The most consequential number in Anthropic's Claude Opus 5.5 launch is not the one most coverage led with. Yes, input and output tokens dropped 20% to $4 and $20 per million. But cache reads fell 60%, to $0.20 per million β and for the long-running agent sessions this model is built for, that is the line item that decides the invoice.
Key takeaways
- Claude Opus 5.5 is priced at $4 per million input tokens and $20 per million output tokens, with cache reads at $0.20 per million β 60% below Opus 5.
- Anthropic deploys the model with production safeguards that can hand cybersecurity tasks to Claude Opus 4.8 and biology or frontier-model-development tasks to Claude Opus 5.
- Opus 5.5 posts 66.4% on Terminal-Bench 4.0 and 1846 Elo on GDPval-AA v2.1, ahead of both Claude Fable 5.1 and OpenAI's GPT-6 Astra.
Why cache pricing decides agent economics
An agent working a codebase overnight does not send fresh context each turn. It resends the same system prompt, the same tool definitions, the same files, over and over. Those replayed prefixes are cache reads, and they dominate the meter on any session measured in hours rather than seconds.
Anthropic's own framing acknowledges as much, describing cache reads as the majority of what agentic and coding work costs. So a 60% cut there does more for an overnight migration than a 20% cut on fresh input ever could. Teams that budgeted Opus 5 by output tokens alone have been modelling the wrong variable.
Throughput moved in the same direction. Output generation is more than 30% faster than Opus 5. A separate Fast mode inside Claude Code and the Claude Platform trades price for latency at $8 and $40 per million tokens, promising up to 2.5x speed. Five-hour usage caps rose on Pro, Max, Team and seat-based Enterprise plans, and subscribers can now bank a rate-limit reset and spend it when a deadline demands it.
When your request stops being Opus 5.5
Buyers should read one footnote twice. Because Opus 5.5 lands near Claude Mythos 5.1 on biology and cybersecurity capability, Anthropic ships it behind the safeguards it built for Fable 5.1. When those safeguards fire, the request does not fail β it gets finished by a different model.
Cybersecurity tasks were completed by Claude Opus 4.8, and biology and frontier LLM development tasks were completed by Claude Opus 5.
That is disclosed rather than hidden, and it is a defensible safety design. It is also a silent capability change inside a product you are paying frontier rates for. Anyone building security tooling on the API should assume a fraction of their traffic is answered by a model two releases old. The practical question is not whether the fallback is justified. It is whether your evaluation harness can tell you when it fired, because the response arrives looking like any other.
The mechanism distorts scoreboards too. Zapier ran the AutomationBench numbers without a fallback configured, so every safeguard intervention scored as a failure β which is how Opus 5.5 ended at 40.0% against GPT-6 Astra's 41.4%. Anthropic argues production deployments would not lose those points. OpenAI's competing chart shows the same machinery from the other side: a Fable 5.1 configuration with Opus 5 fallback reached 31.4%, with OpenAI conceding the cost figure excluded fallbacks that fired on roughly 40% of tasks.
What testers actually measured
Most of the evidence Anthropic published is about volume, not cleverness. One tester closed a 680,000-line code migration inside a day. Another audited and repaired a 200,000-line codebase in under three hours; Opus 5 needed more than 20 hours and 2.5 times the tokens. An internal HAProxy port from C to Rust finished in 9.5 hours against 12 for Fable 5.1, at 51% lower cost.
None of those are intelligence claims. They are throughput claims, and throughput is what turns a model upgrade into a budget line. Customer numbers repeat the pattern: fewer steps, less output, same or better result. Deloitte found the model at its lowest effort setting caught 72% of known bugs in code review, where Opus 5 at high effort managed 56%. Hebbia scored it at 86.6% coverage on expert-graded finance workflows against 60.3% for Opus 5. On public benchmarks, Opus 5.5 records 66.4% on Terminal-Bench 4.0 and 1846 Elo on GDPval-AA v2.1, where Fable 5.1 sits at 1735 and GPT-6 Astra at 1542.
A cadence its own CEO called too fast
The calendar is the uncomfortable part. The Register counted 21 days between Fable 5.1 and Opus 5.5, and 39 between Opus 5 and Fable 5.1 β with CEO Dario Amodei's public call to slow capability gains landing ten days before this launch. Anthropic shipped on the same morning OpenAI released GPT-6 Sol and Luna, three weeks after GPT-6 Astra reached enterprise customers. Full details sit in Anthropic's launch post. Sonnet 5.5 and Haiku 5.5 follow in the coming weeks, which is when the cache-read discount will matter to the largest number of workloads.
FAQ
How much does Claude Opus 5.5 cost?
Opus 5.5 costs $4 per million input tokens and $20 per million output tokens, 20% below Opus 5. Cache reads are $0.20 per million, a 60% reduction. A separate Fast mode in Claude Code and the Claude Platform runs $8 and $40 per million tokens for up to 2.5x speed.
Can Anthropic route my Opus 5.5 request to an older model?
Yes, in narrow cases. Anthropic states that when Opus 5.5's production safeguards intervene, cybersecurity tasks are completed by Claude Opus 4.8 and biology or frontier-model-development tasks by Claude Opus 5. Vetted organisations can apply to Anthropic's Life Sciences Verification Program, with a Cyber Verification Program expanding in the coming weeks.
Is Opus 5.5 better than GPT-6 Astra?
It depends on the task. Opus 5.5 leads on Terminal-Bench 4.0 (66.4% to 57.9%) and GDPval-AA v2.1 (1846 Elo to 1542), while Astra leads on AutomationBench (41.4% to 40.0%) and Terminal-Bench-Science 0.1 (64.6% to 58.7%). The labs ran these evaluations under different harnesses and effort settings, so the margins are not directly comparable.
