Anthropic's Opus 5 Rivals Fable 5 Performance at Half the Cost

Anthropic has rolled out Opus 5, a new model that challenges its own more expensive flagship offering, Fable 5, across most technical benchmarks—yet costs only half as much. The model arrives at $5/$25 per million input/output tokens, unchanged from Opus 4.8, and removes the requirement to opt into extended data retention, making it the default choice for Claude Max users.

Where Opus 5 Stands Against Fable 5#

Opus 5 actually outscores Fable 5 on many standard benchmarks, a trend that underscores how far the model family has progressed. The key differentiator remains sustained autonomy: Fable 5 retains the edge for projects demanding days-long independent operation and error recovery, while Opus 5 suits shorter, self-correcting tasks.

On knowledge work, Opus 5 achieves a score of 1861 on the GDPval-AA v2 benchmark, surpassing both Fable 5 (1747) and OpenAI's GPT-5.6 Sol (1736). It also ranks atop Zapier's AutomationBench—a test of AI agents on real business workflows—with a pass rate double that of its nearest competitor at equivalent cost. For coding, the model reaches within 0.5% of Fable 5's peak performance on CursorBench 3.2 while cutting the per-task expense in half.

On Humanity's Last Exam (with tools), Opus 5 edges out Fable 5 (64.7% vs. 63.9%), though Fable 5 recovers its lead on legal benchmarks and deep software engineering tasks.

Safety and Built-in Guardrails#

Anthropic describes Opus 5 as its most aligned model, grounded by automated behavioral audits and resistance to misuse exploitation. The company deliberately withheld cybersecurity training, yet Opus 5 still approaches Mythos 5's ability to identify vulnerabilities in open-source code—without mastering their exploitation, where Anthropic wants the gap to remain.

Narrower safety classifiers screen requests for targeted cyber tasks like exploit generation and penetration testing, firing about 85% less frequently than Fable 5's broader guardrails. When flagged, Claude.ai and the API fall back to Opus 4.8 by default; enterprises and researchers in Anthropic's Cyber Verification Program can request a less-restricted variant.

Developer Features and Rollout#

Opus 5 supports a "fast mode" delivering output 2.5x quicker at double the cost, available on the Claude Platform and in Claude Code via extra usage credits. Two platform updates launched in beta: automatic fallbacks now route declined requests to alternative models within the same exchange, and developers can toggle which tools Claude accesses mid-conversation without breaking the prompt cache.