On Terminal-Bench 4.0, the benchmark that stands in for “can this model actually finish an agentic coding job,” Anthropic’s new mid-tier Claude Sonnet 5.5 scores 70.6 percent. Its own flagship, Claude Opus 5.5, released six days earlier, scores 66.4 percent. The cheaper model beat the expensive one on the test that sells coding models, and it did so at half the sticker price. That single row is more useful for deciding where Sonnet 5.5 belongs in your stack than anything in the launch copy.
Anthropic shipped Sonnet 5.5 on September 28, calling it a faster, lower-cost complement to Opus 5.5, the flagship it repriced against GPT-6 Sol a week earlier. The headline everyone ran was “cheaper and faster Sonnet,” which is true and also the least interesting part. The interesting part is that the workhorse now lands close enough to the flagship, and ahead of it on some agentic work, that the reason to reach for Opus 5.5 has narrowed to a specific, nameable set of jobs. For everything else, the decision has moved off the price tier entirely and onto a dial most teams never touch on purpose.
The sticker did not move; the bill did
Sonnet 5.5 costs exactly what Sonnet 5 cost: $2 per million input tokens, $10 per million output, $0.20 per million cache reads and $2.50 per million cache writes. Opus 5.5 sits at $4 and $20. So the per-token price is half the flagship’s and identical to the model it replaces. Nothing in the price list changed.
What changed is how many tokens a real job burns. Anthropic’s pitch is that Sonnet 5.5 costs up to 30 percent less per completed task than Sonnet 5, not because the meter dropped but because it finishes with fewer tokens and fewer tool calls. The customer numbers Anthropic published make the mechanism concrete. Balyasny Asset Management reported Sonnet 5.5 answering with about 121,000 tokens per query where Sonnet 5 used 497,000, roughly a quarter of the token spend for a better score. Lovable measured a third fewer tool calls and about half the shell runs to finish the same coding task. Box logged 12 percent fewer total tokens at 2.4 times the speed. Slack saw about 14 percent fewer output tokens with no prompt changes at all.
This is the same lesson the site has been repeating since GPT-6 Astra made cost-per-completed-task the only honest unit: the price per million tokens tells you almost nothing until you multiply by how many tokens a job actually consumes and how many times the agent has to loop. A model that is nominally the same price but does the work in a third of the tokens is, on your invoice, a different and cheaper model. The sticker is a distraction. The token count is the bill.
What the flagship is still for
If the mid-tier model matches or beats Opus 5.5 on agentic coding at half the price, the fair question is why anyone would route to Opus at all. The benchmark table Anthropic published, reproduced across the launch coverage, answers it, and the answer is narrow but real. These are Anthropic’s own reported scores, not an independent lab’s, so read them as the vendor’s characterization of its own models:
| Benchmark (Anthropic’s own testing) | Sonnet 5.5 | Opus 5.5 | Sonnet 5 |
|---|---|---|---|
| Terminal-Bench 4.0 (agentic coding) | 70.6% | 66.4% | 10.3% |
| CursorBench 4.0 | 55.5% | 57.8% | 34.1% |
| FrontierCode 1.1 (Max effort) | 46.2% | 54.4% | 42.4% |
| GDPval-AA v2.1 | 1844 | 1846 | 1449 |
| AA-Briefcase v1.1 | 1811 | 1822 | 1359 |
The pattern is clear once you stop reading it as a horse race. On broad long-horizon knowledge work (GDPval-AA, AA-Briefcase) the two models are within a point of each other, which means for most office-document, analysis and briefing work you are paying double for a rounding error. On agentic coding that runs many small steps, Sonnet 5.5 edges ahead, probably because it can afford to spawn more agents inside the same budget. Opus 5.5 keeps a genuine lead only on the hardest single-shot reasoning, the FrontierCode Max column where careful judgment on one difficult problem beats speed across many easy ones, and it holds a slim margin on CursorBench. Opus is now the model you reach for when a single decision is hard and expensive to get wrong, not the default you run because it is the best. That is a much smaller job than “flagship” implies, and it is the correct way to read the six-day gap between the two releases: Anthropic shipped the expensive model first, then shipped the one most people should actually run.
Effort became the tier
Here is the shift the launch language buries. With Sonnet 5.5, the variable that moves your cost the most is no longer which model you pick. It is which effort level you run the model at.
Sonnet 5.5 exposes five reasoning settings: low, medium, high, xhigh and max. Anthropic’s own framing is that at Medium effort, the default in the Claude apps, the model beats Sonnet 5’s best score for less than a tenth of the cost per task, and at Low effort it still clears Sonnet 5’s best for under a tenth. At High effort, the default on the Claude Platform, Anthropic claims it matches GPT-6 Sol’s best score for about a fifth of the cost per task. Read those together and the effort dial spans something like a tenfold range in cost per task while the model underneath never changes. That is a far bigger lever than the choice between Sonnet and Opus, and almost nobody sets it deliberately.
Worse, the two defaults disagree with each other. The Claude apps and Claude Code default to Medium; the Claude Platform defaults to High. So the same model, called two ways, bills you at two different effort levels out of the box, and the API path is the more expensive one. A team that assumes “we standardized on Sonnet 5.5” has standardized on the model, not on the setting that determines the invoice. This is the same trap covered when the reasoning-effort settings first became a first-class cost control: the dial is where the money is, and leaving it on a vendor default is a decision you did not make.
Choosing on your own workload, not the launch table
The practitioner move is not “switch everything to Sonnet 5.5.” It is to re-derive three defaults, and to do it against your own traffic rather than Anthropic’s benchmark suite.
First, set the effort level before you set anything else. For high-volume, well-scoped work (bug fixes, ticket triage, first-draft documents, the everyday load) start at Low or Medium and only step up if your own evals show the cheaper setting failing. Twenty years of running IT operations taught me that the line item nobody set on purpose is the one that quietly grows, and a per-token bill on an API defaulting to High is exactly that kind of line item. Instrument cost per completed task, not tokens per call, and watch what the effort setting does to it.
Second, demote Opus 5.5 to a routed exception. If the broad-knowledge scores are a wash and Sonnet edges the flagship on agentic coding, then Opus earns its double price only on the hard single-shot reasoning where a wrong answer is expensive. Route to it deliberately for those, not by habit.
Third, verify the cost-per-task claim on your workload before you trust it. Every number above is Anthropic grading Anthropic, on benchmarks Anthropic selected. The token-efficiency gains are plausible and the customer figures are specific, but “up to 30 percent less per task” is a ceiling, not a guarantee, and it depends entirely on your prompts, your tools and how often your agents loop. Run 50 to 200 of your real tasks through Sonnet 5.5 at two or three effort levels, score them the way you actually judge output, and compare cost per completed task against whatever you run today. That is the same discipline that applies to every launch: the leaderboard is a starting hypothesis, not the answer for your traffic.
Two more facts worth logging. Sonnet 5.5 carries the 1-million-token context window standard across the current Claude line and is available with zero data retention on AWS, Google Cloud and Azure under the identifier claude-sonnet-5-5. And it is the first Sonnet to launch with the cyber safeguards and fallbacks Anthropic reserves for its most capable models, which matters if you are routing it into anything security-adjacent.
The launch sold a cheaper, faster Sonnet. What actually shipped is a model that erases most of the reason to pay flagship rates and quietly relocates your biggest cost lever to a dial with two conflicting defaults. The teams that come out ahead this week are the ones that read past “cheaper and faster,” set the effort level on purpose, and priced the second number.