Lab status: workforce rebuilding on measured foundations — qualification program runningread the audit →
KEEN LABS

Keenoble · the flagship

You don't pick the model. That's the whole idea.

Chat, Preview and Code over one shared memory, with every prompt routed across a 335-model library — counted, not rounded — to whatever the evidence ranks best for that particular question. Plus live tools no chatbot alone can offer: web, market data, weather, news, cited on the receipt.

Join the pilot programme →Read the benchmark

01 · Chat

Where ideas form. Smart-routed across the full library — the router picks in milliseconds and you never choose a model from a dropdown.

02 · Preview

Where an idea becomes something you can look at. Live, deployed, and already holding the context you set in Chat.

03 · Code

Where you get surgical. Repo-connected, working from the same project memory rather than from a pasted snippet.

ROUTER · TRACE REPLAYSAMPLE

01 / 05

What's the capital of France?

speed · closed-book fact

The cheapest lane that can be right is the right lane.

Ruled out

  • claude-opus-5capable, and absurd for this — frontier price, no frontier need
  • kimi-k2.6a 200k context window for eight words
  • sonar-prosearch costs a round trip to learn something already known
Groq Llama 3.384ms0.1 sp

02 / 05

Tesla stock price?

live data · needs a source, not a memory

A model would have guessed. A tool call knew, and charged nothing.

Ruled out

  • gpt-oss-120bwould answer confidently from stale training data
  • sonar-prowould find it, and bill for the search
  • claude-opus-5the most expensive way to be out of date
finnhub · tool call142ms0 sp

03 / 05

Generate a TypeScript type for this JSON

code · structured output, small scope

Specialist beats generalist when the task has a right answer.

Ruled out

  • deepseek-v4-prothe deep code tier, for a job that isn't deep
  • llama-3.1-8bfast enough, not reliable enough on exact syntax
  • claude-opus-515× the cost to produce the same interface
Kimi K2.7-Code260ms0.4 sp

04 / 05

Summarise this 200-page document

long context · the whole thing has to fit

Sometimes the constraint is the window, not the intelligence.

Ruled out

  • gemini-3.6-flashcheap, and would have to read it in pieces
  • kimi-k2.7-codewrong specialist — this is prose, not a repo
  • groq llama-3.3the fast lane cannot hold the document
Kimi K2.6890ms2.1 sp

05 / 05

Analyse our Q3 strategy tradeoffs

deep reasoning · multi-step, no single right answer

This is what the frontier is for. It gets used, and it gets billed.

Ruled out

  • gemini-3.6-flashwould produce something fluent and shallow
  • deepseek-v4-proexcellent at code, not the tier for this
  • groq llama-3.3speed is not the scarce resource here
Claude Opus 52.1s3.8 sp

Five representative routing decisions, replayed — not a live feed and not a captured incident log. Figures are the dated snapshot in our scripted replay (2026-08-08). Candidates are ruled out in words because we have not published per-model scores for these prompts, and we are not going to invent some to make the panel look busier.

Memory

Say it once, in any room.

A project is a folder of linked memory nodes — an index map, decisions, preferences, constraints, files. Chat writes to it, Preview reads from it, Code reasons over it. Retrieval follows a path to the two or three nodes that matter instead of shovelling the whole history back into a context window. Pick a room.

PROJECT MEMORY · SAMPLE MAP — nordic-saas/
index.md · mapfoundation/stackdecisions/dark-themepreferences/brandconstraints/budgetfiles/pricingfiles/authartifacts/deploysource memory · mission memory — designed, not shipped
You say in Chat: “Brand colour is emerald, keep the tone dry and Nordic.”
wrote preferences/brand · linked from index.md

Retrieval is a path to two or three nodes, not a dump of the whole memory — that is the difference between shared memory and stuffing a context window, and it is why the receipt says 0.4k tokens instead of 42k. Sample project, drawn to show the mechanism.

Receipts

Every answer tells you what it cost and why.

Model, latency, price, and the reason that model was chosen — attached to the answer, not buried in a usage dashboard you have to go and find. If a run fails its gate, it doesn't bill.

