Anthropic’s New Sonnet Isn’t Built for You. It’s Built for Your Agents

It benchmarks close to Opus 4.8, undercuts it on price per token, and still somehow costs more to run — here’s why, and what to actually do about it.

Photo by Brecht Corbeel on Unsplash

On June 30, Anthropic shipped a Sonnet model that costs more to run than its own flagship. Not on some obscure edge-case benchmark. On the exact test Sonnet is supposed to win by design — the cheap, sensible middle option next to the expensive one at the top.

Independent testing from Artificial Analysis found Sonnet 5 costs $2.29 per task on their Intelligence Index at standard pricing. That’s roughly double what Sonnet 4.6 cost, and about 15% more than Opus 4.8. Same benchmark. Higher bill. For the model that’s supposed to be the budget pick.

Here’s the part everyone’s launch-day coverage glossed over: that’s not really a pricing failure. It’s a symptom. Sonnet 5 spends tokens the way a nervous new hire spends time — checking, re-checking, spinning up help it wasn’t asked for — and depending on what you’re using it for, that’s either the most annoying thing about this release or the most interesting.

I think it’s both. Let me show you why.

The Chart Anthropic Published vs. the Bill You’ll Actually Get

Anthropic’s own numbers aren’t fake. Sonnet 5 hits 63.2% on their agentic coding benchmark, up from Sonnet 4.6’s 58.1%, closing in on Opus 4.8’s 69.2%. On a knowledge-work benchmark called GDPval-AA v2, it actually edges past Opus 4.8–1,618 to 1,615. Terminal-Bench jumped by roughly 20 points over the previous Sonnet. None of that is spin. It’s a real, measurable step up from Sonnet 4.6, and on paper it lands shockingly close to a model that costs substantially more per token.

artificialanalysis.ai

That per-token price is genuinely lower too. Sonnet 5 launched at an introductory $2 per million input tokens and $10 per million output tokens through the end of August, after which it settles at $3 and $15 — the same list price Sonnet 4.6 already carried. Opus 4.8 runs $5 and $25. On the sticker, Sonnet 5 looks like the obvious pick.

The problem shows up once you stop looking at price-per-token and start looking at cost-per-task. Sonnet 5 ships with a new tokenizer, the same one introduced with Opus 4.7, and it maps the same text to anywhere from 1.0 to 1.35 times more tokens than the old Sonnet tokenizer did. That alone eats into the discount. But the bigger factor is that Sonnet 5 just uses more tokens, period. Artificial Analysis measured roughly 40% more output tokens per task at max effort compared to Sonnet 4.6, and close to three times the number of agentic turns on their knowledge-work evaluations. Push it to max effort and it burns through six times more turns than it does at low effort.

So you end up with a model that’s cheaper per token and more expensive per task, which is a genuinely strange place for a “value” release to land.

Why It Can’t Stop Working

CodeRabbit’s independent review of Sonnet 5 put a name to the behavior that’s driving this: the model is chattier than Sonnet 4.6, and it takes noticeably longer even on small requests. Ask for a one-line fix and you’ll sometimes get a test file longer than the change itself. On a big feature, that reads as thoroughness. On a quick edit, it reads as a model that can’t help itself.

There’s a real reason for this, and it’s not just verbosity for its own sake. Sonnet 5 was built to run something close to an evaluator loop — write an attempt, critique it, improve it, repeat — without being told to. Anthropic’s own launch materials describe testers watching the model check its own output unprompted and push a task through to completion instead of stopping halfway, the way earlier Sonnets often did. CodeRabbit’s testing backs this up on the upside: on long, open-ended jobs, that same instinct to double-check itself turns into a real strength, because the model keeps working instead of handing back something half-finished. But that same instinct is also why cranking the effort setting all the way to max, in their tests, roughly doubled the cost without finding meaningfully more bugs. The model doesn’t know when a task is actually small. It just knows how to keep going.

