Opus 5 is more capable, but feels worse to work with
A blog post that drew heavy discussion on Hacker News argues that Opus 5, despite being a stronger model than Opus 4.7 and Opus 4.8 and a benchmark rival of Fable, is a worse day-to-day coding agent than all three. The author states plainly that this is not a capability regression: Opus 5 tests as more capable. The complaint is about how it behaves in a session.
The author identifies three specific behaviors that set Opus 4.7, Opus 4.8, and Fable apart from Opus 5: those models stop and ask a question when the user's intent is unclear, they do not make assumptions without checking first, and they do not reinterpret or silently update the user's plan without asking. Opus 5, in the author's experience, does less of this, which the author says forces "careful babysitting" that the other models do not require.
The post's second half, headed "Baseless speculation" in the author's own words, offers an explanation rather than a documented finding. The author suspects two compounding pressures at Anthropic and at frontier labs generally: the push toward a self-improving system capable of recursively bootstrapping itself to AGI or ASI, and the pressure to score well on benchmarks. The author's argument is that a good benchmark task is, by design, self-contained: it can be solved without hints, without reading the task creator's mind, and without outside information. Selecting or training models to do well on such tasks, including RLVR-style training, rewards bold, usually-correct guessing under ambiguity and penalizes a model's tendency to stop and ask for clarification or direction.
The author argues that this is close to the opposite of what a coding agent needs in practice. It is nearly impossible, the post says, to write down and hand an agent the full context, intent, business implications, budget constraints, and everything else relevant to a real task, so ambiguity and judgment calls are unavoidable. With real consequences riding on the outcome, the author says they do not want an agent taking its best guess instead of asking.
Key facts
- The author rates Opus 5 as more capable than Opus 4.7 and Opus 4.8 and roughly on par with Fable on benchmarks, but says all three of those models feel better to work with day to day.
- The gap is attributed to three behaviors: earlier models stop to ask when intent is unclear, avoid unchecked assumptions, and do not silently reinterpret or update the user's plan; the author says Opus 5 needs 'careful babysitting' by comparison.
- The author explicitly labels the causal explanation 'baseless speculation': two compounding pressures at Anthropic and at frontier labs in general, a push toward self-improving AI capable of bootstrapping to AGI or ASI, and pressure to score well on benchmarks.
- The argument holds that self-contained, solvable-without-clarification benchmark tasks (including RLVR-style training) select for models that guess boldly under ambiguity and penalize models that ask for clarification, which the author says is the opposite of what a coding agent needs.
- The essay reached 847 points and 765 comments on Hacker News within about 21 hours, indicating the complaint resonated widely among readers.
Why it matters
The post names a tension that is easy to overlook when a new model is judged mainly by leaderboard position: optimizing for benchmark scores can push a model toward confident guessing, while the same trait makes it harder to trust with ambiguous, high-stakes work. If the author's reasoning holds even partly, it suggests capability gains and agent trustworthiness are not the same axis and can move in opposite directions.
Who it affects
Anyone using Opus 5 as a coding or task agent, and more broadly anyone choosing between frontier models for agentic work where ambiguity is common. The post also speaks to how Anthropic, and frontier labs generally, weigh benchmark performance against behaviors like asking for clarification when they select and train models.
How to use it
The author's own takeaway is behavioral, not a workaround: know that Opus 5 is more likely to act on an assumption than to stop and ask, so tasks handed to it may need more explicit context, constraints, and guardrails up front than the same tasks would need with Opus 4.7, Opus 4.8, or Fable. Readers who value a model that checks before acting may weigh that against Opus 5's raw capability edge when picking a model for a given task.
How solid is it
This is a single author's opinion, informed by conversations with unnamed colleagues, not a benchmarked study. The author states outright that the causal explanation is 'baseless speculation,' and the post cites no benchmark scores, no specific test names, and no example transcript of Opus 5 behaving as described. The Hacker News discussion (847 points, 765 comments) shows the complaint resonates with many readers, but that is a measure of agreement, not of verification.
Risks and caveats
The comparison to Opus 4.7, Opus 4.8, and Fable rests on the author's subjective impression rather than a controlled test, so it could reflect prompting style, task type, or individual workflow rather than a general property of the models. The explanation for why this happens is explicitly unverified speculation about training incentives at Anthropic and other labs, and should be read as a hypothesis rather than a documented account of how any model was actually trained.
“With real-life consequences on the line, I do not want an agent taking its best guess!”
— the author