Skip to content
Multi-model consensus & judge — one API

Stop betting your output on one model.

Fan a single prompt out to top frontier models in parallel, then let a cross-family judge reconcile them into one answer you can trust. OpenAI-compatible. EU-hosted. You pay for the tokens you spend, plus a transparent consensus margin.

Live consensus, streamedOpenAI & Anthropic SDK drop-inEU-hosted · GDPR
What you get back

One prompt in. Every model's answer + a reasoned verdict out.

A consensus run, shaped exactly like the streamed response from the API and the free public demo.

POST /api/v1/chat/completions · model: tokonomix-consensus · mode: best_of
Prompt

We're a 12-person SaaS. Should we self-host an LLM or stay on hosted APIs for the next 12 months? Give a one-paragraph recommendation.

claude-opus-4.8412 tok

Stay on hosted APIs. At 12 people your bottleneck is product velocity, not inference cost; self-hosting adds GPU ops, eval and on-call burden that will not pay back below roughly €8k/mo in spend…

gpt-5.1388 tok

Hosted, with one caveat: route through a gateway so you can switch providers without code changes. Self-hosting only makes sense once a workload is both high-volume and latency-critical…

gemini-3-pro401 tok

Remain on hosted APIs for 12 months. Revisit only if data-residency or a >70% cost reduction at scale becomes a hard requirement; otherwise the operational tax outweighs the savings…

Judge · best_of

All three converge on "stay hosted", so confidence is high. The winner is rewarded for naming the real decision boundary rather than hedging, and for tying the choice to team size — the most actionable framing.

claude-opus-4.8 94gpt-5.1 89gemini-3-pro 86agreement: high
3 proposers + 1 judgetokens 1 601cost / call ≈ €0.033latency 5.4s

Illustrative example — synthetic data, not a customer run

Why one model is a risk

A confident wrong answer reads exactly like a correct one.

01

Single-model blind spots

One model means one set of blind spots. You can't see the disagreement you never sampled — and a confident hallucination looks identical to a fact.

02

Manual comparison doesn't scale

Pasting the same prompt into three chat windows works once. It does not work in a pipeline, a support flow, or an agent loop running thousands of times a day.

03

Glue code rots

Fan-out, streaming multiplex, a judge prompt, per-model cost accounting, provider failover — that is a system to maintain. Or it is one model slug.

Four synthesis modes

Pick how the answers get reconciled.

Same fan-out, different judge behaviour. Switch with a single field — no re-architecture.

Consensus

Synthesised agreement

default for "give me the best single answer"

Diff

Highlight disagreements

research, due-diligence, risk review

Best of

Judge picks the best response

pipelines that need one clean output

Raw

No synthesis, parallel only

you own the reconciliation

Drop-in

If you can call OpenAI, you can call consensus.

Change the base URL and the model slug. Add an x_council block to pick the pool, the judge, and the mode. That is the integration.

cURL — OpenAI-compatible
curl https://tokonomix.ai/api/v1/chat/completions \
  -H "Authorization: Bearer tok_live_…" \
  -d '{
    "model": "tokonomix-consensus",
    "messages": [{"role":"user","content":"…"}],
    "x_council": {
      "mode": "best_of",
      "models": ["claude-opus-4.8","gpt-5.1","gemini-3-pro"],
      "judge_model": "gpt-5.1"
    }
  }'
MCP — for agents (Claude, Cursor, …)
// tokonomix MCP server
tokonomix_consensus_ask({
  prompt: "…",
  models: ["claude-opus-4.8","gpt-5.1"],
  mode: "full",   // answers + reasoning + verdict
  judge_model: "gpt-5.1"
})

one tool call → multi-model verdict, billed to your Tokonomix balance

Vision — image consensus (new)
IMG=$(base64 -w0 product-shot.png)
curl https://tokonomix.ai/api/v1/chat/completions \
  -H "Authorization: Bearer tok_live_…" \
  -d '{
    "model": "tokonomix-consensus",
    "stream": false,
    "messages": [{"role":"user","content":[
      {"type":"text","text":"Inspect this image for publication-blocking defects. Verdict: DEFECT or CLEAN + findings."},
      {"type":"image_url","image_url":{"url":"data:image/png;base64,'$IMG'"}}
    ]}]
  }'

Inline base64 only (no URL fetch) · jpeg/png/webp/gif · max 5MB/image · non-streaming · vision models auto-selected

What you pay

No fee on single calls. You only pay the fee on consensus.

Ask one model and you pay just its tokens plus a small tier margin — no platform fee. The per-call fee applies only to multi-model consensus checks. 100 consensus checks free every month, no card needed; bundles from €10/month for 500 calls. Every token itemised, nothing hidden.

Single-model call

tokens + margin

No call-fee — only consensus checks carry the per-call fee. You pay the model provider's token price plus your tier margin (+2–5%). Example: a small model on ~4k tokens ≈ €0.001.

Consensus call

call-fee + tokens + margin

The call-fee varies per package (PAYG founders: 2c/proposer + 3c/judge, a 3+1 council = 9c; bundles: counts against your monthly quota; over quota: 1.5c/call). On top: the model provider's tokens + your tier margin.

Bring your own key (BYOK)

call-fee only

On consensus you pay only the per-package call-fee — your own key bills the provider directly, so no token cost and no margin from us. A single-model BYOK call costs nothing.

EU-hosted

Inference and logs in the EU. GDPR-aligned by default.

No lock-in

Standard SDKs. Swap providers without touching your code.

Pay-as-you-go

Prepaid balance. No seats, no minimum, stop anytime.

Observable

Per-model tokens, cost and latency on every call.

FAQ

Honest answers

Does multi-model consensus reduce hallucination?

It's a recall amplifier, not a truth oracle. Several models answer in parallel and an independent judge reconciles them, so an error one model makes is more likely to be caught and disagreement is surfaced rather than hidden. It reduces single-model error — it does not guarantee correctness, because frontier models share training data and can be confidently wrong together. Ground high-stakes facts.

How is the judge kept independent?

The judge is disjoint from the proposers and chosen from a different model family; the proposers answer blind and in parallel, so the judge never scores its own answer. This is a design property of how the council is built, not a benchmarked claim.

Can I keep my data in the EU?

Yes. You can pin a council to EU-hosted models by hosting region, so personal-data prompts are routed only to EU-hosted models. Region routing is per call.

Ship with consensus

Try it free. No sign-up to test.

Run consensus in the browser right now, then grab an API key when you are ready to wire it in.