RECEIPTSAMPLE

What's the capital of France?

routed to
Groq Llama 3.3
because
speed
latency
84ms
charged
0.1 sp

Smallest lane that can be right. Nothing was gained by spending more.

RECEIPTSAMPLE

Tesla stock price?

routed to
finnhub · tool call
because
live data
latency
142ms
charged
0 sp

A tool call, not a model. Nothing to bill, so nothing was billed.

RECEIPTSAMPLE

Generate a TypeScript type for this JSON

routed to
Kimi K2.7-Code
because
code
latency
260ms
charged
0.4 sp

Code specialist. Cheaper and better than the frontier at this exact job.

RECEIPTSAMPLE

Analyse our Q3 strategy tradeoffs

routed to
Claude Opus 5
because
deep reasoning
latency
2.1s
charged
3.8 sp

The frontier, earned. This is the receipt you should see least often.

Every answer the platform returns carries one of these. The four above are samples built from our dated method table (2026-08-08) — real mechanism, illustrative contents. We are not going to paste a customer's prompt on a marketing page to prove we log it.

Compare

Three honest ways to buy AI.

Every cell below describes how an approach is built, not how well it performs — we have not published a head-to-head benchmark, so nothing here claims one. Two of the rows go against us.

How Keenoble compares with a single frontier subscription and with running your own API keys
 One frontier subscriptionYour own API keysKeenoble
Which models you can reachOne vendor's line-upAnything you wire yourself335 measured, 12+ in rotation
Who picks the model per promptThe vendor — it only sells oneYou, every timeThe router, on evidence
What a greeting costsFrontier rates, or a quotaWhatever you chose to callThe cheapest lane that can be right
Memory across chat, preview and codePer-product, if at allWhatever you buildOne project memory, three surfaces
Live data (markets, web, weather)Sometimes, opaquelyAnother integration to ownTool calls, cited on the receipt
Receipt under every answerNoYour own logs, if you wrote themModel, latency, price — always
Incentive to rank models honestlyNone — the rival is the competitorYours, and it's your timeIt's the business model
Published head-to-head benchmarkVendor's own, on vendor's tasksNone unless you run itYes — 54 models, 11 domains, published in full. Ours, not independent
Track record you can checkYears of itn/aMonths, and one published audit

Model vendors sell one model each — ranking a rival honestly means losing a customer to it. Keen Labs buys every model, so honest ranking is the business rather than a threat to it. That is a structural argument, not a result; the method and the dated benchmark example are on /research.

Honest limits

Six reasons you might not want this yet.

Everything above is the argument for Keenoble. This is the cost of admission, stated by us, before you find it yourself.

  • No published head-to-head benchmark. Not one.

    We say the router picks on evidence, and the qualification sweep that produces public, same-test results has not published yet. Until it does, the routing table you see on this site is a dated snapshot of representative decisions — labelled as such everywhere it appears.

  • Our autonomous workforce is paused, by us.

    It reported success it had not earned for three months. We caught it, stopped it, and published the whole thing rather than quietly patching it. Read the audit. It is the least flattering document on this website and the main reason to believe anything else on it.

  • Keenoble has 0 paying customers.

    Active since 2026-05, used daily by its builder, and not yet validated by anyone's money. If a logo wall is what you need before you try something, this is not that.

  • Pro is not shipped. Neither is media generation.

    Pro is designed and unreleased — pricing not announced. Image, video and music models are catalogued in the lab and will arrive as credit add-ons. Both are on the site as coming, and neither is counted in anything we claim today.

  • Two of the four memory layers are drawings.

    Working memory and project memory are shipped and are what the map on this page animates. Source memory and mission memory are designed — the map draws them dashed and says so, because a roadmap rendered in the same ink as a feature is a lie with good typography.

  • We will not show you the routing config.

    We publish findings and method; we do not publish the live routing rules. That is a deliberate limit on how far you can verify us from outside, and pretending otherwise would be its own dishonesty. The findings are the pitch; the config is the moat.

Pilot · five seats

Still here? Take a seat.

One month of Premium free, with a direct line to the founder. Premium is €49/month afterwards, cancel anytime.

Five seats. No card, no spam, one email when your seat opens.