Pay what inference costs.
Plus 20%.
No black-box per-turn fees. No tier gymnastics. We pass through your underlying provider costs at-cost and keep a flat 20% for the orchestration.
Built for developers who think in tokens-per-dollar, not "turns." If your run cost us $0.40 in upstream API calls, you pay $0.48. The $0.08 is what we keep for routing across 22 frontier seats + tool dispatch + trace capture + receipt.
your_cost = underlying_inference × 1.20
No subscription required for usage. Pay only for what you run.
⬢ Live cost estimator
Underlying inference cost
$—
Your cost (cost + 20%)
$—
MirzaTech keeps
$—
Per-run cost to you
$—
FREE TIER
Sandbox
$0
No credit card. No commitment.
- Watch the live 22-seat scripted Parliament demo
- Read public Council deliberation transcripts
- API endpoint for the scripted simulation (zero LLM cost · zero throttle)
- Persona vocabulary: Proponents · Skeptics · Polygeists · Chancellor
USAGE
Pay-As-You-Go
cost + 20%
No floor. No ceiling. No surprise charges.
- Real Parliament runs (22 seats · 5 rounds · agentic tools)
- Real Council annotations (vetted experts · per-step critique)
- Tool layer included: web_search · file_io · python_sandbox
- Per-run cost receipts with full token + provider breakdown
- Receipt-style billing · pay only what you run · no minimum
DASHBOARD
Workspace
$9 /mo
Per workspace · unlimited seats included
- Persistent workspace dashboard for your team
- Saved sessions · search history · receipt archive
- Shared Council expert pool for your team's annotations
- Per-workspace API key (rotation + audit log)
- Usage cost on top per the cost+20% model above
PARTNER
Provider Allocation
credit swap
For LLM provider partnership teams only
- You allocate API credits to MirzaTech's cost-base
- We provide structured agentic-trace + Council annotation bundles
- Per-provider isolated dashboard (raw JSONL export · signed URL)
- Quarterly performance reports for your DevRel + enterprise teams
- Outlier-failure analysis (where N-1 frontier models pass and yours fails)
⬢ Why cost-pass-through?
Frontier model providers charge you per token. So do we — at the actual underlying provider rate, with a flat 20% added for orchestration. This means: if upstream provider prices drop, your bill drops automatically. If a model gets faster + cheaper through caching or speculative decoding, you keep the savings. If we rotate your workload to a cheaper provider that performs equally well, your bill drops, our margin stays at 20%.
The 20% covers: (a) routing across 22 frontier seats, (b) tool dispatch (web_search · file_io · python_sandbox), (c) trace capture + receipt + audit log, (d) Council annotation surface for vetted experts, (e) a sustainable orchestration business so we're around in 12 months when you renew.
Every receipt shows the full token + provider + cost breakdown. No hidden margin. No black-box "turns." If you're priced out, we are too — that's the alignment.
⬢ Pricing FAQ
What about the old $2/turn $4/turn pricing?
Retired 2026-05-11. The "turn" was an opaque cognitive frame that didn't match how developers think about LLM cost. The new model is transparent token economics. Existing $99 PARLIAMENT_REPORT and $5–$199/seat subscriptions remain available for customers who prefer fixed-fee · contact us if you're on the legacy plan.
How do you calculate the underlying provider cost?
Each provider (NVIDIA NIM · Groq · OpenRouter · Novita · etc.) publishes per-token rates. We compute exact cost per call using the provider's published rate × actual tokens consumed. The receipt shows per-seat: model_id · input_tokens · output_tokens · provider · unit_rate · total. You can audit every cent.
What if my run uses tools (web_search · python_sandbox)?
web_search uses Gemini grounding which is free (Mo's existing key allocation). file_io is server-side (no upstream cost). python_sandbox runs on Modal — usage is included in our 20% margin. Tools never add a per-call line item to your bill.
Can I see the cost before I run?
Yes. Every endpoint returns an estimated cost band (low / median / high) before execution. You can set a per-run cost cap; if the run would exceed it, we abort and charge nothing.
What happens to model names? I see "Frontier Tier 1 · 235B-class" instead of model IDs.
Public surfaces show only persona-class labels — we don't expose the seat-to-vendor mapping to anonymous traffic. Once you authenticate, your receipt and dashboard show the full model_id for every seat that answered you. Transparency for paying customers · opacity for the public.