Fikra API platform metrics and how to benchmark your own workload
Platform-wide counts and uptime are published at fikraapi.co.ke/stats and /status. They answer volume and reliability questions. They do not answer which route wins your workload, only your own prompt set does that.
Route-level request counts, cache-hit rates and 30-day uptime are published at /stats and /status. These are the raw numbers. To know which route wins for you: freeze 50 prompts from your own traffic, run them at temperature 0, score accepted-task cost, keep the winner.
What the platform publishes and what it means
fikraapi.co.ke/stats publishes: all-time request count, 30-day request count, prompt tokens, cached tokens, completion tokens, active routes, and per-model counts where the route has more than 100 requests in 30 days. These are ledger-derived aggregates, not estimates.
fikraapi.co.ke/status publishes: 30-day uptime percentage, component health and incident log. Enterprise teams can contact support for contract terms.
What they do not prove: which route wins your task, latency from your network, or quality for your prompt set. A route with 60,000 requests is popular, not automatically right for you.
What Fikra does not publish
No per-route latency percentile (p50, p95, p99). Any page claiming these for a Fikra route is inventing them.
No quality score derived from traffic volume.
No synthetic leaderboard presented as production evidence.
Uptime is published as a percentage over health checks, not as an SLA guarantee.
How to run your own benchmark
Pull the live route list at /v1/models and pin the route IDs you plan to test.
Take 50 prompts from your real production traffic. Include 10 that are hard or failure-shaped.
Run them at temperature 0 through two or three routes. Log request ID, model, tokens, cache hits, cost and wall time.
Score accepted tasks per dollar from your own network. The route with the lowest accepted-task cost wins.
Publish the full distribution to your team. One screenshot of one win is not a benchmark.
Re-run when /models lists a route revision.
curl https://api.fikraapi.co.ke/v1/chat/completions \
-H "Authorization: Bearer YOUR_FIKRA_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"deepseek-v4.1-flash-1017","messages":[{"role":"user","content":"Summarise this ticket in two lines."}]}'
What counts as an accepted task
Accepted-task cost = wallet spend divided by responses that complete the task to your standard. Define your standard before you run the test, not after. A common failure: scoring token count instead of task completion, then discovering the cheapest route misses 20% of tasks.
Payment friction belongs in the benchmark for African teams
If you are in Kenya, the checkout cost is published and exact: $5 minimum charges KES 684.95 and credits KES 665 (KES 133 = $1 checkout rate; the KES 19.95 is the mobile provider's 3% fee). In other markets, your card issuer sets the conversion rate and margin. A declined 3-D Secure attempt costs round-trip time regardless of token price. These belong in your total cost of inference.
Common questions
Are the platform numbers a guarantee for my workload?
No. They describe platform-wide production traffic and health. Run your own 50-prompt test for your workload; that answer is the only one that settles it.
Why do some routes show no per-model numbers?
Below 100 requests in the trailing 30 days, per-model metrics are withheld as 'not enough traffic' rather than estimated. That is a reporting floor, not an error.
Where do these figures come from?
The public usage snapshot at fikraapi.co.ke/stats, derived from backend ledgers, and the public status at fikraapi.co.ke/status. Methodology and retention notes are at /methodology.