跪拜 Guibai
← All articles
Backend · Programmer · Artificial Intelligence

Claude Opus 5 Matches Fable 5's Peak at Half the Cost

By cxuanAI ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

A frontier-class model at half the cost of the top tier changes the economics of agentic coding workflows. Teams that couldn't justify Fable 5's pricing for long-running autonomous tasks now get near-identical capability, and the self-verification behavior means fewer wasted cycles on hallucinated completions.

Summary

Claude Opus 5 launched July 25 at $5/$25 per million input/output tokens — the same price as Opus 4.8 but with capabilities that land within 0.5% of Fable 5 on CursorBench 3.2. On Frontier-Bench v0.1, it doubles Opus 4.8's score at lower per-task cost. ARC-AGI 3 scores are triple the runner-up, and AutomationBench pass rates run roughly 1.5x the next-best model at equivalent spend.

The model's standout behavior is autonomous verification. When a benchmark denied it image-viewing tools, Opus 5 wrote a computer vision pipeline from scratch to extract geometry from raw pixels. Given a real open-source bug with an incomplete community patch, it found the root cause and fixed the edge case rather than papering over the symptom. Anthropic now advises developers to remove old prompt instructions that told earlier models to double-check their work — Opus 5 does it unprompted, sometimes to the point of over-checking.

A 1M-token context window, 128k-token max output, and five thinking-effort levels (low through max) ship by default. Fast mode runs 2.5x quicker at double the price. Claude Max subscribers get Opus 5 as the default model; Pro subscribers also have access. Claude Code lead Boris Cherny notes it is the Claude model most resistant to prompt injection, and automated audits give it the lowest misalignment score among recent releases.

Takeaways
— API pricing is $5/M input tokens and $25/M output tokens, identical to Opus 4.8 and exactly half of Fable 5's $10/$50.
— On CursorBench 3.2 at max effort, Opus 5 trails Fable 5's peak by less than 0.5% while costing half as much per task.
— Frontier-Bench v0.1 scores more than double Opus 4.8's result at a lower per-task cost.
— ARC-AGI 3 scores are triple the second-place model; AutomationBench pass rates are roughly 1.5x the next-best at equivalent cost.
— Opus 5 autonomously verifies its own output — writing CV pipelines to extract data when tools are missing, finding root-cause bugs instead of surface fixes, and building test frameworks when live data is unavailable.
— Anthropic recommends removing old prompt instructions that told earlier models to double-check work, since Opus 5 already does it.
— Context window is 1M tokens, max output is 128k tokens, with five thinking-effort levels from low to max.
— Fast mode delivers 2.5x speed at double the price.
— Claude Max defaults to Opus 5; Claude Pro users also have access.
— Automated behavior audits give Opus 5 a misalignment score of 2.3, the lowest among recent Claude models.
— Vulnerability-finding capability approaches Mythos 5, but exploit-development ability still lags significantly.
Conclusions

Anthropic is compressing its own pricing ladder: a model that nearly matches the flagship ships at half the cost only a month after that flagship launched, which pressures competitors to justify premium tiers.

The self-verification behavior is a meaningful shift from 'model completes task' to 'model validates its own output,' and the official advice to strip old verification prompts suggests this is reliable enough to change how developers write agent instructions.

Pricing Opus 5 identically to Opus 4.8 while delivering Fable-5-class performance effectively obsoletes the previous generation at the same price point overnight.

Making Opus 5 the default in Claude Max signals Anthropic is willing to cannibalize Fable 5 usage for subscription revenue, betting that volume and lock-in outweigh per-token margins.

Concepts & terms
Thinking effort levels
A setting in Claude Opus 5 that controls how much compute the model spends reasoning before responding. Ranges from low (fast, cheap, suitable for simple tasks) through medium, high, xhigh, to max (slowest, most expensive, for hard problems).
CursorBench 3.2
A benchmark that evaluates AI coding assistants on real-world software engineering tasks, measuring both task-completion quality and cost efficiency.
Frontier-Bench v0.1
Anthropic's internal benchmark for testing frontier model capabilities on difficult, novel problems that require reasoning beyond pattern matching.
ARC-AGI 3
A benchmark designed to measure a model's general reasoning ability on entirely new problems it has not seen during training, testing fluid intelligence rather than memorization.
Prompt injection resistance
A model's ability to ignore instructions embedded in user-provided data that attempt to override or manipulate the system prompt or safety guidelines.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