Claude Opus 5 narrows the gap with Fable 5 while keeping agent costs lower
Anthropic has released Claude Opus 5, a new AI model aimed at high-end coding, office work and computer-use tasks at a lower price than its top Fable model. Reuters reported that Anthropic positioned the model as close to Claude Fable 5 on performance while costing about half as much.
According to Anthropic’s Claude Fable page, Fable 5 launched on June 9, 2026, before access was disrupted by U.S. export controls and restored on July 1. Anthropic said the model was built for complex coding and professional work, but also required safeguards around areas such as cybersecurity.
Claude Opus 5 matters because pricing is becoming a serious constraint for AI agents. Axios reported that Opus 5 costs $5 per million input tokens and $25 per million output tokens, matching Opus 4.8 while coming in below Fable 5. That makes the model more relevant for companies running repeated agent tasks, where small cost differences can become large bills.
The benchmark claim getting attention is ARC-AGI-3. The ARC Prize Foundation said Claude Opus 5 scored 30.16% at high reasoning effort on ARC-AGI-3 as of July 24, 2026. The benchmark tests agents on new interactive tasks rather than memorized facts, according to the ARC-AGI-3 paper released in March 2026.
Computer use is the other pressure point. OSWorld 2.0, a benchmark paper released in June 2026, describes 108 long-horizon workflows that require agents to operate software across realistic tasks. The paper said older frontier agents still struggled with hidden state, verification and long task chains. That is why the reported Claude Opus 5 gain on OSWorld-style tasks is useful, but not a full proof of office reliability.
The positive case is clear. A cheaper model that performs near a more expensive one gives developers and enterprises more room to test AI agents without moving every task to the most costly system. This fits the pattern The AI Decode covered in its reporting on AI loops and Claude Code, where coding agents are moving from single prompts toward longer work sessions.
The risk is also plain. Benchmarks are controlled tests, not live business environments. A June 2026 red-team study of Anthropic’s Fable 5 and Opus 4.8 models found that both resisted many automated jailbreaks but still produced confirmed harmful completions under sustained attacks. That does not mean Claude Opus 5 is unsafe by default. It does mean cost and capability claims should be read beside safety testing.
For Anthropic, the next question is whether Claude Opus 5 can hold up in real enterprise use, where tasks are messy, permissions are sensitive and users expect the agent to know when to stop.
