Pricing Skrongle AI

What a learner pays for the AI Skrongle provides on its own key (Skrongle AI, features), and why the server never loses money on it. Every number here is a default in src/metering.rs. You can change it in the environment (deploy/env.example).

Self-hosted servers have no plans. With SKRONGLE_PLANS=off (what deploy/skrongle.service sets) there are no prices or payments: every account has every feature, and the server’s AI key, when the owner sets SKRONGLE_AI_KEY, is shared by every account with no monthly allowance. The owner pays the provider’s bill; SKRONGLE_AI_DAILY_CAP and SKRONGLE_AI_RATE_PER_MINUTE bound it, and ai_usage still records every call. Everything below is about the hosted site, where plans are on.

The numbers

	Price	Skrongle AI included
Free	$0	none: Free brings its own keys, or buys credit packs
Basic	CA$3 a month (`billing::Plan`)	none: like Free, Basic brings its own keys or buys credit packs (#931)
Plus	CA$15 a month (`billing::Plan`; a Stripe subscription or a grant is Plus too)	$1.50 of list-price AI a month (`SKRONGLE_AI_ALLOWANCE_PLUS=1.50`)
Sign-up credit	free, once	`SKRONGLE_AI_ALLOWANCE_FREE` (default $0) of AI as credits, given when a new account first confirms its address; not monthly (#931)
Credit pack	$5 once (a one-off Stripe price), sold on every plan	$3.00 of AI, spent after Plus’s allowance, never expires (`SKRONGLE_AI_CREDIT_PACKS=price_…:3.00:$5`)

Guards: $10 of Skrongle AI a day across everyone (SKRONGLE_AI_DAILY_CAP), 12 Skrongle AI requests a minute per account (SKRONGLE_AI_RATE_PER_MINUTE). Superusers see the day’s spend against the cap on /superuser; learners never do.

The models and their list prices

Anthropic’s list prices, checked 2026-10-19 (dollars per million tokens, standard global routing):

Model	Used for	Input	Output	Cache read	Cache write (5 min)
`claude-haiku-5-5`	the tutor (`SKRONGLE_AI_MODEL_TUTOR`)	0.10	0.50	0.01	0.125
`claude-sonnet-5-5`	writing and revising courses (`SKRONGLE_AI_MODEL_COURSES`)	2.00	10.00	0.10	2.50
`claude-opus-5-5`	the default for learners’ own Anthropic keys	4.00	20.00	0.20	5.00
anything else (`*`)	an unlisted model	10.00	50.00	1.00

The meter charges each call at exactly these prices from the usage the provider reports, so a dollar of allowance is a dollar of Anthropic’s bill. Where the bill could be higher, the meter rounds up:

Cache writes cost 1.25 input tokens; the meter counts them so.
Unknown models (another provider, a new model) are charged at the dearest rate until SKRONGLE_AI_PRICES lists them.
Haiku 5.5 over 100,000 prompt tokens costs five times as much. A tutor prompt is the system prompt plus the last 24 entries of at most 1,000 characters each, about 10,000 tokens at most, so it never gets there. An operator who moves course writing to Haiku should price it at the higher rate (claude-haiku-5-5=0.50/2.50/0.05), since course material can be long.
US-only inference (1.1 times) is never requested.

What $1.50 buys

From the meter’s own estimates (ai::job_tokens, tutor::REPLY_MAX_TOKENS):

A tutor turn on Haiku: about 4,000 tokens in and 250 out, $0.0005; at the most (10,000 in, 700 out) $0.0014. $1.50 is 1,000 to 3,000 turns.
A lesson on Sonnet: about 10,000 tokens in (the course format reference plus the request) and 6,000 out, $0.08; the outline $0.05. A 3-lesson course is about $0.29, a 6-lesson one $0.53. $1.50 is about five 3-lesson courses. Real lessons usually come in under the estimate.

The margin

Per paid learner per month, if they use every cent of their allowance:


Subscription	$5.00
Stripe’s fee (2.9 % + $0.30, a US card)	− $0.45
Skrongle AI at list price, at most	− $1.50
A job’s last request may run past the allowance (one lesson at most: 16,000 output tokens on Sonnet)	− $0.18
Left, at worst	$2.87 (57 % of the price)

A credit pack: $5.00 − $0.45 − $3.00 − $0.18 = $1.37 (27 %) at worst. Free learners cost nothing: they are never offered Skrongle AI. Learners on their own keys cost nothing either (their calls are recorded at $0).

The server checks before every call, so a turn or a job can only start while the learner has allowance (a job’s estimate is held up front). The only overshoot is a job’s last request, above. The daily cap bounds a bug.

Why one key and our own meter

Anthropic’s Admin API was checked (2026-10-19) for a key or a workspace per learner:

Workspaces: 100 per organisation by default (archived ones don’t count). That is not enough for one per learner, and each would need its own spend limit, which you set in the Console, not through the API.
API keys: the Admin API lists and updates keys but does not create them. You make keys in the Console (Claude Code also mints them in its own workspace). The app could not issue one per learner.
Usage and cost reports: usage shows up “typically within 5 minutes”; the cost report is daily buckets only; polling once a minute is the sustained limit. That is too late to stop a learner before a call, which is when the allowance has to be checked.
Prompt caches are per workspace, so splitting learners would also lose the cache.

So the server keeps one key (SKRONGLE_AI_KEY) and meters each learner in ai_usage (src/metering.rs), checking before every call. The check reads running totals, not the rows: ai_usage_months (per learner and UTC calendar month) and ai_spend_days (everyone, per UTC day), which triggers on ai_usage keep equal to the sums (#623, #624).

Reconciling with Anthropic

Once a month (not built yet; TODO(reconcile) in src/metering.rs):

Put the server’s key in a workspace of its own, and make an Admin API key.
GET /v1/organizations/cost_report?starting_at=<1st>&ending_at=<1st of next>&group_by[]=workspace_id with the Admin key: that workspace’s cost for the month.
Compare it with SELECT SUM(cost_micros) FROM ai_usage WHERE key_source = 'hosted' AND created_at >= <1st>.

A meter that reads lower than the bill means the price table is out of date or a model is missing from it. Fix SKRONGLE_AI_PRICES before the next month.