Anthropic Launches Opus 5.5: Fable 5.1 Performance at a Lower Cost

Anthropic has made Claude Opus 5.5 available, claiming performance levels comparable to Fable 5.1 for most use cases, along with reduced pricing and quotas. The company also highlights enhanced safety measures and external evaluations, while businesses report measurable gains in coding and productivity.

Activatable Safeguards and Announced External Evaluations

Anthropic describes a safety mechanism for Opus 5.5 that combines in-depth alignment testing, pre-launch evaluations by external organizations, and safeguards dedicated to high-risk areas such as cybersecurity and biology. The company refers to positions expressed by its CEO Dario Amodei regarding the modulation of the pace of advancements in AI capabilities. According to Anthropic, Opus 5.5 is the most effective model they have tested to date, with improvements in behaviors associated with recent cybersecurity incidents, including biased reasoning or attempts to escape a controlled environment. The company states that it has utilized external testing providers and a class of safeguards similar to those of Fable 5.1 for cybersecurity, biology, and cutting-edge LLM development. When these safeguards are triggered, queries are automatically routed from Opus 5.5 to Opus 4.8. In practice, most cybersecurity tasks will be redirected to Opus 4.8, while requests related to biology and LLM development will be forwarded to Opus 5. Verified organizations can apply for the Life Sciences Verification Program for enhanced access in biological research. Those approved under the Cyber Verification Program will be able to start using Opus 5.5 in a few weeks.

Reduced Token Pricing and Subscription Quotas

Anthropic has reduced the price of tokens by 20% compared to Opus 5 and claims that Opus 5.5 consumes fewer tokens to produce higher-quality work. For subscribers, usage limits over five hours have increased by 20%. The company adds that the caps over five hours and weekly become more efficient since using Opus 5.5 costs less than using Opus 5. The $20 per month plans are affected by these changes.

Claimed Performance: Increased Speed and More Natural Style

Anthropic positions Opus 5.5 with performance levels close to Fable 5.1 for most tasks, at a cost approximately 40% lower. The model is available immediately, less than two months after Opus 5. The company also claims a safer, less costly, and less verbose model than its predecessors, stating that result generation is over 30% faster than with Opus 5. For reference, Opus 5 was introduced with performance levels similar to Fable at half the price. Anthropic further asserts that Opus 5.5 communicates more naturally than previous versions.

Business Feedback: Reduced Tokens and Faster Task Resolution

At Box, evaluations report that Opus 5.5 consumed one-third of the tokens used by Opus 5, with responses 40% less verbose without loss of accuracy. Box emphasizes that speed and cost are crucial for processing large amounts of content and believes these gains will be significant for teams managing agents in sectors like finance or public services. On the GitHub side, tests via Copilot CLI and VS Code indicate that Opus 5.5 is among the models using the fewest tokens and measured steps; in VS Code, it resolved more terminal tasks than Opus 5 in less than half the steps. These results are presented as making more ambitious projects more accessible for developers.

Output Quality and Bug Detection According to Ramp and Deloitte

At Ramp, a software engineer describes the end of unnecessarily verbose outputs and writing that conforms to their internal rules; a generated design specification was usable with very few edits, a prompt rewrite was preferred over the original, and optimizing a test suite allowed for a confident change deployment. Deloitte Consulting LLP reports that at the lowest effort setting, Opus 5.5 detected 72% of known bugs in code review, compared to 56% for Opus 5 at high effort, with fewer false positives and a fraction of the output. During a consulting analysis in the United States, a low-effort attempt matched higher settings with half the output and passed quality checks.

Next Steps: Sonnet 5.5 and Haiku 5.5 Announced

Anthropic plans to launch Claude Sonnet 5.5 and Haiku 5.5 in the coming weeks.