DeepSeek served directly. The cheapest routes in the world, even cheaper than the official DeepSeek API.
Three pinned builds, all on a 1,048,576-token context, priced to be the cheapest available route, not bundled into a mid-tier plan. One OpenAI-compatible endpoint, DeepSeek-style automatic prompt caching billed at a fraction of fresh input.
Open-model inference you can audit.
The cheapest DeepSeek routes in the world, even cheaper than the official DeepSeek API, priced so an agent or an operator can compute the bill before it exists, with the evidence published. No subscriptions, no card gates, no silent model swaps.
Price
Cheaper than the official DeepSeek API: pinned DeepSeek builds at $0.045 to $0.05 per 1M input, cached input at $0.025, and the lowest DeepSeek output rate of any live route.
See the DeepSeek line →Agent rails
x402 lets an agent pay per request with no signup, no card and no human in the loop. MCP and A2A are built in. Same /v1 endpoint, same rates for agents and keys.
Agent gateway docs →Cost control
A paid quote endpoint returns what a call will cost before you run it. Machine-readable pricing.md, a public calculator, and a prepaid wallet. The meter stays auditable.
Cost calculator →Evidence
Live usage and uptime as JSON, published traffic rules, a public benchmark for every model (failures included), an OpenAPI spec and llms.txt. Check everything.
Public benchmarks →| Model | Input $/1M | Output $/1M | Cached input $/1M | Context |
|---|---|---|---|---|
| deepseek-v4.1-flash | $0.05 | $0.11 | $0.025 | 1.0M |
| deepseek-v4.1-flash-1017 | $0.05 | $0.11 | $0.025 | 1.0M |
| deepseek-v4-flash-0731 | $0.045 | $0.17 | $0.025 | 1.0M |
Prices refreshed from the live model catalog. Machine-readable rates: pricing.md.
Observed market floors, October 2026
The table below compares listed Fikra rates with the cheapest provider floors on major gateways on October 20, 2026. The Fikra column is the same registry that answers GET /v1/models, the pricing page and the docs: one price everywhere, verifiable with a single request. My comparison numbers move as competitors move, so run your own spot check via the quoted source before you commit a workload.
| Model | Fikra in / out | Cheapest observed gateway floor |
|---|---|---|
| deepseek-v4.1-flash-1017 | $0.05 / $0.11 | $0.10 / $0.60 |
| deepseek-v4-flash-0731 | $0.045 / $0.10 | $0.05 / $0.16 |
| deepseek-v4.1-flash | $0.045 / $0.09 | $0.036 / $0.14 |
Floors change weekly. Fikra passes most cache savings to you, which the headline price rows on most gateways do not show.
Cached input at $0.025/M.
DeepSeek builds reuse the longest stable prefix of your prompt automatically. Cached tokens bill at $0.025/M against the $0.045-0.05 fresh input rate, which is why code agents, RAG pipelines and long system prompts get the most out of the line: the parts you repeat cost 50% or more less every turn.
No flags to set. Keep your system prompt stable and the cache does the rest. Cached and fresh tokens are itemized separately in your usage logs.
from openai import OpenAI client = OpenAI( base_url="https://api.fikraapi.co.ke/v1", api_key="your_fikra_key" ) response = client.chat.completions.create( model="deepseek-v4.1-flash-1017", messages=[...] )
Which build should I pick?
deepseek-v4.1-flash is the day-to-day flagship of the line: the latest stable schedule. deepseek-v4.1-flash-1017 is the October 17 build with better reasoning, coding and vision numbers and lower task cost. deepseek-v4-flash-0731 is the pinned older build kept for API stability.
How is cached input billed?
Cached tokens bill at each model's cached rate (see the table above). Fresh and cached tokens are priced separately and itemized in both the usage response and your billing history.
Do you keep the old builds around?
Yes. The DeepSeek line is served from pinned builds with all three listed as long-term offerings. If a build ever has to retire, it will be announced on the feed with a sunset date rather than pulled.
What can I pay with?
Cards, USDC (x402, on Base) and mobile money. Agents can pay per request over x402 with no signup at all, against the same endpoint. The prepaid wallet top-up starts from $1.
Is there a free way to test?
The free-flagship tier has retired and fikra-free-deepseek now bills at DeepSeek-line rates, and every new account activates with a first top-up, minimum $1. Your prompt, not ours, budgets the test.
Where does traffic run?
Fikra is operated by Lacesse Ventures, routed through a hardened upstream grid of verified providers, with documented retention rules at /.well-known/traffic-rules and live usage/uptime at /usage. DeepSeek ZDR handling is available for compliance-sensitive workloads.
Run the whole replay set on /v1/batches.
50,000 DeepSeek chat requests in one JSONL upload, OpenAI-shaped, billed at the same $/1M rates as interactive, with cached input still automatic on batch lines. Evals, labeling runs and backfills fit in one upload.