Midnight Release of Claude Opus 5: Using an Old Project to Hunt Bugs Yields Surprising Findings

Friday 1 AM Popup

At 1 am on a Friday, the author noticed a notification that Claude Opus 5 had been fully launched. An old pagination bug that had caused occasional crashes despite a community patch was chosen as a test case.

Opus 5 Analyzes Before Rewriting

Unlike previous AI tools that would rewrite code after a superficial scan, Opus 5 first generated test inputs to locate the root cause. The prompt given was:

"The paginate function has been patched for negative inputs but still crashes online. Do not modify it yet; try all inputs that could cause a crash and tell me the root cause."

Opus 5 ran several inputs and identified that floating‑point page numbers caused a crash, which the community patch had missed because it only guarded against negative values.

def paginate(items, page, page_size=10):
    if page < 0:
        page = 0
    start = page * page_size
    end = start + page_size
    return items[start:end]

data = list(range(25))
print("page=2.7:", paginate(data, 2.7))

Running this produced a

TypeError: slice indices must be integers or None or have an __index__ method

, confirming the floating‑point issue.

Second Task: Batch Split Edge Case

The author gave Opus 5 a batch‑splitting function that the community patch handled empty lists but not a zero batch size.

def batch_split(items, batch_size=5):
    if not items:
        return []
    return [items[i:i+batch_size] for i in range(0, len(items), batch_size)]

orders = ["A1","A2","A3","A4","A5","A6","A7"]
print("batch_size=0:", batch_split(orders, 0))

The execution raised ValueError: range() arg 3 must not be zero, showing that a zero batch size crashes the function.

Third Task: Market‑Data Parsing Demo

Using a publicly reported demo, Opus 5 built a test tool to verify code that parses exchange market data before proceeding, illustrating a “test‑first, then implement” workflow.

It builds its own test tool, confirms the code can correctly parse data, then proceeds.

This mirrors the author’s earlier experience with AI models that either stopped after a superficial rewrite or, like Opus 5, iterated between testing and fixing.

Performance and Pricing Comparison

Opus 5 is priced at half of Fable 5 yet, on the CursorBench 3.2 benchmark, reaches the highest effort tier and outperforms GPT‑5.6, staying within 0.5 % of Fable 5’s peak. On the OSWorld 2.0 desktop‑operation test, it achieves the best score of Fable 5 with only about one‑third of the cost.

The model automatically falls back to Opus 4.8 for security‑sensitive requests, and its Fast mode doubles speed at double price, so users should enable it only when time is critical.

Cost‑Effectiveness in Daily Use

Because the price is halved, the author feels comfortable keeping Opus 5 running for routine debugging, refactoring, and code reading, reserving the more expensive Fable 5 for particularly hard problems.

All test scripts, prompts, and detailed bug‑hunting notes are collected at the author’s public page (https://mp.weixin.qq.com/s?__biz=MzUxNTczOTczNQ==∣=2247504356&idx=1&sn=00fcf93441cef1ed44dc6269bef1794e&scene=21#wechat_redirect) for anyone who wants to avoid costly trial‑and‑error.

Final Thoughts

The biggest shift is not the model’s intelligence but its affordability, which removes the hesitation to keep it running. While today Opus 5 may be the cheapest high‑performing model, tomorrow a newer model could reclaim the lead. The practical advice is to master the inexpensive tool now rather than obsess over who is “best”.