Claude Opus 5.5’s September 22 launch offers two different explanations for a lower bill: reduced token prices and fewer tokens per task. Keeping those separate helps teams understand what they can budget from the price sheet and what they still need to assess in their own work. Anthropic’s announcement
The published input and output rates are $4 and $20 per million tokens, respectively, 20% below Opus 5. Cache reads fall from $0.50 to $0.20 per million tokens, a 60% reduction. Cache writes cost $5 per million tokens. Those are concrete listed rates. The broader claim that typical workloads cost 40% less at default settings comes from Anthropic’s tests and also depends on improved efficiency.
That distinction matters for long agent sessions, where cached context, generated output and repeated attempts contribute differently to the total. My inference is that a team with substantial cache reuse could see a different cost pattern from one producing lengthy uncached answers. The announcement alone cannot predict either bill.
Speed introduces another choice. Anthropic says standard output generation is more than 30% faster than Opus 5. A separate fast mode promises up to 2.5 times the speed, but its listed input and output rates are $8 and $40 per million tokens. A purchase decision therefore needs to include the value of lower latency, rather than assuming every mode carries the launch’s price reduction.
The release also increases five-hour usage limits on Pro, Max, Team and seat-based Enterprise plans and provides subscription users a rate-limit reset they can save. These subscription changes are distinct from metered token pricing. The page does not provide one universal allowance that applies to every plan.
Anthropic positions Opus 5.5 near Fable 5.1 for most work, while cautioning that benchmark margins have become less reliable guides to practical differences. Safeguards also affect some workloads and benchmark runs. The useful buying question is whether the model completes the intended work at an acceptable cost and latency—not whether one headline percentage applies to every session.