Anthropic Launches Claude Opus 5 at $5 per Million Tokens, Half the Cost of Fable 5

Positioning: Near-Frontier Intelligence at Half the Cost

Anthropic launched Claude Opus 5 on July 24 at a price of $5 per million input tokens and $25 for output. This is the same rate charged for Opus 4.8, but with radically different performance: in Frontier-Bench v0.1, an internal benchmark for reasoning and automation tasks, Opus 5 doubled the score of Opus 4.8 at the same cost. In CursorBench 3.2, focused on LLM-assisted coding, it achieved 0.5% of the maximum score of Fable 5, Anthropic's flagship model which costs twice as much: $10 for input and $50 for output per million tokens.

The context window has increased to 1 million tokens, with synchronous output of up to 128,000 tokens in a single API call. The timing is strategic: OpenAI's GPT-5.6 Terra was released on July 9, and Google's Gemini 2.5 Pro is already operating in the same capability range. The mid-tier has become the real battleground among closed AI providers, as this is where most enterprise workloads operate in production.

The Mechanism: The Five-Level Effort Dial

Ranging from "low" to "maximum", the Opus 5 API calibrates the depth of reasoning per request. Tasks like triaging, extraction, or simple routing do not require the same processing power as a legal analysis or code review. The model adds a fast mode at $10 per million input tokens, 2.5 times faster than the standard, prompt caching at $0.50 per million tokens, and a 50% reduction via Batch API.

This explicit control of effort per call has no equivalent in GPT-5.6 Terra. For developers building agents with hundreds of chained steps, the adjustment between "maximum" and "medium" in intermediate steps can represent a 30% to 40% reduction in total execution cost without measurable loss in the quality of final outputs.

What Launch Partners Measured

Harvey, a legal AI platform used by major law firms in London and New York, reported a 26% reduction in the volume of tokens generated compared to Opus 4.8 in maximum reasoning mode, with equivalent quality assessed internally. In high-volume legal contracts, the difference directly impacts the cost per dossier.

The CEO of Zapier reported that Opus 5 achieved 100% approval in the company's internal AutomationBench, a test simulating end-to-end churn prevention flows with multiple chained tools. Previous versions, including Opus 4.8, did not reach the same result in this evaluation.

Two Markets Where the Cost Equation Changes

In the United States, the launch compresses the mid-tier: the architectural decision has shifted to cost per task, not model ranking position. The cheaper GPT-5.6 Luna, at $1 per million input tokens, does not reach the analytical depth of Opus 5. The Sun, flagship of the same OpenAI family, is more expensive and serves different use cases. For integrators who have monetized the last generation of engagements based on Opus 4.8, Opus 5 redefines the cost baseline without requiring a change in providers.

In India, TCS, Infosys, and HCL Technologies operate development centers utilizing Claude as the backbone for enterprise automation for European and North American clients. The control of effort per request alters the equation in large-scale projects: workloads for automatic validation and triaging, which currently consume uniform inference capacity, can now be managed by task complexity. The cost of inference changes from being constant to a controllable variable in the P&Ls of engagements, directly impacting AI-as-a-service contract margins.

Anthropic has not disclosed adoption projections or a timeline for subsequent versions of Opus. The available data consists of benchmarking details and those reported by launch partners.