet 5 costs more per task than Opus 4.8.

Put it together and the CFO takeaway is one sentence: the "2.5x cheaper" line is true on the price sheet during the promo, and can be false on the invoice the moment you measure a real task. The discount narrows from 2.5x to 1.67x on September 1 while token consumption stays 3x heavier. If your team migrates the whole agent layer to Sonnet 5 "because it's cheaper" and never re-measures cost-per-task in September, the risk isn't that you didn't move — it's that you moved without counting. That's how a cloud bill grows for no visible reason.

The rule that survives all of this: compare price per finished task on your workflow, not price per token on a slide.

6. What Dies, What Lives

7. What to Build This Week

You have a free trial with a deadline. Use it as a measurement lab, not a shopping spree.

By August 31 you'll own something more valuable than a discount: a real unit-economics table for your own agents.

8. The B2C / B2B Split

Frequently Asked Questions