xpersona 3rd party models (gpt/fable) are small/micro oss models

[ashleymcdonald]

as per title, the models listed as gpt and fable are not.

Fable

• Listed on models.dev/providers/xpersona at $3/$18 per MTok
  (official Fable 5 pricing is $10/$50; below Anthropic's own batch floor of $5/$25)
• Langfuse trace of a request to their endpoint, model field "claude-fable-5", shows:
  • usage reports reasoning_tokens (169) additive with output (271) — OpenAI-style
    accounting; Anthropic includes thinking in output_tokens and reports no separate
    reasoning count
  • input_cached_tokens = 13312 (exactly 13×1024) — consistent with OpenAI/vLLM
    128-token block caching, not Anthropic's exact-prefix cache_read counts
  • no cache_control breakpoints were set client-side, yet cache reads occurred
  • response quality inconsistent with a Mythos-class model (fails the canonical
    car-wash reasoning probe)

related pr #3044

[github-actions]

I've reviewed the issue and the relevant files. This is not an actionable model catalog data fix — it needs maintainer review, and I did not open an automated fix.

Why this needs maintainer review

The issue alleges that xpersona is misrepresenting its models — that the "claude-fable-5" and "gpt-5.5" listings are actually small/micro OSS models rather than the genuine Anthropic/OpenAI models they claim to be. The supporting evidence is forensic (Langfuse token-accounting traces, cache-block arithmetic, pricing-below-floor arguments, and a quality probe).

This is a provider-integrity / policy determination, not a concrete metadata correction:

• No specific verified field correction is requested. The issue doesn't say "change price X to Y" or "fix limit A to B" with an authoritative source backing it. It asks the maintainers to conclude the provider is fraudulent and act on that conclusion. The appropriate "fix" (removing models? delisting the provider? changing pricing?) is a judgment call, not a mechanical data update.
• The central claim cannot be verified against provider documentation. The allegation is that xpersona's own documentation/API is deceptive. There is no first-party authoritative source that states "xpersona serves fake models" — the evidence is analytical and requires human interpretation and trust decisions.
• Pricing-below-floor is suspicious but not dispositive. A reseller pricing below a batch floor could indicate fraud, but could also reflect subsidies, loss-leaders, or volume deals. Concluding fraud from this is a policy call, not something I can safely encode as a data correction.
• The issue references related PR Add Claude Fable 5 model for Xpersona #3044, indicating an ongoing maintainer discussion already exists around this provider.

This falls squarely into "policy/product discussion" territory. Deciding whether to delist, flag, or otherwise act on a suspected fraudulent provider is a decision only the maintainers can make. I've left all TOML files unchanged.