- Anthropic cut Claude Opus 5.5 API pricing to $4 per million input tokens and $20 per million output tokens, with cache reads down 60% to $0.20 per million, targeting high-frequency agentic workloads.
- Safety classifiers can reroute Opus 5.5 agent requests to Opus 4.8 or Opus 5, disclosed via the API response’s model field and a fallback content block, but easy to miss if evals or logging don’t check for it.
- Rerouting can still cause SLA and cost surprises for teams whose monitoring doesn’t check the response’s model field, since a call logged against the Opus 5.5 endpoint may have actually been served by an older model.
Buried in Anthropic’s Opus 5.5 release documentation is a clause that should stop enterprise architects before they commit workflows: safety classifiers can redirect individual agent requests to older models, a switch Anthropic discloses in the API response itself but that’s easy to overlook if your tooling isn’t checking for it. The 20% price cut is real, but the routing behaviour means model provenance now depends on whether your integration checks the response’s model field, not just the endpoint you called.
How the Routing Works
The pricing structure sets input tokens at $4 per million and output tokens at $20 per million, down from $5 and $25 under Opus 5. Cache read pricing drops 60% to $0.20 per million tokens, a reduction aimed at high-frequency agentic workloads. Anthropic says Opus 5.5 generates output more than 30% faster than its predecessor and requires fewer tokens per task, which the company says combines with the lower rates to produce roughly 40% savings on typical workloads, though those figures come from Anthropic’s own estimates rather than independent testing.
The routing caveat sits in the release documentation. When a request in a multi-turn agent workflow triggers specific safety protocols, Anthropic’s internal classifiers can redirect it to Opus 4.8 or Opus 5. The switch is disclosed, the response names the model that actually served it, and a fallback content block records the handoff, but nothing forces an integration to read either signal. The triggers cited cover domains including cybersecurity and biology. A workflow built and tuned against Opus 5.5’s output characteristics may unknowingly run part of its task graph on a different model, with different reasoning behaviour and different cost dynamics. Rerouting to an older model also introduces variance into projected cost savings for any workflow segment that crosses a safety threshold.
Audit Trails and SLA Exposure
For enterprise deployments running agentic workflows across financial analysis, compliance review or code generation, the governance exposure is specific. Internal AI policies and client SLAs typically specify which model handles which data. Enforcing that specification requires checking the response’s model field on every call, since a request logged against the Opus 5.5 endpoint can be served by an older model. Audit logging built only around the endpoint name, rather than the field Anthropic actually returns, will miss the substitution.
Anthropic has previously documented that specific model IDs map to fixed model snapshots, framing this as a stability guarantee. The safety classifier layer sits outside that guarantee. Model weights for a given ID remain fixed; the serving infrastructure around them, including classifiers and request routers, can change and alter observable output. That distinction matters for teams who have tuned prompts and structured workflows against specific model behaviour. A prompt optimised for Opus 5.5’s output format may produce degraded or malformed results when the same request is handled by Opus 5 or Opus 4.8.
Benchmarks and the Transparency Gap
On benchmarks including Terminal-Bench 4.0, Anthropic claims Opus 5.5 outperforms rivals on specific tasks, though those comparisons come from Anthropic’s own published results. The prior Opus 4.5 release posted an 80.9% score on SWE-bench, giving context for the trajectory Anthropic is pursuing on coding-heavy agent tasks.
The release follows Anthropic CEO Dario Amodei’s public call for the AI industry to prioritise safety and commit to external evaluation. The rerouting mechanism is consistent with that position, and so is disclosing it through the API response, though the disclosure is easy for teams to miss if they aren’t specifically checking for it. Enterprises evaluating Opus 5.5 for mission-critical deployments should treat routing as a procurement variable, not a footnote, and press Anthropic for aggregate fallback-rate data across safety categories before committing workflows that depend on consistent model behaviour.
