The pricing changed more than the model did
Anthropic shipped Claude Sonnet 5.5 on September 28, 2026, positioning it as a clear upgrade over Sonnet 5 that runs 30% faster and costs up to 30% less for most work. The published rate is $2 per million input tokens and $10 per million output tokens, with up to 90% savings from prompt caching and 50% from batch processing. There’s also US-only inference at 1.1x if your workload needs to stay domestic.
The headline performance claims matter, but for anyone running agents at volume, the interesting part is what this does to your model-tiering math. A mid-tier model that’s simultaneously faster and cheaper per task changes which workloads justify an Opus-class model at all.
Token efficiency is the compounding variable
The customer quotes Anthropic published cluster around one pattern: fewer steps, fewer tokens, same or better quality. Some reported numbers worth noting, all from the vendor’s own page:
- A finance workload suite showed roughly 121k tokens per answer versus 497k for Sonnet 5
- One team’s coding evals saw a third fewer tool calls and about half the shell runs per task
- Slack’s offline evals showed about 14% fewer output tokens with no prompt changes
If even half of that holds on your traffic, the effective discount compounds well beyond the sticker price cut. Tool calls and shell runs aren’t just token spend — they’re wall-clock latency in an agent loop, which is usually what users actually feel.
This echoes what I covered in GPT-6 Sol and Luna on Bedrock: the tiering decision is increasingly about matching a model’s cost profile to a workload class, not picking one flagship. Sonnet 5.5 strengthens the middle tier.
What to actually test before switching
Anthropic describes Sonnet 5.5 as a faster, lower-cost complement to Opus 5.5, aimed at well-scoped everyday work — coding, review, drafting, investigation. The thinking-effort control in the API is the lever to experiment with: low effort for cheap routine calls, higher for the ones that matter.
A practical migration test: take your three highest-volume agent flows, replay a few hundred real tasks through claude-sonnet-5-5, and compare output tokens and tool-call counts against your current model — not just pass/fail quality. That’s where the 30% claim either shows up or doesn’t.
One caveat: these are vendor-published figures and vendor-selected customer quotes. The per-customer numbers are directionally consistent with each other, which is a good sign, but they’re still self-reported. Run your own evals before rerouting production traffic.
The grounded takeaway: Sonnet 5.5 doesn’t introduce a new capability class — it reprices the middle of the stack. If your tiering map hasn’t changed since earlier in 2026, this is the moment to revisit which tasks you’re overpaying a frontier model to handle.
Sources
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.