Claude Opus 5 Costs $5 per Million Input Tokens: Anthropic Pulls Agent Competition Back to Cost

Over the next three to six months, Anthropic’s efficiency claim will fail to carry from formal benchmarks into production if companies do not see a lower cost per successful task. The company has not disclosed API volume, enterprise retention, or independent production-cost data. Buyers therefore need to test the claim with deployment-level success rates, token consumption, and time spent on human rework.

Anthropic released Claude Opus 5 on July 24, 2026 and made it available through the Claude API, Claude Code, Claude.ai, and major cloud platforms. The API price is $5 per million input tokens and $25 per million output tokens, unchanged from Opus 4.8. Fast mode runs at roughly 2.5 times the default speed and costs twice the base rate. Opus 5 is now the default model for Claude Max and the strongest model available to Claude Pro subscribers, positioning it for routine agent and professional workloads rather than only a small number of maximum-difficulty requests.

Opus 5 competes with Fable 5 on cost per completed task

Anthropic says Opus 5 leads the tested models on Frontier-Bench v0.1 while delivering more than twice the performance of Opus 4.8 at a lower cost per task. On CursorBench 3.2 at the max effort setting, it comes within 0.5% of Fable 5’s peak score at about half the cost per task. TechCrunch likewise identified the lower price and lighter restrictions as central to the model’s position, noting that Opus 5 outperformed Fable 5 on several benchmarks included in Anthropic’s announcement.

Those results have a defined boundary. Test setup, tool access, and the definition of failure all change the cost of a successful task, and a vendor launch package cannot replace evaluation on a company’s own repositories, permissions, and latency requirements. Anthropic cited improvements reported by early customers in debugging, financial research, and long-running agent tasks. It did not publish a standardized sample, control group, or total number of failed attempts, so those examples do not establish an economy-wide productivity gain.

Fewer classifier interventions still leave explicit cyber limits

Anthropic expects the cybersecurity classifiers for Opus 5 to intervene around 85% less often than those for Fable 5. The model may find vulnerabilities in source code, but the safeguards still block binary vulnerability scanning, penetration testing, and exploit generation. Anthropic’s own tests also place it substantially behind Mythos 5 at developing exploits. A beta automatic-fallback feature can route a flagged API request to another model instead of returning an error immediately.

“Less restrictive” therefore describes a revised boundary between permitted defensive work and blocked high-risk actions. Anthropic’s automated behavioral audit gave Opus 5 an overall misaligned-behavior score of 2.3, the lowest among its recent models, but this remains a company-designed test. External users will need to measure false-block rates, completion after fallback, and whether output quality survives a handoff between models.

The launch puts the measurable variables on one cost chain: tokens consumed per task, fallback events, and human corrections required before completion. If Opus 5 continues to lead formal benchmarks during the next three to six months but enterprise success costs and rework time do not fall, the $5 input rate will mean only that the list price held steady. It will not demonstrate a reduction in the total cost of agent work.

Sources:

• Anthropic: Introducing Claude Opus 5
• TechCrunch: Anthropic launches Opus 5