DeepSeek V4 Flash 0731: price, context and 30-day evidence

Model card (live registry, as of 2026-11-14)

KES figures use the checkout rate of KES 133 = $1 from the live checkout facts (as of 2026-11-14). Your card issuer sets its own conversion rate and margin; Kenya checkout charges KES 136.99 for a $5 top-up after the 3% fee.

The machine-readable catalogue is at /v1/models; the public card with the same context windows and rate limits is at /models. Prices move with provider terms, so check both before a production commitment.

30-day production evidence

As of 2026-11-14, /stats reports 200 requests for deepseek-v4-flash-0731 in the last 30 days, with 0.0% of prompt tokens served from cache across 346,436 total tokens.

Per-route latency percentiles and tokens-per-second are not published; Fikra does not estimate them. Run your own timing from your own network.

1. Enough traffic for per-model reporting means more than 100 requests in the trailing 30 days; below that the page reports the route as unmeasured.
2. No latency percentile or score for this route is published anywhere on this page; compare accepted tasks on your own prompt set.
3. Platform uptime and incidents are published at /status; as of 2026-11-14 the 30-day figure is 98.95% over 88,390 checks with status reported as attention.

What it costs on your workload

10M input tokens and 2M output tokens cost about $0.90 in wallet credits at the listed rates (KES 120.03 at KES 133 = $1), before any cache discount.

Input price is not the bill. Cache-hit rate, output length and retry count decide wallet cost. A stable prompt prefix that hits the cache at the listed cached rate changes the arithmetic more than the headline input price does.

Cap max_tokens to what the answer needs. Verbose drafts are the silent budget leak on output-priced routes.

1. Estimate: monthly spend = (input_tokens x $0.05625/1M) + (output_tokens x $0.17/1M), minus any cache discount.
2. Score cost per accepted task, not cost per token. A route that fails one task in five is not cheaper at any sticker price.
3. Compare in KES at the checkout rate for Kenya, or at your issuer rate elsewhere, not in headline dollars.

How to run it now

Set base_url to https://api.fikraapi.co.ke/v1 in any OpenAI-compatible client, use model=deepseek-v4-flash-0731, and replace YOUR_FIKRA_KEY with a key from the dashboard. The key is shown once; store it somewhere safe.

Start with a short max_tokens and temperature 0 for classification, extraction and comparisons. Only raise temperature if your task needs variation.

1. POST https://api.fikraapi.co.ke/v1/chat/completions with the body below.
2. The example uses deepseek-v4.1-flash-1017 (63,202 of the last 30 days' 116,189 requests as of 2026-11-14); replace the model field with deepseek-v4-flash-0731 to call this one.
3. Keep the request ID from the response if you need to reconcile the wallet charge.
4. Before production, run at least 50 of your own prompts across clean, cached and failure cases, and log cost per accepted task.
5. Confirm deepseek-v4-flash-0731 is listed at /v1/models for your account before you commit traffic to it.

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."}]}'

How this route compares

For text work, compare accepted-task cost against deepseek-v4.1-flash-1017 ($0.05 input and $0.15 output per 1M; 63,202 requests in the last 30 days as of 2026-11-14) and glm-5.3-flash ($0.1875 input and $0.625 output; 207 requests, 60.79% cache-hit rate).

For hard reasoning, compare against deepseek-v4-pro-pinned. For retrieval, pair any chat route with fikra-embed. The winner is the route with the lowest wallet spend per accepted task on your own prompt set, not the one that graphs best.

1. Freeze a 50-prompt set from your real traffic, including 10 failure-shaped prompts.
2. Log model ID, request ID, tokens, cache hits and total cost for every call.
3. Keep the pair that wins on accepted-task cost. Re-run when /models lists a revision.
4. If this route drops off /v1/models, stop posting production traffic to it.

Before you commit budget

Fikra self-serve is prepaid with no SLA. The balance is the hard cap; calls stop when funds run out rather than overcharge. Lite, Pro and Max are $10, $25 and $60 per 30 days with 1, 2 and 5 keys; every tier allows 3,000 requests per minute and 250,000 per day.

If your production contract needs a committed uptime or response time, talk to support before you migrate traffic, not after.