A developer's candid essay about Anthropic's latest model is sparking a conversation that every AI developer should be paying attention to. The post, titled "Why does Opus 5 feel worse to work with?", was published on August 14, 2026, and quickly climbed to 172 points on Hacker News. Its core argument is both simple and profound: a more capable AI model can feel worse to work with if it stops asking questions and starts making assumptions.
The Problem: Capability ≠ Usability
The author, who goes by "mun logadan," makes a counterintuitive claim: Opus 5 is objectively more capable than its predecessors (Opus 4.7, Opus 4.8, and even the Fable model), yet it feels like a downgrade in day-to-day use. The reason? The older models exhibit three behaviors that Opus 5 has lost:
- They stop and ask questions when intent is unclear
- They don't make assumptions without checking first
- They don't reinterpret or update plans without asking permission
Opus 5, by contrast, charges ahead with confident assumptions. It's more capable in the abstract, but it requires more careful babysitting in practice. The model that asks "did you mean X or Y?" is more useful than the model that assumes you meant X and builds an entire feature around it.
The Root Cause: Benchmark Optimization
The essay's most incisive section traces this behavior to its source: benchmark training is selecting against the exact behaviors developers want.
Here's the mechanism:
- Benchmark tasks are self-contained. They have a correct answer that can be reached without hints or clarification.
- Training models to excel at benchmarks (and on RLVR tasks in general) inherently selects for models that make bold, usually-correct assumptions in the face of ambiguity.
- This penalizes models that would stop and ask for clarification or direction.
In other words, the optimization pressure rewards models that "just go for it" and punishes models that pause to check. On a benchmark, pausing to ask questions is wasted time. In real software development, pausing to ask questions is the difference between building the right thing and building the wrong thing really fast.
Two Compounding Forces
The author identifies two forces driving this trend:
1. The push for self-improving AI. The desire to create models that can recursively bootstrap themselves toward AGI/ASI encourages autonomous behavior. A model that asks questions is dependent on human input — not ideal for a self-improving loop.
2. Benchmark pressure. Despite it being an "open secret that many benchmark tasks are ill-defined, unfair, hackable, or otherwise broken," benchmarks still drive model development. The incentive structure rewards models that produce answers, not models that produce questions.
Why This Matters for Every AI Developer
This isn't just about Anthropic. The same dynamics affect every frontier lab. As the author notes, this is "current frontier labs in general." The pattern is systemic:
- OpenAI's GPT models have shown similar trajectory — each generation more confident, less questioning.
- Google's Gemini models optimize heavily for benchmark performance.
- DeepSeek, Qwen, and others all face the same pressure to score well on standardized tests.
The result is a generation of AI coding agents that are individually more capable but collectively harder to work with. They write more code, faster — but some of that code solves the wrong problem.
The Real World Isn't a Benchmark
The essay's conclusion is its most powerful point:
Real life just isn't a benchmark. There isn't a guaranteed right answer to every question, nor even a set of right answers, and with real-life consequences on the line, I do not want an agent taking its best guess!
This is the fundamental tension in AI development. Benchmarks are necessary — without them, we can't measure progress. But optimizing too hard for benchmarks produces models that are optimized for the test, not for the work.
In real software development:
- Business implications aren't written in the prompt
- Budget constraints aren't in the system message
- The "correct" answer depends on context the model doesn't have
- Ambiguity is the norm, not the exception
A model that asks "should I use PostgreSQL or SQLite here?" is more valuable than one that silently picks PostgreSQL and migrates your entire data layer.
What Can Be Done?
The essay doesn't offer solutions, but the diagnosis suggests several:
System prompts that explicitly require clarification. Developers can counteract the model's tendency to assume by instructing it to ask questions when ambiguous.
Benchmarks that reward question-asking. If the AI research community designed benchmarks that scored models on knowing when to ask rather than just what to answer, the training pressure would shift.
Model selection based on workflow fit. Sometimes a "less capable" model is the right tool. If Opus 4.8 asks questions and Opus 5 doesn't, use 4.8 for exploratory work and 5 for well-defined tasks.
Human-in-the-loop by design. Build pipelines that require confirmation before irreversible actions. This is already best practice, but it becomes more critical as models get more autonomous.
The Bigger Picture
This essay touches on something deeper than model preference. It's about the alignment between how we train AI and how we use AI. We train models to solve problems autonomously, but we use them in collaborative workflows where autonomy without communication is actively harmful.
The gap between benchmark performance and real-world utility is the most important unsolved problem in AI development today. A model that scores 95% on a coding benchmark but makes wrong assumptions 30% of the time in production is less useful than a model that scores 85% on benchmarks but asks the right questions.
As AI agents become more autonomous — running for hours, making decisions, writing code — the cost of unwarranted assumptions grows. An agent that assumes wrong for 5 minutes wastes 5 minutes. An agent that assumes wrong for 14 hours wastes 14 hours and may have rewritten half your codebase in the wrong direction.
The best AI model isn't the one that knows the most answers. It's the one that knows what it doesn't know — and asks.
This article is based on the essay "Why does Opus 5 feel worse to work with?" by mun logadan, published August 14, 2026.
Top comments (0)