Anthropic released Claude Sonnet 5.5 as the second model in the Claude 5.5 family, one week after Opus 5.5. The pitch is not a new intelligence class. It is a mid-tier model that finishes well-scoped work faster and with fewer tokens. List price is unchanged: $2 per million input tokens, $10 per million output, $0.20 per million cache reads. Anthropic says typical tasks cost up to 30% less than Sonnet 5 because the model uses fewer tokens to get to a usable answer, and that it generates output more than 30% faster. That is vendor workload math, not a price cut.
What shipped
Sonnet 5.5 is live on Claude.ai (web, iOS, Android), the Claude API as claude-sonnet-5-5, Amazon Bedrock, Google Cloud, and Microsoft Foundry. Context window stays at 1 million tokens. Max output is 128K. Knowledge cutoff is June 2026. Adaptive thinking is on by default. API users can still set effort. Claude apps default to Medium effort. The Claude Platform defaults to High.
Anthropic positions it as the complement to Opus 5.5. Opus is for work that needs careful judgment. Sonnet 5.5 is for well-scoped everyday jobs: bugs, features small enough to review, polished one-pagers, slides that follow a template, spreadsheets, and design passes that need less cleanup. Haiku 5.5 is promised in the coming weeks for high-volume, cheap work.
On Anthropic’s own numbers, Sonnet 5.5 scores 70.6% on Terminal-Bench 4.0 against 10.3% for Sonnet 5 and 66.4% for Opus 5.5 at the highest effort setting. That jump is large enough to notice even if you distrust vendor benches. Independent scoring from Artificial Analysis puts Sonnet 5.5 near Opus 5.5 on several knowledge-work and coding tables at max effort, with a caveat: at that setting it burns a lot of output tokens. The “30% cheaper” claim and the “heavy at max effort” measurement can both be true. Effort setting is the dial.
Two other details matter if you run agents. Sonnet 5.5 is the first Sonnet to ship with the cyber safeguards and fallbacks Anthropic reserved for its strongest models. High-risk cyber requests can fall back to Sonnet 5. Developers moving from Sonnet 5 should read the migration notes. Preserved thinking is tighter, and turning thinking off between tool calls is not a silent drop-in.
Why a small team cares
Most agency spend is not Opus-class work. It is a client email that needs a reply in brand voice, a WordPress fix, a deck for Thursday, a support ticket draft, a first-pass audit of a messy repo. That is Sonnet work. If the mid-tier model is faster and cheaper on those jobs, the win shows up as more finished artifacts per afternoon, not as a blog post about AGI.
If I were running a five-person shop that already bills Claude usage, I would change the default today and keep Opus or Fable for architecture and judgment calls. Do not pay flagship rates for a bug that fits in one file. Same unit price plus fewer tokens is a margin change if you bill it out, and hours back if you eat the cost. Leaving Sonnet 5 selected after today is leftover inertia.
Hype vs useful
The hype is that Sonnet just beat the flagship and you should rewrite your stack. Useful is: change the default in Claude and in your API router, run last week’s real jobs against 5.5, and keep Opus for the 10% of work that actually needs it.
Do not treat Terminal-Bench 70.6% as a promise it will rewrite a brownfield monolith unsupervised. Review the diff. The design claim is worth a test on one real slide deck, not a new “AI designer” product line.