Claude Opus 5 dropped on July 24, 2026, and within 48 hours two things were simultaneously true: it had claimed the top spot on virtually every major AI leaderboard, and developers were publicly complaining that it was a pain to work with. Product leader Claire Vo said it plainly — "Opus 5 is here…and I hate working with it."
That tension is worth unpacking carefully. Because if you dismiss the model based on the complaints, you'll miss genuinely frontier-grade performance at half the cost of Fable 5. And if you reach for it without understanding its quirks, you'll burn tokens and patience in equal measure.
The Benchmark Reality
Let's start with what's actually true about the numbers.
Frontier-Bench v0.1 — Anthropic's agentic terminal coding benchmark — Opus 5 scores 43.3%, nearly 10 points ahead of Fable 5 (33.7%) and more than double Opus 4.8's 21.1%. On the tasks that matter most for autonomous coding agents, Opus 5 isn't just better than Fable — it's in a different league.
ARC-AGI-3 — the gold-standard novel reasoning benchmark — Opus 5 scores 30.2%, independently verified by the ARC Prize Foundation. What's striking isn't just the number: the Foundation noted Opus 5 "used advanced logical reasoning to turn ARC-AGI-3 layouts into algebraic notation," developing its own mathematical representation rather than pattern-matching. That's qualitatively different reasoning.
IMO 2026 — Opus 5 achieved a gold-medal performance (42/42 points) on International Mathematical Olympiad problems.
Independent validation from Artificial Analysis agrees: Opus 5 ranks first on both the Agentic Index and Intelligence Index (61 vs Fable 5's 60), and second on the Coding Index — just 0.3 points behind GPT-5.6 Sol.
Real-world tests back this up. In a 3D F1 racing game built from a single HTML file, Opus 5 produced a polished, functional result with minor physics issues. Fable 5 was solid but simpler. GPT-5.6 Sol and Kimi K3 produced unplayable builds with missing track geometry. In a personal finance dashboard test, Opus 5 delivered clean architecture, React Router, and a working SQLite backend where others either reinvented the wheel or reached for an unnecessarily complex stack.
So Why Are Developers Frustrated?
The complaints aren't about intelligence. They're about personality.
It bails on contradictions. Give Opus 5 instructions that point in different directions — even slightly — and it will stall or ask for clarification rather than making a reasonable call. Fable 5 would pick a path and explain its reasoning. Opus 5 treats ambiguity as a stopping condition.
It takes initiative you didn't ask for. The model runs long, adds caveats, restructures things you wanted to keep as-is. Product teams building tight workflows report the model doing "extra work" that breaks downstream steps.
"Claudeslop." Vo coined this term for idiosyncratic language patterns — strings of non-sequiturs, formulaic sentence endings, a distinctive cadence that makes outputs feel machine-generated in ways that earlier Claude models managed to avoid.
In a blind taste test, the same reviewers who complained ranked Opus 5 above every competitor model, including Fable 5 and GPT-5.6. Brilliant, but annoying is the most accurate summary.
Pricing: Where the Math Gets Interesting
$5 / $25 per million input/output tokens — same price as Opus 4.8, half the cost of Fable 5 ($10 / $50). Fast mode (extended thinking disabled) runs at $10 / $50.
The 1M-token context window means you can feed it entire codebases. And crucially, Opus 5 has no data retention requirements, unlike Fable 5 — a meaningful advantage for privacy-sensitive deployments.
For agentic workflows — where a task might require 20-50 model calls — the cost difference between Opus 5 and Fable 5 compounds fast. If Opus 5 outperforms Fable 5 on your benchmark while costing half as much, the business case is clear.
How to Actually Work With It
The frustrations above are real but mostly predictable, which means they're also avoidable.
Write tight, non-contradicting instructions. Opus 5 exposes sloppy prompts that Fable 5 would paper over. Use this as a forcing function: if your system prompt is vague, Opus 5 will tell you. Treat stalls as signals to improve your prompts, not as model failures.
Set scope boundaries explicitly. Tell it what not to do. "Modify only the function I specify. Do not refactor surrounding code." Opus 5 respects explicit constraints more reliably than Fable 5 — you just have to state them.
Use it for the right jobs. Opus 5 excels at:
- Long-horizon agentic coding tasks
- Complex reasoning with many dependencies
- Tasks where you want the model to flag ambiguity rather than guess wrong
- Privacy-sensitive deployments where data retention is a constraint
Consider a different model when:
- You need fast creative iterations where its verbosity slows you down
- Your instructions are inherently ambiguous and you want the model to bridge the gap rather than stop
- Latency is critical (Opus 5 runs slower than Sonnet 5 by design)
The Routing Strategy
The smartest teams aren't choosing one model — they're routing.
- Opus 5 for deep agentic tasks, complex code generation, reasoning chains
- Sonnet 5 for fast interactive tasks, chat interfaces, first drafts
- Haiku 4.5 for classification, routing, high-volume extraction
Opus 5's adjustable effort level (thinking tokens) gives you a dial: low-effort for quick answers, high-effort for the problems that need the full model. This is the architecture Anthropic was hinting at when they said they want Opus 5 "used every day."
The Bottom Line
Claude Opus 5 is the best model for autonomous coding agents and complex reasoning tasks — full stop. The benchmark gap over Fable 5 on Frontier-Bench (43.3% vs 33.7%) is too large to dismiss, and independent leaderboards confirm it.
The "brilliant but annoying" verdict is accurate, but it's also a symptom of misuse. Teams that write precise instructions, route tasks deliberately, and treat its stalls as feedback rather than failures will find Opus 5 delivers frontier-grade results at half the price of the alternative.
The developers who hate it are often running their Fable 5 prompts unchanged. Write for Opus 5's strengths and you'll ship better work faster — even if you occasionally want to close the terminal in frustration.