Claude Opus 5 lands with frontier-level claims, $5 input pricing, and new API tools
AI寒武纪 · wechat · 2026-07-25
A long Chinese roundup of the Claude Opus 5 launch. It says Opus 5 is positioned for heavy daily use, with roughly frontier-level intelligence at half the price of the higher-tier model and major gains on software engineering, knowledge work, computer use, and life-science tasks.
Key claims
- It reportedly leads Frontier-Bench, roughly doubles Opus 4.8 on performance at lower cost, and comes close to the top model on CursorBench 3.2 at about half the cost.
- It also claims strong results on ARC-AGI 3, Zapier automation tasks, and OSWorld 2.0 computer-use benchmarks.
- The post highlights a practical demo where the model reconstructs a 3D object from a part drawing by building its own vision pipeline, plus examples of fixing open-source bugs and creating a market-data validation tool.
API and usage notes
- Pricing is said to match Opus 4.8: $5 / 1M input tokens and $25 / 1M output tokens.
- A Fast mode is available at 2.5× speed for 2× the base rate.
- New API features include changing tools mid-conversation without invalidating cache, and automatic fallback to another best-available model if a request is blocked by a safety classifier.
- The article also summarizes prompt-engineering advice: let the model report all issues in code review, use tools like cropping and visual verification for vision tasks, constrain scope on small tasks, and cap sub-agent usage so costs do not explode.
Takeaway
The author’s conclusion is that Opus 5 looks good enough to use broadly, especially for coding and multi-step agent work.
Related event: Anthropic Releases Claude Opus 5 with Leading Benchmark Performance(77 posts)→
More from coding & agent
- Kimi Code hits a usage cap on a $200 plan during an Agent Swarm run — xeophon · 2026-07-25
- Perplexity ships a CLI that gives coding agents web search access — AravSrinivas · 2026-07-25
- Opus 5 Codes 3D Colosseum Game with a Single Prompt — chrisfirst · 2026-07-25
- A simple proxy trick helps debug agent skills by intercepting every call — Daniel_Farinax · 2026-07-25
- GPT-5.6 Sol edges Opus 5 on DeepSWE with 72.7% vs 68.8% — rohanpaul_ai · 2026-07-25
- Nimbus launches as an open-source Astro framework for agent-ready docs — irvinebroque · 2026-07-25