Claude Opus 5 Early Reviews: Efficiency King

Anthropic released Claude Opus 5 on July 24. Four days post-launch, early reviews are positive. The model scored 61 on the Artificial Analysis Intelligence Index, edging past the company's own flagship Fable 5 (60), marking the first time an Opus-class model has topped the benchmark. Pricing remains at $5 per million input tokens and $25 per million output, unchanged from Opus 4.8.

However, user discussions focus less on benchmark scores and more on operating costs. Across Reddit and X, feedback highlights the model's low token consumption. In the launch post, Anthropic positioned Opus 5 as a daily-use model offering near-Fable intelligence at half the price, a characterization that early adoption patterns seem to validate.

Praise for Token Efficiency: 'The Usage Counter Barely Moves'

Token consumption is the most discussed topic on r/ClaudeAI. In a thread titled Opus 5 Token Usage is Amazing, Max-plan subscribers reported difficulty reaching their limits. One user noted that after running heavy agent sessions for days, they had consumed only 2% of their weekly quota.

The model's new effort setting drives the token efficiency gains. Comparative reviews show that Claude Opus 5 on a low-effort setting matches the output quality of Opus 4.8 on high effort, and fewer failed runs mean fewer retries. For subscribers, the same monthly fee now stretches across several times more work.

Coding Benchmarks Near Fable 5 at Half the Cost

Coding benchmark data supports these efficiency reports. On CursorBench 3.2 at maximum effort, Claude Opus 5 lands within 0.5% of Fable 5's peak score at half the cost per task, and on Frontier-Bench it more than doubles Opus 4.8's result. On OSWorld 2.0, a computer-use benchmark, it beats Fable 5's best figure at just over a third of the price.

The New Stack reported that this pricing structure shifts the cost-to-performance ratio for agentic coding, making it viable for development teams to run agents on complex tasks for hours. The efficiency pitch is not simply a lower price, but rather more hours of high-performance coding per dollar.

General-Purpose Work Still Belongs to Fable 5

However, these advantages are primarily limited to coding and execution tasks. In CodeRabbit's benchmark review, Opus 5 performed better at code generation than at code review. While its precision improved, its overall issue coverage fell and the volume of minor criticisms quadrupled compared to the baseline.

For knowledge-intensive judgment calls and multi-agent orchestration, users still rate Fable 5 significantly higher, noting that Opus 5 tends to jump to conclusions Fable would flag. Consequently, an emerging workflow utilizes Fable 5 for planning and orchestration while employing Opus 5 for execution. This division of labor — Fable 5 handling the orchestration, Opus 5 the build — casts Opus 5 as a specialized execution tool rather than a general-purpose flagship.

The Next Battle: August GPT-6 Rumors and a 5.1 Follow-Up

The timing of this praise matters as much as its content. Anthropic had spent the first half of the year losing users to OpenAI's Codex, battered in turn by the Sonnet 5 backlash, a lukewarm Opus 4.8, and the subscription-exclusion controversy around Fable 5. With token efficiency now the model's calling card, early reviews point to that exodus slowing.

Attention now shifts to August. An X post claims OpenAI could ship GPT-6 as soon as August, while chatter about a Fable 5.1 point release circulates on the Anthropic side. Neither is official, yet both are widely treated as likely scenarios. For Anthropic, preparing to go public, the stakes could hardly be higher — the next question is whether the momentum Claude Opus 5 restored can hold through August.