On this page
Yesterday's headlines all sounded alike: Anthropic launches Sonnet 5.5, 30% faster, up to 30% cheaper.
That second claim is easy to misread. The price list did not go down. Sonnet 5.5 still costs $2 per million input tokens and $10 per million output tokens, exactly the same as Sonnet 5 (Anthropic, SiliconANGLE).
So where does "cheaper" come from? The answer is the most useful lesson in this launch, and it applies even to businesses that never write a line of code.
What happened, by date
- Around June 2026: Sonnet 5 launched, roughly three months ago according to TechCrunch (TechCrunch).
- September 22, 2026: Anthropic released Opus 5.5, the top model of the 5.5 family (SiliconANGLE).
- September 28, 2026: Sonnet 5.5 launched, available on every major platform including Amazon Web Services, Google Cloud, and Microsoft Azure. Anthropic also said Haiku 5.5 is coming "in the coming weeks," with no specific date (Anthropic).
Just one week between the expensive model and the cheaper one. That cadence is deliberate, and we'll come back to it at the end.
What "30% cheaper" actually means
Anthropic says Sonnet 5.5 "costs up to 30% less for most work" compared with Sonnet 5 and runs more than 30% faster (Anthropic). The savings aren't in the unit price. They're in how many tokens and tool calls the model needs to finish the job: fewer steps, more actions bundled into each turn.
A few numbers Anthropic cites from early-access customers. To be clear, these are figures published by the vendor and its customers:
- Zendesk: tickets resolved 20% faster.
- Base44: Opus 5 needed an average of 7.7 iterations to complete a task; Sonnet 5.5 needed 3.6.
- Another customer reported 121K tokens per task versus 497K.
Here's the point I want you to take away: don't compare price per token, compare price per outcome. A model that costs 20% more per token but finishes in half the steps is still cheaper. Conversely, a "budget" model that flails for ten rounds before getting it right costs far more than it looks.
Independent data points the same way, but also urges caution. Artificial Analysis scores Sonnet 5.5 at 56 on its intelligence index, versus 38 for Sonnet 5 and 58 for Opus 5.5. The cost of running the same evaluation suite ranges from roughly $0.41 to $7.60 depending on the chosen "effort" level (Artificial Analysis). That page contradicts itself in its pricing table, so I'm only using the scores and that cost range, nothing else from it.
That spread of nearly 19x tells you something: with the same model, cost depends heavily on how you configure it. If you're buying AI services, ask your vendor "which effort level do you use, and why?" not just "which model do you use?"
The benchmark table: read it before you applaud
According to Anthropic's table, Sonnet 5.5 scores 70.6% on Terminal-Bench 4.0, beating even Opus 5.5 (66.4%), and reaches 1844 Elo on GDPval-AA v2.1, just two points behind Opus 5.5 (1846) (Anthropic).
Very impressive. But three things need to be read alongside it.
One: this is Anthropic's own table. TechCrunch notes plainly that there was no independent verification of these claims (TechCrunch). Independent measurements will arrive over the next few weeks; compare again then.
Two: where Opus still wins, the table says so. On FrontierCode 1.1, Sonnet 5.5 scores 46.2% while Opus 5.5 scores 54.4% at the Xhigh setting. Anthropic's footnote even admits Sonnet 5.5 scored lower at Max than at Xhigh because some code-review tasks timed out (Anthropic). I respect that they disclosed it. But it shows that "98% of Opus's score" is an average; the hardest problems still belong to the expensive model.
Three: when a jump is too big, be suspicious first. In the same table, Sonnet 5 scores 10.3% on Terminal-Bench and Sonnet 5.5 scores 70.6%. A sevenfold leap between two generations three months apart most likely reflects a change in how the benchmark is run or how tools are used, not a model that is purely "seven times smarter." Indeed, Artificial Analysis's composite intelligence index went from 38 to 56: a big step, but nowhere near seven times. When a number jumps, ask what it measures before you ask how good it is.
What few people noticed: deliberate downgrading
This is the detail I find most worth discussing, and it sits in the safety section, not the performance section.
Sonnet 5.5 is the first Sonnet model to ship with cybersecurity safeguards similar to Opus 5.5. Anthropic writes that higher-risk cybersecurity tasks will "visibly fall back to Sonnet 5," meaning requests are automatically routed to the older model, and the user can see that happen (Anthropic). It's also the first Sonnet with a classifier that prevents extraction of its reasoning chain.
The business implication is concrete. If you build a tool on Sonnet 5.5 and some requests cross a sensitivity threshold, output quality may differ from what you tested. Anthropic also acknowledged a pre-launch deployment bug that affected structured output, since fixed. If your product depends on a strict output format, rerun your test suite rather than assuming "the new version" must be better.
Three things to do after this news
- Calculate cost per result. Pick 20 real tasks from your business, run them on both models, and measure the total cost and the number of manual fixes. That's the number you actually pay.
- Don't default to the most expensive model. For well-defined, narrow work such as bug fixes, drafting documents, or building spreadsheets, a mid-tier model is now good enough. Anthropic itself describes Sonnet 5.5 as strongest at exactly these jobs. Keep the premium model for the genuinely hard problems.
- Ask your vendor about configuration. Effort level, token limits, automatic model fallback: all of it shows up on the monthly bill.
What it signals
One week between Opus 5.5 and Sonnet 5.5, with Haiku 5.5 on the way. The labs are closing the gap between price tiers. When a mid-tier model scores nearly as well as the flagship, an AI product's competitive edge no longer comes from "using the strongest model." It comes from how you use the model: the right job, the right configuration, with measurement.
If you're considering bringing AI into your website or sales process, that's the part I usually work on with clients. See AI integration services or get in touch to talk specifics.
