If you ship bots for a living, the Opus 5.5 launch is a footnote and the Sonnet 5.5 announcement is the actual story.
Here is what Anthropic confirmed. Opus 5.5 landed on Tuesday, described by the company as “the strongest-performing model we’ve tested to date” on coding and knowledge work benchmarks. Alongside it came word that Sonnet 5.5 and Haiku 5.5 are coming “in the coming weeks” with similar improvements. No firm date for either. Anthropic also shipped Claude Opus 5.5 at $4 per million input tokens and $20 per million output, a 20% cut to list prices, with reported scores of 66.4% on Terminal-Bench 4.0 and 1846 Elo on GDPval-AA v2.1.
That is the whole verified set. Everything else floating around right now is inference, mine included, and I will label it as such.
Why the middle tier is the one that matters
Every bot I have put into production follows the same cost curve. The flagship model writes the hard parts, designs the tool schema, and untangles the one gnarly edge case that broke at 2am. Then it gets swapped out. What actually serves traffic — the classifier, the router, the summarizer, the retrieval reranker, the thing that answers the same twelve questions eight thousand times a day — runs on the mid tier.
That is not cost-cutting theater. It is architecture. A frontier model on a high-volume path is usually a sign that someone skipped the step where you decompose the problem. Once you break a bot into small, well-scoped calls, most of those calls do not need a frontier model. They need a fast one that follows instructions exactly and does not surprise you at the 99th percentile.
So when Anthropic says the Sonnet tier is getting the same treatment as Opus, that is the sentence worth circling. Improvements at the top move demos. Improvements in the middle move unit economics for anyone running real volume.
The price signal, read carefully
The only price numbers we have are for Opus 5.5: $4 in, $20 out, down 20% from the prior list. Anthropic has not published Sonnet 5.5 pricing, and I am not going to guess at a number and dress it up as reporting.
What I will say is that a 20% cut at the top of the lineup tells you something about the direction of travel. Companies do not usually trim the flagship and then hold the line below it. If the pattern holds, the mid tier gets cheaper too. If it does not, that is genuinely interesting news and worth writing about when the numbers land.
What to do in the waiting period
“Coming weeks” is not a date, which means you have a window. Use it to make the swap boring:
- Pin your model strings. If a model ID is hardcoded in six files, you do not have a migration path, you have a scavenger hunt. One constant, one place.
- Build the eval set now. Fifty to two hundred real examples from your logs, with expected outputs, scored automatically. Without this you are judging a new model on vibes and whichever three prompts you happen to remember.
- Log tokens per request path. You cannot evaluate a price change if you do not know which endpoint burns your budget. Tag by feature, not just by day.
- Measure latency at p95, not average. Faster models are sold on averages. Users notice tails.
- Re-test your tool calls specifically. In my experience, structured output and function calling behavior shift more between model versions than prose quality does. That is where migrations break.
None of this is exciting work. All of it turns a release announcement into a same-afternoon upgrade instead of a two-week project.
A note on the naming mess
Worth flagging honestly: the reporting around this cycle is inconsistent. One source dates Opus 5.5 to a Tuesday release, another to 22 September. There is also coverage of a Claude Sonnet 4.6 release in circulation.
For builders, the practical lesson is version discipline. When model families ship this fast and get covered this loosely, the name in a headline is not a reliable identifier. The exact API model string is. Record it in your eval results, your logs, and your incident notes. Six months from now, “we upgraded to 5.5” will mean nothing and the string will mean everything.
My read
Opus 5.5 is the model that wins the benchmark argument. Sonnet 5.5, whenever it lands, is the model that changes what you can afford to build. A cheaper, faster mid tier does not just shave your bill — it moves work back onto paths you previously had to hand-code around, because the model call was too expensive to justify.
That is the thing I am watching for. Not the leaderboard. The price sheet, and whether the tail latency holds up when I point real traffic at it.
🕒 Published: