Claude Opus 5.5 Replaces Opus 5 and Opus 4.8
September 23, 2026 · InfraScout Team
Claude Opus 5.5 is now selectable on Anthropic connections, and it takes the place of both Opus 5 and Opus 4.8 in the model catalog. One million tokens of context, 128k of output, and adaptive thinking across every effort level from low to max.
Where it sits on cost
Opus 5.5 is cheaper than Opus 5 on every line of the bill:
| Claude Opus 5 | Claude Opus 5.5 | |
|---|---|---|
| Input, per million tokens | $5.00 | $4.00 |
| Output, per million tokens | $25.00 | $20.00 |
| Cache writes (5 minutes / 1 hour) | $6.25 / $10.00 | $5.00 / $8.00 |
| Cached input, per million tokens | $0.50 | $0.20 |
The last row matters most for assessment work. A long investigation re-reads its own accumulated tool output on every turn, so cached input is the majority of what a serious conversation spends, and Opus 5.5 charges less than half of what Opus 5 did for it.
That makes Opus 5.5 the natural first step above the balanced tier. It costs a fraction of Claude Fable 5.1, which stays at the top of the catalog for the investigations Opus 5.5 cannot crack. For quick questions, inventory lookups and routine checks, the faster and cheaper models are still the right answer — see Choosing a Model for how the tiers lay out.
Your existing connections
Adding a model to the catalog never moves a connection onto it, and taking one out never breaks a connection already using it. A connection on Opus 5 or Opus 4.8 keeps working and keeps being reported at exactly what it costs. The AI Providers view now marks it with an Outdated model badge pointing at Opus 5.5, and switching is a single edit to the connection's model.
Opus 4.8 used to stay in the picker as a fallback for security reviews that tripped Opus 5's stricter safety classifiers. The catalog no longer carries a dedicated fallback tier. If a legitimate review keeps getting declined, move it to a connection on a different model — and if you already registered an Opus 4.8 connection for exactly this, it stays available to you.
Opus 5.5 starts at a medium reasoning effort, its own documented default. It reasons more per effort level than Opus 5 did, so medium is not a step down in thoroughness.
Everything you rely on works with it
Tool search, the live task budget, long-run context compaction, Task Agents and agentic workflows all behave as they do on the rest of the catalog. We verify each model's capabilities against the live API when we add it, rather than inferring them from release notes, and Opus 5.5 was no exception.
Two smaller fixes came out of that verification, and you benefit from them whichever model you run:
- Conversation titles are complete again. On the models that always reason before answering, the short call that names a conversation could run out of room partway through, leaving a title cut off mid-word. Titles now come back whole, and a truncated one is never saved.
- Your data region pin now covers tool-group planning. Before each turn, a short call decides which tool groups the turn needs. It now honors the connection's data region like every other call the connection makes.
To use it, an administrator picks Claude Opus 5.5 on an Anthropic connection under AI Provider Connections, or edits an existing Opus 5 connection onto it.
Questions or feedback? Reach us at info@infrascout.cloud.