Anthropic releases Opus 5.5 with reduced pricing and updated safeguards

Anthropic has released Opus 5.5, the most capable tier in its Claude artificial intelligence lineup.

2 min read

Claude Opus 5.5 promotional graphic with title text.
Promotional graphic for Claude Opus 5.5 showing the model title over a horizon aesthetic.

Anthropic has released its newest artificial intelligence model, Opus 5.5, which sets a new performance level in coding and knowledge work. Opus is the highest tier in Anthropic's three-tier Claude lineup, sitting above the middle tier, Sonnet, and the fastest tier, Haiku.

Key facts

Input and output tokens
$4 and $20 per million
Cache reads
$0.20 per million tokens
Fast mode input tokens
$8 per million
Fast mode output tokens
$40 per million
FrontierCode score
54.6%
CursorBench score
52.5%
Coding Benchmark and Performance Comparison
FrontierCode
Claude Opus 5.5: 54.6%
Claude Opus 5: —
GPT-6 Astra: 53.3%
CursorBench
Claude Opus 5.5: 52.5%
Claude Opus 5: 46.6%
GPT-6 Astra: —
Terminal-Bench 4.0
Claude Opus 5.5: 52.3%
Claude Opus 5: 51.8%
GPT-6 Astra: —

Performance and Pricing

The new model outperforms the larger Fable model on multiple benchmarks and successfully completed several informal tasks where Fable failed. Alongside these capability gains, Opus 5.5 is more affordable than its predecessor. Output tokens cost $20 per million tokens, down from $25 in the previous version, with similar price reductions across other usage metrics. The system also runs faster due to a reduction in the computing power needed to serve it.

Communication styles have also been adjusted. Opus 5.5 tends to use less specialized jargon and places important information earlier in its messages.

Timeline and Future Releases

The launch follows the release of Opus 5 just two months prior. Anthropic announced that Sonnet 5.5 and Haiku 5.5 will follow in the coming weeks with comparable improvements.

Safety and Governance

Opus 5.5 matches the biology and cybersecurity capabilities of Mythos. Because of these advanced capabilities, the release follows the same safety protocols used for the Fable model. These rules restrict usage related to discovering software vulnerabilities and developing biological agents.

The release arrives after CEO Dario Amodei publicly supported pacing the development of artificial intelligence capabilities to allow safety research and policy to keep pace. Safety training for Opus 5.5 included alignment testing and pre-release evaluations by independent organizations such as METR and Frontier Design.

Source: Original announcement

Worth a read?

Report

Comments