Anthropic released Claude Sonnet 5.5 today, September 28, 2026. PressBot 1.12.0 ships the same day. claude-sonnet-5-5 is already selectable in the visitor chatbot, the Pro admin agent, Shield, and the blog content pipeline, and it’s the new default for the admin agent.
Same price, a much better agent
Sonnet 5.5 costs exactly what Sonnet 5 did: $2 per million input tokens and $10 per million output tokens. It keeps the 1M-token context window.
What changed is the work it can do. Anthropic’s own benchmark numbers show the jump is largest on agent-style tasks, where a model has to use tools over several steps:
- Terminal-Bench 4.0 (agentic coding in a terminal): 70.6%, up from 10.3% for Sonnet 5. Opus 5.5 scores 66.4%.
- OSWorld 2.1 (operating a computer to finish tasks): 80.1%, up from 57.0%, close to Opus 5.5’s 81.8%.
- GDPval-AA (real knowledge-work tasks): 1844, level with Opus 5.5’s 1846.
Anthropic also says it answers more than 30% faster and uses fewer tokens per task, so a job can cost less than it did on Sonnet 5 even at the same rates. Those are their numbers, not ours. Run your own workload before you believe any of them.
Why it’s the new admin agent default
The admin agent is a tool-using job. “Find the posts with no meta description, draft new ones, and put them in a plan for me to review” is five or six tool calls in a row, and that is the kind of work Sonnet 5.5 improved most. At Sonnet 5’s price, keeping Sonnet 5 as the default would mean leaving the better model on the shelf.
So from 1.12.0, new installs and sites that never picked a model get Sonnet 5.5 for the admin agent. The same goes for the blog pipeline’s Writer and Critic, where following a brief and judging a draft both benefit from a stronger model.
Your saved pins stay put
If you chose a model yourself, PressBot doesn’t touch it. Sonnet 5 stays selectable, labeled previous. Switching models changes behavior, speed, and spend, and you should make that change on purpose. When you’re ready, open the admin agent’s model selector and pick Claude Sonnet 5.5.
The visitor chatbot default stays Claude Haiku 4.5. Sonnet 5.5 is available there, unlike Opus 5.5, because its price is reasonable for public traffic. But at twice Haiku’s rates, it only earns its place if your visitors ask questions that need real reasoning, such as comparing products or working through a multi-step support problem.
Why launch-day support took real code
Adding a new model is usually one line: a new ID in a list. Sonnet 5.5 took more, because it changes how the API behaves.
Claude 5 models think before they answer by default, and that thinking counts against the reply’s token budget. PressBot turns thinking off for visitor chat and for short pipeline steps, so a quick answer stays quick and a 256-token classification step doesn’t spend its whole budget reasoning. On Sonnet 5 that setting was disabled. Sonnet 5.5 rejects disabled with an error. Its lowest setting is a new one, between_tools.
An integration that only swapped the model ID would have failed on every visitor message. PressBot 1.12.0 sends between_tools to Sonnet 5.5 and keeps disabled for the models that still accept it. We tested visitor chat with site answers and the admin agent’s tool loops against the live API before shipping.
One change you may notice: between tool calls, the admin agent now shows fewer “working on it” notes. Sonnet 5.5 returns those notes in a different format. The final answer isn’t affected.
How we’d pick for a real session
- Everyday admin work: Sonnet 5.5. It’s the default for a reason.
- A wrong call would cost money or a client: Opus 5.5 for security questions, multi-step restructures, or anything that touches a live client site. It’s close to Sonnet 5.5 on several of the benchmarks above, so test before you pay double.
- The hardest long jobs: Fable 5.1, after Opus 5.5 has fallen short.
- The visitor widget: Haiku 4.5, unless your visitors’ questions need more.
Update and try it
PressBot 1.12.0 is available now. Pro plans are still $29, $59, and $99 per year, and Sonnet 5.5 runs on your own Anthropic key, so model spend goes to your Anthropic account, not to us.
If you already run Pro, update to 1.12.0, pick Claude Sonnet 5.5 in the admin agent, and give it a real task you’d normally send to Sonnet 5. Compare the result and your Anthropic usage page. Full notes are in the changelog. New to PressBot? Start with the free chatbot or see what Pro adds.
