Post-fact billing architecture, unbilled safety contract, and API parity¶

A major architectural update to Nexus Hub: full transition to transparent post-fact billing, introduction of the protective unbilled contract for interrupted requests, audit payload redaction for privacy, and strict alignment with official OpenAI, Anthropic, and Gemini specifications.

New capabilities¶

True post-fact billing architecture¶

Balance debits now occur strictly post-fact — only after a generative request completes successfully with confirmed token counts returned by the model.
Removed upfront balance reservation and hold mechanisms, eliminating balance locks during concurrent or parallel requests.
All balance deductions are immutably recorded in the ledger with precise breakdowns for input, output, cache read, and cache write tokens.

Protective unbilled contract¶

Established a strict safety contract: if a request is aborted by the client, times out, or terminates with an ambiguous error without verified token counts, it is marked unbilled.
Incomplete, dropped, or ambiguous requests never trigger balance deductions, protecting users from paying for interrupted network calls.

Privacy protection and audit payload redaction¶

Enforced strict payload redaction in audit logs: user prompts, system instructions, chat messages, and model completions are never stored in the database in plain text.
Audit records store only anonymized metadata, request durations, completion statuses, and token usage metrics for operational diagnostics.

Native per-request pricing support¶

The model catalog and pricing engine now natively support models billed on a fixed per-request basis alongside standard per-token pricing.

Improvements¶

Strict API specification parity¶

Response schemas, HTTP status codes, and Server-Sent Events (SSE) streams are fully aligned with official OpenAI (Chat Completions, Completions), Anthropic (Messages), and Gemini APIs.
Seamless compatibility with external tools: verified reliable operation with Codex CLI, Claude Code, OpenCode, VS Code, Zed, Cursor, and custom bot frameworks.

Status monitoring synchronization¶

Public status page canary probes are synchronized with the active model lineup; retired models were removed from health checks.

Fixed issues¶

Protocols and streaming¶

Fixed occasional blank responses and schema mismatches in Responses Compact fields.
Fixed token usage finalization on early stream termination.
Eliminated spurious pricing errors during valid prompt caching fallback scenarios.

Pricing presentation¶

Corrected pricing displays in the bot and catalog for hybrid models, preventing token-based and per-request rates from mixing.

Existing API keys, balances, and base URLs remain valid and active without modification.
Details regarding post-fact billing and the unbilled contract are documented in pricing and balance.
Consult the supported endpoints and API reference for endpoint details.