Anthropic has launched Claude Opus 5, a new flagship large language model available today that the company says delivers substantially better capability and cost-efficiency than the prior Opus generation. The model is positioned as the new default on Claude Max and the strongest option on Claude Pro, and Anthropic has opened API access with the same public token pricing as Opus 4.8.
Confirmed: Claude Opus 5 release, availability and pricing
Anthropic states that Claude Opus 5 is available immediately across its platforms and via the Claude API. Public pricing is unchanged from Opus 4.8: $5 per million input tokens and $25 per million output tokens. Opus 5 is offered in a Fast mode that runs about 2.5× the default speed at twice Opus 5’s base price. Anthropic also made two features available in beta: mid-conversation tool changes on the Claude Platform and API automatic fallbacks that reroute flagged requests to another model.
How Opus 5 works and operational changes
Opus 5 introduces an adjustable “effort” setting that lets users trade token cost and latency for higher reasoning effort; Anthropic shows how performance scales against cost on internal effort plots. For safety and operational control, Opus 5 uses updated cyber classifiers that are less restrictive than those on Fable 5: they allow vulnerability finding in source code but block binary-based scanning, penetration testing and exploit generation. Anthropic expects the classifiers to intervene roughly 85% less often than Fable 5.
When requests are flagged in Anthropic’s hosted products (Claude.ai, Claude Code, Claude Cowork), Opus 5 will fall back to Opus 4.8 by default. API customers can enable automatic fallbacks so flagged requests route to the next available model rather than failing. For enterprises and researchers requiring fewer restrictions, Anthropic’s Cyber Verification Program (CVP) provides access to a version of Opus 5 with reduced security constraints.
Why it matters
Anthropic presents Opus 5 as a cost‑efficient performer for software engineering, knowledge work and scientific research. Internal results cited by Anthropic show Opus 5 topping Frontier‑Bench v0.1, nearly matching Fable 5 on CursorBench 3.2 at roughly half the cost per task, and delivering large gains on specialized evaluations such as ARC‑AGI 3 and Zapier AutomationBench. The company reports material improvements in life‑sciences evaluations, including organic chemistry and protein‑function tasks, and says visual output capabilities are stronger.
For product teams and developers, unchanged pricing combined with higher per‑token performance and a Fast mode gives predictable upgrade economics: teams can migrate to a model Anthropic says is both more capable and more efficient without immediate price increases. The API fallback and CVP paths create operational options for balancing safety restrictions against productive workflows in security and biology projects.
Safety trade-offs and market positioning
Anthropic’s internal behavioral audits report Opus 5 as its most aligned model to date, scoring 2.3 on overall misaligned behavior and showing lower rates of deception and susceptibility to misuse than Opus 4.8, Sonnet 5 and Fable 5. At the same time, Anthropic notes Opus 5 did not push the frontier in risky dual‑use capabilities and remains behind Mythos 5 on both biological research and offensive cybersecurity tasks.
The company’s OSS‑Fuzz results illustrate that Opus 5 can identify software vulnerabilities at rates similar to Mythos 5 but is considerably less capable at developing exploits. Anthropic’s approach is therefore to permit helpful vulnerability discovery in many contexts while blocking higher‑risk exploit generation, using classifiers plus fallback behavior to route blocked requests to Opus 4.8 or to CVP participants.
Positioning-wise, Anthropic frames Opus 5 as approaching Fable 5 intelligence on many fronts while being less restrictive and substantially cheaper to run at task level. That combination aims to make Opus 5 attractive for daily developer and enterprise workflows where cost, latency and safety controls all matter.
What to watch next
Many of Anthropic’s performance claims come from internal runs and early‑access customer reports; independent third‑party benchmarks and tests will be important to validate those results in diverse production contexts. Monitor how frequently Anthropic’s classifiers intervene in real workloads and how often fallbacks to Opus 4.8 are triggered, since that behaviour determines both user experience and safety margins.
Also watch uptake through the Cyber Verification Program and whether enterprise customers push for broader CVP access. Track how Fast mode adoption balances the 2× price premium against 2.5× speed in latency‑sensitive applications. Finally, Anthropic’s System Card and any follow‑up technical documentation will be the definitive source for the detailed safety evaluations and limits the company alludes to.
Opus 5 gives developers and enterprise teams a higher‑capacity Opus model at the same public token price as Opus 4.8, with configurable effort, faster modes and new fallback and CVP controls shaping how capability and safety are balanced in practice.
Source: Anthropic