Claude Opus 5 ships at the old price, with 100k context and stronger agent scores

卡尔的AI沃茨 · wechat · 2026-07-25

A long Chinese review argues that Claude Opus 5 is strong enough to move into the same tier as top frontier models for planning and execution — at a much lower price.

Main points

The author says Opus 5 surprised them by launching sooner than expected and outperforming or matching expectations in several areas, especially multi-step reasoning, agentic coding, legal, and health tasks.
Pricing is kept at $5 / million input tokens and $25 / million output tokens, the same as Opus 4.8, with a fast mode that is 2.5× faster at 2× price.
It supports 100k context and claims knowledge updated to May 2026.
In the author’s testing, Opus 5 comes within 0.5% of Fable 5 in maxeffort mode, and beats other models on performance/cost in high / xhigh / max settings.
The post also highlights an ARC-AGI-3 score said to be more than 3× GPT-5.6’s, which the author interprets as strong performance in unfamiliar, rule-discovery environments.

Coding-agent workflow notes

The author now routes work as: Fable 5 for planning, GPT-5.6 and Opus 5 for execution, and Fable 5 again for integration.
They also report using GPT-5.6 Sol Pro in the web UI for large code reviews to save quota.
A major section discusses slimming down Claude Code: Anthropic reportedly removed about 80% of the system prompt, and the author argues fewer hard rules, examples, and redundant constraints can improve agent performance.
OpenAI’s internal guidance experiments are cited as well: simpler system prompts allegedly improved coding-agent scores by 10–15%, cut tokens by 41–66%, and reduced cost by 33–67%.
The post recommends using /doctor to audit unused skills, MCPs, and plugins, then classifying them into high-frequency, project-level, and low-frequency archived tiers instead of deleting them outright.

Original post →

Claude Opus 5 ships at the old price, with 100k context and stronger agent scores

More from coding & agent