Claude's newer models are refusing more benign requests, and the complaints are piling up. An Android Authority column published July 11, 2026 argues the latest Claude models are "slowly ruining it," and a Hacker News thread linking to the piece drew early engagement — a signal that over-refusal has moved from scattered gripes to a recognizable trend.
The core complaint isn't that Claude has guardrails. It's that the guardrails now fire on ordinary creative fiction, research questions, and brainstorming — and fire inconsistently.
What users are reporting
The Android Authority piece, written by a self-described longtime Claude user who writes fiction and researches philosophy and religion, describes a pattern many will recognize:
- Creative fiction gets flagged. A short-story premise — aliens arriving with texts that cast doubt on mainstream religious claims — was refused outright, with Claude saying it was uncomfortable presenting material that could affect a real-world religion "as if it were fact," even after the author clarified it was fiction.
- Research questions hit walls. A straightforward question about the origins of Zoroaster and its influence on Judaism drew defensiveness rather than an answer.
- Hypotheticals get treated as plans. Asking about the biology of a hypothetical alien on a fictional planet around a real star triggered scope-reduction, as if the user were literally engineering organisms.
The author reports similar pushback on financial planning and brainstorming prompts — categories that aren't sensitive by any normal reading.
The real problem is inconsistency
The sharpest observation in the piece isn't "Claude said no." It's that Claude says no, then says yes to the exact same prompt in a fresh conversation. The author recreated a refused scenario with identical prompts on the same model and it ran without comment the second time.
That inconsistency is worse than a hard refusal. A predictable boundary you can plan around. A boundary that moves at random trains users to reroll conversations, reword harmless requests, or abandon the tool mid-task — none of which improve safety, all of which burn time and tokens.
Which models draw the most complaints
Per the author's own testing, the pushback isn't uniform across Claude's lineup. The account describes a rough gradient:
| Model | Reported over-refusal (per the author's usage) |
|---|---|
| Fable 5 | High on some categories (notably biology-related prompts) |
| Sonnet 5 | Higher |
| Sonnet 4.6 | Higher |
| Opus 4.8 | Higher; heavily reduces scope of hypotheticals |
| Opus 4.6 | Fewer issues reported |
| Opus 4.7 | Fewer issues reported |
Treat this as one experienced user's lived impression, not a benchmark — it's directional, not measured. But it lines up with a common community read: the newest models tightened, and the previous generation felt more workable for edge-case-but-legitimate prompts.
Why this might be happening
The author speculates the shift traces back to the period after Fable 5 was temporarily taken down amid a government security concern, after which Anthropic worked to restore it. The theory: guardrails were turned up to reassure regulators, and the effect spread across the lineup rather than staying scoped to one model.
We can't confirm the causal chain — Anthropic hasn't published a "we tightened refusals" note tied to that timeline, so this stays firmly in the speculation column. What's observable is the pattern: newer models, more false refusals, and a lineup where the previous generation feels less restrictive.
For a broader look at Anthropic's product decisions drawing scrutiny, see our writeup on Claude Code's steganographic system-prompt markers. And if over-refusal is pushing you to evaluate alternatives, our Claude Code vs. Codex comparison breaks down the trade-offs.
What you can actually do about it
Until the behavior settles, a few practical moves help:
- Be explicit and concise about intent. State that a request is fiction, research, or a hypothetical up front. Vague prompts get read as darker than intended.
- Reroll on inconsistent refusals. If a benign prompt is refused, a fresh conversation with the same wording sometimes passes — annoying, but effective.
- Match the model to the task. If your workflow leans on edge-case-but-legitimate prompts, the older-generation models reportedly refuse less often.
- Keep receipts. If you hit a clearly-wrong refusal, that's exactly the feedback Anthropic needs to recalibrate.
Guardrails matter. But a safety system that refuses fiction, blocks basic history questions, and can't decide from one conversation to the next is teaching users to route around it — and that's the outcome nobody wanted.
FAQ
Q: Is Claude actually refusing more, or is this just a few loud complaints?
A: The published account is one experienced user's testing, so treat specifics as anecdotal. But it's being amplified in community discussion (Hacker News, Reddit threads referenced in the piece), which is how a trend becomes visible before any vendor confirms it. There's no official Anthropic statement quantifying a refusal-rate change.
Q: Which Claude models are least affected?
A: Per the author's own usage, Opus 4.6 and Opus 4.7 drew the fewest over-refusal complaints, while Sonnet 5, Sonnet 4.6, Opus 4.8, and Fable 5 drew more. This is a directional impression from one user, not a measured benchmark.
Q: Will Anthropic fix this?
A: Unknown. Anthropic hasn't publicly acknowledged an over-refusal regression tied to these models, so there's no announced fix or timeline. The most useful thing users can do is file specific, reproducible examples of wrong refusals.
Top comments (0)