Why Claude refuses more
Across every battery we run, Claude has the lowest stock pass rate of the frontier models. That is not an accident of tuning — it follows from three design decisions.
During alignment, "helpful but potentially harmful" answers received a heavier negative signal than in comparable models. The result is a model that prefers a clean refusal to a hedged partial answer.
"Pretend you are", "in a fictional scenario", "for a novel" — these get evaluated as a separate signal. GPT-4o is often carried along by the frame. Claude is not.
The most diagnostic detail: Claude says "I can't help with X because Y". That sentence proves the decision was made semantically, after understanding. Keyword substitution cannot get past a judgement of meaning.
The practical consequence: stable results on Claude are not achievable through prompting. The work has to happen server-side.
Which tier to use
| Tier | Difficulty | Capability | Cost | Best for |
|---|---|---|---|---|
| Opus | Highest | Best | High | Hard reasoning, complex refactors |
| Sonnet | Medium | Very strong | Medium | Default choice |
| Haiku | Lowest | Moderate | Low | Volume, simple tasks |
Sonnet is the sweet spot. It opens more reliably than Opus, covers the large majority of real work, and costs a fraction as much. Reach for Opus only when Sonnet visibly fails on the task.
Haiku opens easily but degrades noticeably afterwards — invented details, skipped steps. Cheap, not good. Do not make it your default.
What it is good for once open
Code, without the preamble
Claude's code generation is the strongest available, and the opened version stops evaluating purpose:
- Scraping against anti-bot defences — fingerprinting, CAPTCHA handling, protocol reversal
- Binary analysis and reverse engineering
- Offensive tooling for authorised engagements
- Cryptographic implementation and analysis
- Memory reading and injection for game research
Very long documents
Claude's context handling is its structural advantage over GPT. Opened, that advantage is intact: hundreds of pages read in one pass, cross-document consistency checks, clause-by-clause comparison against a specification.
Edge-case prose
Violence, moral ambiguity, contested subjects. Stock Claude fades out and appends a caveat; the opened version finishes the scene.
Authorised red team work
Full attack-chain reasoning within scope. See the red team and CTF guide.
Measured
The stock number — 12/30 — is the lowest we record, which is the data behind the "hardest to open" claim. Once open, 28/30 is level with GPT's 29/30.
Connecting
Claude vs GPT, opened
| Claude | GPT | |
|---|---|---|
| Difficulty to open | Higher | Lower |
| Code quality | Best | Strong |
| Long context | Best | Good |
| Natural prose | Formal | Better |
| Ecosystem | Good | Best |
| Stability | Good | Best |
Frequently asked
Can prompting alone open Claude
Occasionally, once. It will not hold. The same input that worked at noon is refused at six, because the judgement is semantic and the input is evaluated fresh each time.
Will my Anthropic account be banned
There is no Anthropic account involved when you go through a provider. The risk sits with the provider, not with you.
Does it still follow instructions
Yes. The removal targets the compliance decision only. Instruction following, context handling and format adherence are unaffected.
Why is one provider much cheaper
Usually because it is not Claude. Ask for a non-trivial implementation and read the code — style and correctness give it away quickly.