Anthropic has released Claude Opus 5.5, cutting the standard API price to $4 per million input tokens and $20 per million output tokens while claiming performance comparable to Claude Fable 5.1 on most work.
The release was initially flagged through secondary reporting, but Anthropic has now confirmed it. Its Claude Opus 5.5 launch announcement says the model costs 40% less than Opus 5 on typical workloads, generates output more than 30% faster and is available across Anthropic’s platforms as well as Amazon Web Services, Google Cloud and Microsoft Azure.
The 40% figure needs some unpacking. Anthropic has cut the actual input and output token prices by 20% compared with Opus 5. The larger claimed saving comes from Opus 5.5 also using fewer tokens to finish a task. For developers, that makes cost per accepted task a more useful metric than token price alone.
Claude Opus 5.5 pricing is 20% lower per token than Opus 5
| Model | Input per 1M tokens | Output per 1M tokens | 100k input + 10k output example |
|---|---|---|---|
| Claude Opus 5.5 | $4 | $20 | $0.60 |
| Claude Opus 5 | $5 | $25 | $0.75 |
| Claude Fable 5.1 | $10 | $50 | $1.50 |
The worked example assumes identical token consumption. It therefore shows the direct pricing difference, not Anthropic’s claimed 40% reduction in real workload cost.
That difference is important for budgeting. Opus 5.5 is 20% cheaper than Opus 5 at the base token rates, but 60% cheaper than Fable 5.1. If Opus 5.5 can complete the same jobs with fewer tokens and fewer retries, the effective gap can widen further.
Prompt caching has also changed substantially. Cache reads cost $0.20 per million tokens on Opus 5.5, down from $0.50 on Opus 5. That can disproportionately affect coding agents and other long-running systems that repeatedly reuse large amounts of context.
There is a pricing trap for latency-sensitive deployments, though. Anthropic’s optional fast mode can deliver up to 2.5 times the speed, but raises the rate to $8 per million input tokens and $40 per million output tokens. Teams modelling production cost should therefore separate standard Opus 5.5 and fast-mode usage rather than using one blended headline rate.
Fable 5.1-level performance does not automatically make Fable redundant
Anthropic says Opus 5.5 performs at the level of Fable 5.1 on most work and reports higher Opus 5.5 results on several of its published coding and knowledge-work evaluations.
Its reported Terminal-Bench 4.0 result is 66.4% for Opus 5.5 against 55.8% for Fable 5.1, while GDPval-AA v2.1 is reported at 1846 for Opus 5.5 against 1735 for Fable 5.1. Anthropic also cautions that benchmark differences between frontier models are becoming less reliable indicators of real-world differences.
That’s what procurement teams should pay attention to. A model that leads a benchmark can still lose on your workload because of different tool behaviour, retries, context use, output length or failure modes. DIY AI’s AI model comparison is designed around those broader trade-offs, not treating one benchmark as a universal ranking.
For an existing Fable deployment, the sensible test is not simply “does Opus 5.5 score higher?” It is whether the cheaper model reaches the same acceptance threshold on your own difficult tasks. At a 60% lower base token price, the gap is large enough that even teams satisfied with Fable now have a strong reason to run that evaluation.
The biggest Opus 5.5 limitation may be invisible model routing
Opus 5.5 is not simply an unrestricted replacement for Opus 5. Anthropic has applied safeguards similar to those used with Fable 5.1 because the new model has stronger capabilities in areas such as cybersecurity and biology.
For many cybersecurity requests, Anthropic says the system can transparently reroute the task to Opus 4.8. Biology requests can also encounter additional safeguards. This creates an important evaluation problem: a developer may select Opus 5.5 but not necessarily have every part of a sensitive workflow executed by Opus 5.5.
Teams testing security, scientific or other restricted workflows should therefore record which requests are blocked or rerouted, not just whether the final API call returns successfully. Otherwise, a benchmark can appear to measure Opus 5.5 while actually measuring a mixed system.
Anthropic has also removed the option to run Opus 5.5 with thinking completely disabled. That may affect integrations built around tightly controlled latency or deterministic cost ceilings, even if the lower token rates improve average economics.
Claude Code users are watching session quality as closely as benchmark scores
Early discussion around the launch is clustering around a more practical set of complaints: whether Opus 5.5 follows instructions more consistently, produces less unnecessarily long prose and stays focused during extended coding sessions.
Those concerns line up with Anthropic’s own release messaging. The company says Opus 5.5 has been adjusted to put important information earlier, follow writing rules more closely and produce clearer output than Opus 5.
Subscription economics are separate from API pricing. Anthropic says it is increasing five-hour usage limits on Pro, Max, Team and seat-based Enterprise plans and providing users with a rate-limit reset that can be saved. The announcement does not reduce that change to a simple token allowance, so it would be premature to translate the API price reduction directly into 20% or 40% more Claude Code usage.
Our Claude Code pricing guide covers why subscription limits and metered API costs need to be assessed separately.
How DIY AI will test whether Opus 5.5 is actually cheaper work
DIY AI has not yet run an independent Opus 5.5 evaluation, so Anthropic’s benchmark and efficiency numbers should remain vendor-reported claims until they are reproduced on representative work.
A useful comparison needs to keep the job fixed while changing the model. For coding, that means using the same repository state, permissions, instructions and acceptance tests across Opus 5.5, Opus 5 and Fable 5.1. Tests should include a contained bug fix, a multi-file change, code review and at least one long-running task where context management becomes material.
The important measurements are not limited to pass rate. We would record elapsed time, input and output usage, tool calls, failed attempts, retries and the amount of developer correction required before the result can be accepted.
Cost per accepted result then becomes the useful number. A model charging less per token can still cost more if it needs repeated attempts. Conversely, a more expensive model can be economical if it reaches an acceptable answer with substantially less intervention.
Opus 5.5 changes the model choice even before independent testing is complete
For new Anthropic API deployments, include Opus 5.5 in the first round of evaluation rather than starting with Opus 5 by default. Existing Opus 5 users have a straightforward reason to regression-test it: the base rate has fallen, and Anthropic claims better performance at the same time.
The more interesting decision is for Fable 5.1 customers. Fable still costs $10 per million input tokens and $50 per million output tokens, while Opus 5.5 is $4 and $20. Anthropic’s own wording stops short of saying the models are interchangeable on every workload, but a 60% price difference is too large to ignore where Opus 5.5 clears the same quality bar.
The launch therefore changes the economics more clearly than it settles the performance question. Anthropic has made Opus substantially cheaper. The next job is to establish whether developers can keep the same acceptance rate while spending less money and less time getting there.