This is also, not coincidentally, where the sub-agent behavior comes in. Sonnet 5 is tuned to decompose bigger jobs into pieces and hand them off, rather than trying to hold the entire task in one long reasoning chain. That’s a genuinely different posture from most mid-tier models, and it’s part of why Anthropic is positioning this release around reliability on long agentic chains rather than any single headline score. The catch is that decomposition itself costs tokens. Every sub-agent you spin up carries its own context, its own reasoning pass, its own overhead — and a model that reaches for that pattern by default, even on tasks that didn’t need it, is going to look expensive next to one that just answers the question.

The Refusal Rate Went Up. So Did the Wrong Kind of “No.”

Buried in the system card is a detail that matters more than it sounds like it should. Sonnet 5 refuses malicious requests more reliably than Sonnet 4.6 did — that’s the intended outcome, and Anthropic reports real gains in resisting prompt injection and agentic misuse. But the same document notes a slight increase in what Anthropic itself calls “wet blanket” responses: replies that are more discouraging, dismissive, or moralizing toward the user than the situation actually called for. Translation: it’s a little more likely to say no to something it should have just done.

For anyone running Sonnet 5 as an unsupervised piece of a larger pipeline — which is exactly the use case Anthropic is selling it on — that’s worth knowing before you wire it in. A model that occasionally declines a benign, well-scoped sub-task is a debugging headache in a way that a model which occasionally makes a small mistake usually isn’t.

The Timing Wasn’t an Accident

Sonnet 5 didn’t launch into a quiet news week. Three weeks earlier, Anthropic had released Fable 5 and Mythos 5, its first Mythos-class models, only to have the Commerce Department order a global shutdown three days later over national-security concerns tied to the models’ cybersecurity capabilities. Every customer, everywhere, lost access overnight. The standoff dragged on for eighteen days, with mounting pressure from tech executives worried the delay was handing free runway to Chinese open-source labs. The controls came off the same week Sonnet 5 shipped, with Fable 5 access restoring the following day.

That sequencing tells you something about what Sonnet 5 actually is. It’s the model Anthropic could ship without a fight. It’s deliberately, explicitly capped on offensive cybersecurity capability — in Anthropic’s own published test, it never produced a working exploit against a real Firefox vulnerability, a result they frame as a design choice rather than a shortfall. Sonnet 5 is also the first Sonnet-tier model to ship with real-time cybersecurity safeguards built in. Anthropic isn’t hiding the comparison. It’s the safe, mid-tier release timed almost exactly to when the flagship it’s standing in for finally got its passport back.

What This Model Is Actually For

None of this makes Sonnet 5 a bad model. It makes it a badly-matched one if you’re using it the way you’d use any other Sonnet — as your everyday driver for quick questions and small edits. That’s where the extra thinking and extra tokens cost you the most and buy you the least.

Where it earns its keep is exactly the opposite scenario: long, loosely-scoped, multi-step work you’d rather hand off and check on later. Bug hunts across a messy codebase. A pipeline that touches three systems and needs to verify its own output at the end. Anything where “it kept going instead of stopping halfway” is worth more to you than “it answered fast.” Anthropic’s own prompting guidance for the model backs this up directly — it recommends front-loading well-specified, complete instructions in a single turn rather than clarifying progressively, because ambiguous prompts fed in gradually are exactly what triggers the token-hungry, over-checking behavior. Feed it a vague task in pieces and you’ll pay for every round of guessing. Feed it a complete brief up front, and the same instinct that inflates your bill on a one-line fix becomes the thing that gets a real multi-step job done without babysitting.

If you’re building agent pipelines where Sonnet 5 is one link in a chain rather than the whole conversation, this is a genuinely strong option, especially while the introductory pricing holds. If you’re chatting with it directly the way you used to chat with Sonnet 4.6, expect it to take a little longer and think a little harder about things that didn’t need thinking about.

Anthropic built a model that knows how to work without you watching. That’s the whole pitch, and it’s a real one. It just means the version of Sonnet 5 you actually want depends entirely on whether you’re the one giving it instructions, or whether something else is.