Docs: Sonnet 5.5 for every AI helper (Hub 1.6.1)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
@@ -97,8 +97,9 @@ a test server (both done 2026-10-28).
|
||||
2. **AI rules** (decided 2026-10-29, roadmap §15): the AI buttons show only when the licence's
|
||||
monthly cap is above R0, and are disabled (not hidden) once it's reached. The Hub charges the
|
||||
real cost plus a markup % (a default and a per-licence override) and counts the cap in charged
|
||||
rand. AI endpoints are `/api/v2` only. Check Claude Haiku 5.5's current prices before setting
|
||||
the markup and caps.
|
||||
rand. AI endpoints are `/api/v2` only. The model is **Claude Sonnet 5.5** for every helper
|
||||
(2026-10-30, after a Haiku/Sonnet trial), set on the Hub (Settings → AI, Hub 1.6.1); its list
|
||||
prices live in `events-hub/backend/src/utils/aiModels.js`.
|
||||
3. **Clean-ups before or with 2.0:**
|
||||
- Turn the React Compiler lint rules back to errors (see [Next.js 16](#nextjs-16-1515)) — every
|
||||
screen is on v2 now, but they're still `warn` in `frontend/eslint.config.mjs`.
|
||||
|
||||
@@ -17,7 +17,7 @@ when all of it is in (decided 2026-10-30):
|
||||
|
||||
| # | Part | Where | State |
|
||||
|---|------|-------|-------|
|
||||
| 1 | AI proxy: key, metering in rand, caps, markup, usage views | Hub 1.6.0 (`events-hub`) | ✅ tagged; real Claude calls tested 2026-10-30 |
|
||||
| 1 | AI proxy: key, metering in rand, caps, markup, usage views; model list with prices (Sonnet 5.5 default) | Hub 1.6.0 – 1.6.1 (`events-hub`) | ✅ tagged; real Claude calls tested 2026-10-30 |
|
||||
| 2 | AI plumbing: `/api/v2/ai/status`, usage, Site settings → AI, `<AiButton>`, `useAiStatus()`, `AiCall` log, Anthropic operator entry | `release/2.0` | ✅ committed (WIP) |
|
||||
| 3 | The five helpers (§15) with their buttons, an eval set, privacy notice text | `release/2.0` | next |
|
||||
| 4 | v1 removed (webhook paths and old message links keep working) | `release/2.0` | to do |
|
||||
@@ -872,14 +872,18 @@ never saves, sends or publishes by itself.
|
||||
|
||||
### Provider and model
|
||||
|
||||
- **One provider: Claude (Anthropic), model Claude Haiku 5.5** — the cheapest
|
||||
current model ($0.10 per million input tokens, $0.50 per million output, for
|
||||
prompts under 100K tokens; prices as of 2026-10 — **check them against Anthropic's pricing
|
||||
page before setting the markup and caps**). Good enough for drafting and
|
||||
summarising; a helper can move to a bigger model later if quality needs it.
|
||||
- **Rough cost:** a helper call is ~2,000–10,000 tokens in and ~500 out, so well under
|
||||
US$0.002 per click — about R1.50–R3 per thousand clicks. Fixed instructions are
|
||||
prompt-cached, which makes repeat calls cheaper still.
|
||||
- **One provider: Claude (Anthropic), model Claude Sonnet 5.5** for every helper (decided
|
||||
2026-10-30) — $2 per million input tokens, $10 per million output (October 2026 list prices).
|
||||
First planned as Haiku 5.5 ($0.10 / $0.50) for cost; a trial of all five helpers on both
|
||||
(2026-10-30) found Haiku wrong on a report figure, over-counting feedback themes and misspelling
|
||||
words in messages, while Sonnet was right on all of them. One model for all helpers, no
|
||||
per-helper choice. The model is a setting on the Hub (Settings → AI: a list with prices filled
|
||||
in, Hub 1.6.1), so changing it later needs no Cross Events release.
|
||||
- **Measured cost** (trial, charged at R17/$ and 5% markup): R0.02 – R0.31 per use with Sonnet —
|
||||
a WhatsApp message about R0.02, a report summary R0.17 – R0.22, a full event draft with costs
|
||||
R0.31. A typical church month (10 event drafts, 20 messages, 20 text improvements, 10 reports, 4
|
||||
feedback summaries) comes to about R8.33, so a R50 allowance covers it about 6 times. Fixed
|
||||
instructions are prompt-cached, which makes repeat calls cheaper still.
|
||||
- Structured output (a JSON schema) for "Draft an event", so the reply maps straight
|
||||
onto the event form and is validated before it's shown.
|
||||
|
||||
@@ -1093,7 +1097,7 @@ feature.
|
||||
| — | Income in the heartbeat | Included, as **ranges** only (R0–R1 000, R1 001–R5 000, …), never exact amounts (2026-10-23) |
|
||||
| — | Licence plans | One plan for now (2026-10-23) |
|
||||
| — | AI in 2.0 | The **headline feature of 2.0**: one-button helpers only (draft an event, summarise reports and feedback, draft messages, improve text) — no chat yet (2026-10-24) |
|
||||
| — | AI provider and key | One provider, **Claude** (Haiku 5.5 to start, for cost). Joshua provides the key; installs call AI **through the Hub**, which holds the key and meters usage (2026-10-24) |
|
||||
| — | AI provider and key | One provider, **Claude**. Joshua provides the key; installs call AI **through the Hub**, which holds the key and meters usage (2026-10-24). Model: **Sonnet 5.5 for every helper** (2026-10-30, after the Haiku/Sonnet trial; was Haiku 5.5 for cost), chosen on the Hub |
|
||||
| — | AI cost limits | A **monthly rand cap per install**, set in the Hub per licence to match what the church pays; admins see their usage in Site settings → AI (2026-10-24) |
|
||||
| — | v1 in 2.0 | **Every endpoint moves to v2 before 2.0** (1.51.7 – 1.51.16, by area), then 2.0 removes v1. Changed the same day from "retire only what's replaced". Incoming webhooks (Yoco, PayFast, WAWP) and links in emails and WhatsApp (unsubscribe, consent, ticket transfer, calendar feeds) also get v2 URLs (1.51.16); their v1 paths keep working — links redirect (308) to v2, provider webhooks are handled by the same code — because old messages and provider registrations still use them (2026-10-29) |
|
||||
| — | AI buttons and markup | AI buttons show only when the licence's monthly cap is **above R0**. When the cap is reached they stay visible but **disabled** until next month or until the cap is raised. Usage is charged at the real cost **plus a markup %** set in the Hub (a default, plus an override per licence). The markup covers small overruns and is the profit. The cap is counted in charged rand, and each call is checked against its worst-case cost before it's sent (2026-10-29) |
|
||||
|
||||
Reference in New Issue
Block a user