FAQ
13 questions, answered the way we'd want them answered.
Including the ones with answers that cost us something. Every claim below resolves to a page on this site — terms, privacy, security, or the audit, which lists every dated, sourced claim we make.
Looking for how to do something rather than what is true about it? The guides are the step-by-step version — one task per article, with the exact on-screen labels.
The product
What is smart fusion routing?
The measure-route-re-measure loop behind Keenoble. Every answer logs which model handled it, what it cost, how fast it was, and the outcome. That evidence re-ranks the 335-model library on quality per krona, speed per task, and reliability. When your prompt arrives it is classified — everyday question, code, long document, deep reasoning — and sent down the cheapest lane that can answer it well; where the best answer isn't a model at all (a stock price, the weather, a cited source), the router calls a live tool instead of guessing with an LLM. We publish this method and our findings; we don't publish the current routing configuration.
Which models does Keenoble include?
Keen Labs measures a 335-model library (counted, not rounded, growing weekly) — every Claude, every GPT, DeepSeek, Kimi, Qwen, Grok, Gemini, Llama, and Perplexity's live-search Sonar, plus models nobody markets. 12+ are in active daily rotation at any time; the router promotes and demotes based on evidence, not brand. The full library is included in every plan — no tier hides models behind it.
How is this different from ChatGPT, Cursor, or Lovable?
Those tools each route within one model family or ship one surface — a chat product, a code editor, a build tool. Keenoble routes across every vendor's models, not just one company's, and Chat, Preview and Code share one memory (working, project, source, and mission memory), so what you say in Chat is already known when you open Preview or Code. Vendors that sell a model also have a reason not to rank a rival model honestly; Keen Labs buys every model, so honest ranking is the business, not a conflict with it.
When do image, video and music models arrive?
They are catalogued in the lab and not shipped. The plan is credit add-ons on top of Premium rather than a separate subscription. They are marked "coming" everywhere they appear on this site, and nothing we claim today counts them — including the model library figure, which counts models measured by the catalogue pipeline, not media generation you can buy.
Have you published a benchmark I can check?
Yes, as of 2026-08-17. The first qualification sweep ran every model in the v1 roster across eleven task domains — one task set, identical conditions, graded by deterministic checks (executed code, parsed JSON, matched values, detected language) wherever one applies, and by a blinded jury only for open-ended reasoning. Pass rate, cost per task, latency and reasoning overhead are published per model at /research#results, and the raw export the numbers come from is committed in the repository as data/findings-public.json so you can check the page against it. Three caveats we state on the page rather than bury: it is our benchmark and not an independent one; the depth is uneven, so some rows rest on around 120 scored tasks and others on 11, and each row shows which; and it is one run, which makes it directional rather than statistical. The failure findings at /research#failures are arguably more useful than the pass rates — including two models we route to that answer benign requests with an empty HTTP 200.
Money
What are Sparks?
Sparks are Keenoble's usage credit: what a chat reply, a preview step, or a code edit costs to run, priced by which model or tool actually answered. Premium (€49/month) includes 1,000 Sparks. Measured against our own production cost log, everyday answers that were charged anything averaged 0.1148 Sparks, which makes the allowance about 8,708 of them; the single dearest answer we have logged cost 0.8 Sparks, which would make it 1,250. That range is the honest answer, and /pricing shows the division rather than asserting a result.
What happens when I run out of Sparks?
You can buy a Spark pack; unlike the monthly allowance, purchased packs stay until you use them. Unused monthly Sparks expire at the end of each billing period. What does not happen is a silent downgrade to a cheaper model to stretch your quota — and you don't have to take that on trust, because the receipt under every answer names the model that produced it, so a swap would be visible on the answer itself.
Can I cancel anytime?
Yes — Premium is a monthly plan with no lock-in, cancelled from your own subscription settings in the Stripe customer portal, effective at the end of the current period. No forms, no phone calls, no retention flow. The Pro tier is designed and unreleased, with pricing not announced (see /pricing for its "coming" status); Keen OS Command, the custom workforce tier, is deployed in scoped engagements rather than sold as a subscription.
Do you give refunds?
On your first subscription purchase you have the 14-day right of withdrawal under Swedish distansavtalslagen; because the service starts immediately at your request, a withdrawal in that window is refunded minus a proportional deduction for the Sparks you already used. Renewals are not new contracts and carry no withdrawal right — cancelling stops future billing. Spark packs are delivered instantly at your request and are non-refundable once delivered, which you acknowledge at purchase. Full text on /terms; disputes can also go to ARN.
Data and privacy
Where does my data live? Is it hosted in the EU?
Keen Labs is a Swedish company (Landskrona) and GDPR applies to everything. Your account data, projects and memory sit in Postgres on Supabase behind row-level security scoped to your account; authentication is Clerk, payments are Stripe, hosting is Vercel. What we will not claim is EU-only processing: when you send a prompt it goes to whichever model provider answers it, and that includes providers outside the EU/EEA, in the United States and in China. Transfers rest on those providers' own safeguards. Plainly said — if you don't want a prompt processed by a given country's provider, don't put personal data in prompts. The current subprocessor list is on /security.
Do you train models on my prompts?
No. We do not train our own models on your content, we do not sell personal data, and we run no advertising. We do keep one telemetry row per model call — which model, token counts, cost, latency, success or failure — because that measurement is what makes the routing better. It is measurement of the service, not profiling of you, and anything we publish from it is aggregated: never your prompts, never your content. Your prompts are still processed by the model provider that answers them, under that provider's API terms.
What don't you publish, and why?
Three things. The live routing configuration — we publish findings and method, but not the rules themselves, which is a deliberate limit on how far you can verify us from outside; the findings are the pitch and the config is the moat, and pretending otherwise would be its own dishonesty. Customer prompts and content — published findings are aggregated, always. And anything we cannot source: a claim with no record and no date does not go on this site, which is why /research#audit lists every claim that is on it, along with what we explicitly do not claim.
The lab
Who is behind Keen Labs, and how do I get access?
One person: Keen Andersson Lang, sole trader (enskild firma) in Landskrona, Sweden, with the registration number on every invoice and on request. The lab has zero paying customers — the products are live, the revenue is not, and we say so rather than implying traction we don't have. You can start on Keenoble free, without a card. Pilot access is five seats: one month of Premium, free, with a direct line to the founder — request one from /beta or the form on /pricing.
Having problems with the platform? Contact us at support@keenlabs.pro. Something missing here, or an answer that doesn't match what the product actually does? Same address — it gets fixed in public. Step-by-step instructions live at /guides; why the lab works this way is on /mission.