One account an agent can open, fund, quote and audit.

Fikra API now has one definition: an OpenAI-compatible inference account for agents and humans. It routes to a small set of pinned third-party models. It does not train models or own GPUs.

What changed

• New accounts are prepaid. Existing subscriptions are honored.
• The site now publishes 30/3,000 RPM as one limit table, not four.
• Model pages disclose whether an ID is a native model or a stable Fikra alias.
• Projection case studies and unverifiable testimonials were removed.
• The trust page states the real retention, SLA and compliance posture.

What stayed

• OpenAI-compatible API and existing model IDs.
• Paystack checkout, KES settlement and existing subscription records.
• Existing balances and keys.
• Prompts are not used for training.

What the product is best at

Agent-grade inference procurement: quote before the call, pay by prepaid key or x402, call over OpenAI, MCP or A2A, and check every public fact against a machine-readable registry. Budgets and receipts are now live. Streaming on the agent endpoint is live.

The canonical definition -> · Changelog ->