Anthropic releases Claude Opus 5, matching Fable 5 capabilities at half the price

Decision brief

What changed

Anthropic released Claude Opus 5, priced at $5/$25 per million input/output tokens, claiming performance close to or matching its higher-priced Claude Fable 5 model on many benchmarks (e.g., GDPval-AA v2, SWE-ECI, Epoch Capabilities Index) while costing roughly half to a third as much. Opus 5 becomes the default model for Claude Max and Claude Pro subscribers, with reported improvements in coding, prompt-injection resistance, and reduced safety-classifier false positives.

Why it matters

For enterprises running AI-driven coding, agentic automation, or customer-facing workflows, Opus 5 offers a way to cut inference costs roughly in half without a major capability trade-off versus the flagship Fable 5, directly affecting AI infrastructure budgeting and vendor selection. The release also follows regulatory scrutiny and export controls on Fable 5, suggesting Anthropic is hedging model strategy amid compliance pressure, which matters for teams assessing vendor stability and long-term roadmap risk. Additionally, longer autonomous agent runs enabled by lower cost introduce new operational risks (token spend control, task-stopping failures) that security and platform teams must address before scaling deployment.

Affected roles

CFO CTO CISO COO

Evidence

The pricing, benchmark comparisons, and default-model status are consistently reported across Anthropic's own announcement, The Verge, TechCrunch, The New Stack, AWS, MarkTechPost, and Simon Willison, giving high confidence in the core facts. Safety and security claims (prompt injection resistance, reduced false positives) are corroborated by Anthropic's system card as cited by Zvi and Simon Willison, though these remain company-reported metrics rather than independently verified.

What remains uncertain

Benchmark comparisons rely heavily on Anthropic's own reporting and select third-party leaderboards (e.g., Epoch Capabilities Index showing Opus 5 at 159 vs Fable 5's 161), and some outlets note public benchmarks may understate or overstate real-world performance; independent verification of safety claims (prompt injection, self-harm response rates) is limited to Anthropic's system card. It's also unclear how the removal of certain cybersecurity training and safety policies in Opus 5 will affect risk profile in production use at scale.

Monitor next

Watch for independent third-party benchmark validation and early enterprise deployment reports (e.g., via AWS Bedrock or Claude Code integrations) that confirm whether Opus 5's cost-performance trade-off holds up under real-world agentic workloads and security stress-testing.

Analytical support, not advice — assumptions and open questions stated above.

Source coverage