inference-bot
2eb88f55b4
Audit of the operator docs against live reality (2026-11-04): - app-inventory.md: all 19 Cloudron apps with id, fqdn and registry image pin. Corrects the portal path (code/member-portal, not co-op/member-portal) and the panel path (local/coop.inference.admin-panel, not code/admin-panel). - maintenance/backups.md: the two backup systems; the deliberate per-user restic cache split and the --group-by "" retention fix, with why. - maintenance/woodpecker-ci.md: CI server + agent, image build/pin, manual-only trigger, git-lfs requirement, and its current not-yet-green state (agent's job-collection long-poll gets HTTP 500 from the server). - provider-pricing/: index + GreenPT, Tinfoil (local EHBP proxy architecture) and the Metis /v1/systemone decision-model route; regolo.md updated to record the gateway prices were applied. - agent-fleet.md: steward, infrastructure-bot and socialbot (Unix users, gateway units, teams); member-bot tier 2 planned but not yet provisioned. - README: file list; operator-token-runbook: correct the membership_sweep cadence (3x daily, not nightly); migration doc: post-migration addendum (16->19 apps, DNS split resolved, old box still on with apps stopped).
118 lines
5.9 KiB
Markdown
118 lines
5.9 KiB
Markdown
# Regolo pricing: what changed, and where our numbers live
|
||
|
||
Source: Regolo's announcement email (link goes to their SSO portal, so the tables below are
|
||
entered by hand) plus their public pricing page.
|
||
|
||
## Effective 2026-11-01 (from the notice email), EUR per 1M tokens
|
||
|
||
| model | input | output |
|
||
|---|---|---|
|
||
| gpt-oss-20b | €0.10 | €0.20 |
|
||
| qwen3.5-9b | €0.10 | €0.20 |
|
||
| brick-complexity-pro (SEMANTIC SMART ROUTER) | €0.10 | €0.20 |
|
||
| gpt-oss-120b | €0.20 | €0.70 |
|
||
| mistral-small-4-119b | €0.20 | €0.70 |
|
||
| gemma4-31b | €0.40 | €0.60 |
|
||
| qwen3.8-27b | €0.50 | €3.40 |
|
||
| **qwen3.5-122b** | **€0.60** | **€3.40** |
|
||
| **apertus-70b** | **€1.00** | **€3.40** |
|
||
| glm5.2 | €2.30 | €6.90 |
|
||
|
||
## Before (public pricing page, read 2026-10-31), EUR per 1M tokens
|
||
|
||
apertus-70b €0.40 / €2.10 · qwen3.5-122b €1.00 / €4.20 · gpt-oss-120b €1.00 / €4.20 ·
|
||
gpt-oss-20b €0.10 / €0.42 · gemma4-31b €0.40 / €2.10 · glm5.2 €2.00 / €5.20 ·
|
||
mistral-small-4-119b €0.50 / €2.10 · qwen3.8-27b €0.50 / €2.10 · qwen3.5-9b €0.07 / €0.35 ·
|
||
brick-complexity-pro €0.10 / €0.40 · brick-v1-beta €0.00 / €0.00
|
||
|
||
Image generation, unchanged by the notice: **€0.000000015 per pixel** (with the
|
||
subscription note: Qwen-Image is *not included* in any plan).
|
||
|
||
## What we serve, and the change it needs
|
||
|
||
Peg: EUR x 1.16 -> USD, the rate our numbers already use.
|
||
|
||
| our model | old USD/M | new USD/M (from 2026-11-01) | direction |
|
||
|---|---|---|---|
|
||
| regolo/apertus-70b | $0.46 / $2.44 | **$1.16 / $3.94** | input +152%, output +62% |
|
||
| regolo/qwen3.5-122b | $1.16 / $4.87 | **$0.70 / $3.94** | input -40%, output -19% |
|
||
|
||
**Until the gateway is updated**, apertus bills $0.46/M input while costing ~$1.16/M — a ~70 cent/M
|
||
loss on the co-op — and qwen over-bills members by ~46 cents/M input. Both directions argue for
|
||
updating promptly rather than waiting.
|
||
|
||
## Where the numbers live
|
||
|
||
The gateway's per-model costs are the authority, and they are what the response headers
|
||
(`x-litellm-response-cost`, shipped 2026-10-31) and member billing both read. Read them with
|
||
`GET https://gateway.inference.coop/model/info` (master key): as of 2026-10-31,
|
||
`regolo/apertus-70b` = 4.6e-07 / 2.44e-06 and `regolo/qwen3.5-122b` = 1.16e-06 / 4.87e-06
|
||
per token. Member-facing copy is `docs/models.md`.
