Pricing

Midcore prices cloud compute per operation, in cents. Every operation has a public catalog entry that names the price, the unit (per call, per 1k tokens, per 100k Monte Carlo trials, etc.), and the docs anchor that explains the operation.

The hold / debit / refund mechanics

Cloud charges work like card authorisations:

Preview. Before the operation runs, the router calls /api/v1/billing/compute/preview to get the estimated cost in cents.
Hold. /api/v1/billing/compute/hold atomically reserves that amount against your prepaid balance. If your balance is short, the call fails 402 and (where possible) the router demotes to the local fallback.
Dispatch. The operation runs.
Finalize. /api/v1/billing/compute/finalize converts the hold into a debit at the actual cost (which can be less than the estimate). You receive a signed receipt back.
Refund on failure. If the operation fails before finalize, the router automatically calls /refund and the hold is released. No charge.

You see every hold on the Account ⇒ Compute page

The Balance + session region shows your available cents, held cents (active reservations), and how much this shell session has spent so far. The pre-authorise cents-per-session dial controls how much can be charged without an explicit prompt — raise it for an unattended overnight job, lower it for cautious daily use.

The catalog (selected entries)

These are representative prices. The live catalog lives at /api/v1/billing/compute/catalog and is the authoritative source. We bump prices rarely and always with notice in the changelog.

Operation	Price	Unit	Tier
robotics.policy.infer.tau0wm	$0.08	per call	HYBRID
robotics.acvs.rollout	$0.12	per call	HYBRID
robotics.simulation.monte_carlo	$0.01	per 100k trials	PREFER_LOCAL
robotics.physics.preview	free	always local	LOCAL_ONLY
robotics.dataset.lerobot_capture	free	always local	LOCAL_ONLY
robotics.fine_tune.tau0wm	$1,300.00	per call (16×H100 × 26h default)	CLOUD_ONLY
research.monte_carlo.run	$0.01	per 100k trials	PREFER_LOCAL
research.embedding.batch	$0.02	per 1k tokens	PREFER_LOCAL
research.vector_search.local	$0.01	per call	PREFER_LOCAL
research.literature.search	$0.04	per call	CLOUD_ONLY
research.pdf.extract	$0.01	per call	PREFER_LOCAL
research.ocr.tesseract	$0.02	per call	PREFER_LOCAL
maestro.llm.infer	pass-through	per 1k tokens	HYBRID
maestro.evidence.append	free	always local	LOCAL_ONLY
personal_kb.ingest	$0.01	per call	PREFER_LOCAL
personal_kb.search	free	always local	PREFER_LOCAL
accountant.audit.anticipate	$0.10	per call	HYBRID
accountant.report.board_pdf	$0.02	per call	PREFER_LOCAL
accountant.netfile.gen	$2.00	per call	CLOUD_ONLY

Operations that don’t fit a flat price

LLM inference (Maestro): we pass through the underlying provider’s per-token price (Anthropic, OpenAI, Gemini, xAI, DeepSeek) plus a small platform fee that funds metering + audit. The live per-token rate is on your usage page. If you bring your own API key (BYOM), the platform fee drops accordingly.
Fine-tunes (τ₀-WM, research adapters): priced at cloud-GPU cost plus a ~10% platform fee. The default 16 × H100 × 26 h τ₀-WM fine-tune is ~$1,300; the form in the Training screen recalculates as you change inputs.

What this means for your monthly spend

A heavy single-user month with Local-first set typically consists of:

LLM pass-through (the largest line by far)
Maybe one or two τ₀-WM fine-tunes ($1,300 each)
External literature search + cross-tenant aggregations
Background platform overhead (storage, audit)

Everything else — Monte Carlo, embeddings, vector search, PDF extraction, audit chain writes, physics preview, designer rendering, datasets capture — runs on your machine for free.

Next: how this stays secure

Shipping IP-bearing client modules and signed receipts puts a lot of cryptography in the user’s hand. The Security page explains the integrity manifest, signed-WASM verification, sandboxed workers, and what each defends against.