AI Cost Transparency

Last updated: 2026-02-21

OneCount Pty Ltd · ACN 695 536 415 · ABN 29 695 536 415

This page shows the arithmetic behind OneCount's AI unit economics. It is intentionally conservative and static: where the repository does not prove an exact token ceiling, the assumption is stated in the table instead of presented as a precise system limit.

Grounded implementation facts

The checked-in website Ask action calls the Supabase Edge Function implementation ai-intelligence with type: "business_query" from app/actions/ask.ts.
The Edge Function sets RATE_LIMIT = 200 and checks ai_usage for the prior hour before accepting a request.
business_query uses OPENAI_COMPLEX_MODEL and defaults to gpt-4o. Other structured ai-intelligence actions default to gpt-4o-mini.
The business-query runner allows up to four tool-call rounds plus a final answer. Each OpenAI call caps output at max_tokens: 1500, but the built prompt and tool results do not have a single exact input-token cap in code.
Public pricing below uses OpenAI list prices checked on 2026-07-05: gpt-4o at $2.50 per 1M input tokens and $10.00 per 1M output tokens; gpt-4o-mini at $0.15 per 1M input tokens and $0.60 per 1M output tokens.

Worst-case dashboard assistant COGS

The table models the highest-cost website path that is actually wired here: business_query on gpt-4o. The monthly figure uses a 30-day month. Free and Starter do not list AI features in lib/pricing.ts; Pro, Business, and Enterprise do. The Edge Function's rate limit is per organization, not per tier.

Tier	AI packaging shown in site pricing	Worst-case request assumption	Cost per request	Worst-case org cost per hour	Worst-case org cost per 30-day month
Free	No AI feature listed.	Not modelled as packaged access. If an active or trialing org is exposed to this endpoint anyway, the same technical cap applies because the function is not tier-specific.	-	-	-
Starter	No AI feature listed.	Not modelled as packaged access. If an active or trialing org is exposed to this endpoint anyway, the same technical cap applies because the function is not tier-specific.	-	-	-
Pro	AI-assisted invoice details for review and review cues for recorded variance.	Assumes max path: 5 OpenAI calls/request (4 tool rounds + final answer), 20,000 input tokens/call as a finance assumption, and the code-capped 1,500 output tokens/call.	((20,000 x 5) / 1,000,000 x $2.50) + ((1,500 x 5) / 1,000,000 x $10.00) = $0.325	200 requests x $0.325 = $65.00	$65.00 x 24 x 30 = $46,800.00
Business	Everything in Pro. The public plan copy does not separately promise website-assistant access.	Same technical limit as Pro. No separate per-tier AI rate limit was found in the Edge Function.	((20,000 x 5) / 1,000,000 x $2.50) + ((1,500 x 5) / 1,000,000 x $10.00) = $0.325	200 requests x $0.325 = $65.00	$65.00 x 24 x 30 = $46,800.00
Enterprise	Everything in Business. The public plan copy does not separately promise website-assistant access.	Same technical limit as Pro. No separate per-tier AI rate limit was found in the Edge Function.	((20,000 x 5) / 1,000,000 x $2.50) + ((1,500 x 5) / 1,000,000 x $10.00) = $0.325	200 requests x $0.325 = $65.00	$65.00 x 24 x 30 = $46,800.00

Fair-use policy caps (per tier)

Separately from the technical rate limit above, the checked-in implementation defines the following fair-use caps for AI business queries per organisation per rolling 30-day period. The repository path checks the organisation's tier and recorded usage before a model call. This is an implementation description, not independent verification of the configuration or deployment state in production. At the stated ~$0.325-per-query stress-test assumption, the configured caps imply roughly $42 (Pro), $81 (Business), and $163 (Enterprise). The separate 200 requests/hour limit is the checked-in technical cap; extrapolating it across every hour of a 30-day month is a stress-test assumption.

Tier	AI business queries / month (fair use)
Free	0 — AI not included
Starter	0 — AI not included
Trial	10
Pro	130
Business	250
Enterprise (private sales)	500

Structured AI calls

Structured ai-intelligence actions default to gpt-4o-mini, not gpt-4o. At the same token and rate assumption, the arithmetic would be: ((20,000 x 5) / 1,000,000 x $0.15) + ((1,500 x 5) / 1,000,000 x $0.60) = $0.0195 per request, or 200 requests x $0.0195 = $3.90 per org-hour. That comparison is shown only to separate the lower-cost default model from the website assistant path above.

Other OpenAI paths found

The repository also contains invoice and shelf-count Edge Functions. They are listed here for completeness, but they are not rolled into the tier table because their image/input token counts cannot be converted into a code-proven dollar ceiling from the repository alone.

Path	Model grounding	Rate limit grounding	Known code caps	Cost posture
Invoice text parsing	`invoice-ai` defaults `OPENAI_MODEL` to `gpt-4o-mini`.	`RATE_LIMIT = 200` requests/org/hour.	`MAX_INPUT_CHARS = 20_000` and `max_tokens: 4096`.	Exact tokens are not knowable from a character cap, so no precise dollar ceiling is stated.
Invoice vision parsing	`invoice-ai` uses `OPENAI_VISION_MODEL` and defaults to `gpt-4o`.	`RATE_LIMIT = 200` requests/org/hour.	`MAX_VISION_BASE64_CHARS = 20_000_000` and `max_tokens: 4096`.	Image-token usage depends on image detail and payload shape, so no precise dollar ceiling is stated.
Shelf-count vision	`shelf-count-ai` hardcodes `gpt-4o`.	`RATE_LIMIT = 100` shelf-count vision calls/org/hour.	Up to 5 images, `MAX_BASE64_BYTES = 6_000_000` per image, and `max_tokens: 4096`.	Image-token usage is not determinable from base64 byte limits alone, so no precise dollar ceiling is stated.

What is not proven by code

The repository proves the model defaults, website path, output-token cap, and per-org hourly request cap. It does not prove a hard input-token ceiling after business overview and tool results are assembled. Treat the $46,800 monthly figure as a deliberately conservative finance stress test under the stated 20,000-input-token assumption, not as a contractual maximum.