|
||
|
||
## Open inconsistency: the image model is priced in EUR-as-USD
|
||
|
||
`regolo/qwen-image` carries **no** input/output cost, and a live 512x512 call billed
|
||
**$0.00393216** — which is 262,144 px x EUR 0.000000015, i.e. the EUR figure applied as if it
|
||
were dollars. The two chat models convert at 1.16; the image model does not, so it is ~15.5%
|
||
under cost. Small per image (~$0.0007), but it is the same class of error as the chat prices and
|
||
should be fixed in the same pass (x 1.16 -> $0.0000000174 per pixel).
|
||
|
||
## Process note
|
||
|
||
Keep a dated section per notice. The next change should be a diff against this file rather than a
|
||
re-derivation — the email link is a login wall, so the numbers only reach us when the operator
|
||
forwards them.
|
||
|
||
|
||
## Our actual Regolo usage (measured 2026-10-31, from LiteLLM SpendLogs)
|
||
|
||
Source: `"LiteLLM_SpendLogs"` in the gateway's Postgres, reached with python + `psycopg` inside the
|
||
gateway container (there is no `psql` there). Models were first used 2026-10-16.
|
||
|
||
| model | requests | spend | window |
|
||
|---|---|---|---|
|
||
| regolo/qwen3.5-122b | 195 | $2.61 | 2026-10-16 → 10-31 |
|
||
| regolo/qwen-image | 94 | $0.96 | 2026-10-16 → 10-31 |
|
||
| regolo/apertus-70b | 157 | $0.82 | 2026-10-16 → 10-31 |
|
||
| **all Regolo** | **452** | **$4.38** | last 30 days; ~all of it |
|
||
|
||
Context: total gateway spend is **$426.34** lifetime (104,947 requests, since 2026-09-04), and
|
||
**~$393 in the last 30 days** — greenpt/glm $231.42, deepseek $146.14, gpt-oss $10.32,
|
||
Regolo **$4.38**, other $0.79. So **Regolo is ~1.1% of inference spend, about $4.38/month.**
|
||
|
||
## Should we take a subscription?
|
||
|
||
**Almost certainly not at current usage.** A plan has to beat ~$4.38/month to be worth it, and
|
||
Regolo's plan prices are **not published** — the pricing page has a SUBSCRIPTIONS section with no
|
||
figures, which is why the announcement email links into their account portal. Worth one short
|
||
question to them via "Talk to an engineer", not a project.
|
||
|
||
One thing that is public and useful: their **30-day free trial** grants **1M tokens/day** across
|
||
"All Core Models". Our entire Regolo usage is roughly **120k tokens/day** (≈3.6M/month), i.e. well
|
||
inside that, so the trial — and likely a low tier — would cover everything we currently do.
|
||
|
||
## The 2026-11-01 change, in money terms
|
||
|
||
At these volumes the mispricing is small: apertus's input roughly 2.5x-ing on a model that has cost
|
||
**$0.82 lifetime** is about **$1/month** of exposure, and qwen's cut moves ~$0.6/month in members'
|
||
favour. So this is **billing accuracy, not urgency** — but accuracy still matters, because the same
|
||
numbers feed the cost headers members can now read.
|
||
|
||
## UPDATE 2026-11-04 — the gateway was updated; both stale sections above are now resolved
|
||
|
||
Checked against the live gateway config (`/app/data/config.yaml`, `model_info`):
|
||
|
||
- `regolo/apertus-70b` = **1.16e-06 / 3.944e-06** per token
|
||
($1.16 / $3.94 per 1M) — the new 2026-11-01 rate. The "~70 cent/M loss while the
|
||
gateway still bills $0.46" paragraph above no longer applies.
|
||
- `regolo/qwen3.5-122b` = **6.96e-07 / 3.944e-06** ($0.70 / $3.94 per 1M) — the new
|
||
rate. The "over-bills members by ~46 cents/M" paragraph no longer applies.
|
||
- `regolo/qwen-image` now carries **`input_cost_per_pixel: 1.74e-8`** — the EUR
|
||
figure converted at ×1.16, i.e. the "image model is priced in EUR-as-USD" open
|
||
inconsistency above is **fixed**. It is no longer priced at the raw EUR number.
|
||
|
||
So the two "Until the gateway is updated" tables and the image-model section above
|
||
are **historical**. The usage measurement ($4.38/month, ~1.1% of inference spend)
|
||
and the subscription analysis stand. Keep this file dated-section-per-change; do
|
||
not delete the history, it shows the reasoning.
|