Current limits and honest boundaries

API limitations

• Stable public compatibility is OpenAI-shaped Chat Completions, not the complete OpenAI platform.
• Responses, native Gemini, files, batches, fine-tuning, realtime, and legacy completions are not product-key surfaces.
• Native Anthropic Messages is provider-specific and does not translate the current NanoGPT/AIsa OpenAI-wire pool.
• Embeddings, image generation, audio, rerank, and moderation routes are registered but do not have a verified stable end-to-end product path.
• The developer message role and many newer Chat parameters are not forwarded.
• Full multi-turn tool-call compatibility is incomplete because assistant tool_calls are not preserved on the follow-up request.
• Structured output repair is best effort; applications must validate JSON themselves.
• Non-admin /v1/models is intentionally opaque and returns auto rather than the full internal registry.
• BYOK covers the upstream credential only. The account switch decides whether an attached key is used and detaching removes it, but Singular credit billing applies either way, and only providers you have attached a key for are affected.

Payment limitations

• x402 and MPP default off and require a live deployment configuration and healthy verifiers.
• Payment requests require an explicit model and non-streaming request.
• Native Anthropic Messages and Responses cannot be paid through these per-call paths.
• Current x402 EVM upto behavior needs alignment with the current Permit2-based specification before release.
• MPP imr* methods are Singular-specific extensions with no published first-party client in this repository.
• No current test proves a real wallet → live facilitator/on-chain settlement → inference → receipt round trip; unit tests mock external settlement.
• Per-call payment is not a key-balance or reseller-wallet top-up and has no documented automatic-refund guarantee.

App limitations

• Composer attachments are capped at 4 files per message and 3.5MB per file, and are limited to PNG, JPEG, WebP, and GIF images, PDF, and plain text, Markdown, or CSV. Other types are rejected, and a long extracted document is truncated before it reaches the prompt.
• An attached image needs a vision-capable model. Auto selects one; pinning a text-only model with an image attached can fail at the provider.
• Voice notes are capped at 90 seconds and are answered by an audio-capable catalog model, so a pinned text-only model cannot serve an audio prompt. A browser without microphone support cannot record.
• Only pinned Space files enter thread context. A stored file that is not pinned is kept for reference and is never injected.
• Media 'Regenerate' re-runs the original prompt as a new billed request and adds a branch sibling; it does not edit the existing media in place.
• Library pin state is not fully account-synced. Message likes are account-persisted and do survive reload and sign-in elsewhere.
• Funding history reflects the platform's payment records. Value added outside those records, such as a satisfied per-call payment challenge, is not listed there.
• Account export downloads JSON immediately and excludes asset file bytes; it is not a ZIP and is not delivered by email.
• Account deletion is synchronous and irreversible: it revokes API access and removes stored data before the sign-in identity is deleted, with no recovery path afterwards.
• Anonymous guest issuance requires Turnstile integration in production and should not be advertised as universally available until the active deployment includes that UI path.

Catalog limitations

The storefront feed is reliable for model ID, provider/name presentation, positive retail prices, and positive context values. Missing price, modality, BYOK, capability, and latency fields may be normalized or defaulted. Do not use them as definitive model capability evidence.

Release principle

Documentation follows verified behavior, not aspirational route names. A feature should move out of this page only after source wiring, contract tests, and a live smoke agree.