# Public sources (accessed 2026-11-07)

Every number here comes from a public status page or pricing page, read on 2026-11-07. Nothing is estimated. If a figure couldn't be found, it says "not found."

## 1. Provider incidents, 2026-08-09 to 2026-11-07 (about 90 days)

| Provider | API-affecting incidents | Impact split | Published uptime | Filtering rule | Source |
|---|---|---|---|---|---|
| Anthropic | **29** (27 if two non-inference incidents are left out: delayed credits on Oct 1, Google connectors on Aug 20) | critical 2, major 19, minor 8 | Claude API component, 90 days: **99.77%** (60 days 99.87%, 30 days 99.88%) | `components[]` includes "Claude API (api.anthropic.com)" and `created_at` is on or after 2026-08-09. This matches the 29 events in the component's uptime bar. | https://status.claude.com/api/v2/incidents.json, https://status.claude.com/uptime_showcase?components=k8w3r06qmzrp |
| OpenAI | **18** incidents on the "APIs" component group (11 of them touched Chat Completions or Responses) | degraded_performance 16, partial_outage 2 (Images Sep 15, API Login Oct 20), full_outage 0 | APIs group: **99.97%**, labelled "Partial outage on 2 days. Degraded performance on 18 days." Degraded time does not lower this figure. | At least one component-impact record on the APIs group (14 components) in the window. The ChatGPT, Codex, FedRAMP and Ads groups are excluded. | https://status.openai.com/proxy/status.openai.com/component_impacts?start_at=2026-08-09T00:00:00Z&end_at=2026-11-07T23:59:59Z, https://status.openai.com/ |
| Google AI Studio / Gemini API | **3** (Oct 30 "Major outage affecting API key traffic": all API-key requests failing 17:12–19:48; Oct 1 replicated voices; Sep 14–16 Batch not finishing in 24 h) | not published | not found | Classified by title and description; the page has no component tags. | https://aistudio.google.com/status |
| Google Vertex "Gemini on Agent Platform" | 0 in the window (the most recent was 2026-02-27) | n/a | not published | Product history page | https://status.cloud.google.com/products/Z0FZJAMvEB4j3NbCJs6B/history |

**Example incidents:**
- Anthropic, 2026-10-23 (critical): "Elevated errors for Claude Fable 5.1, Sonnet 5.5, and Haiku 5.5"
- Anthropic, 2026-08-16 (critical): "Service disruption on Claude services"
- OpenAI, 2026-09-17: "We are seeing elevated error rates across API models"
- OpenAI, 2026-10-21: "Degraded performance in ChatGPT, Codex, and APIs"

**Caveats:**
- Each provider counts incidents differently: Statuspage for Anthropic, incident.io for OpenAI, and a hand-classified page for Google. Don't compare the counts across providers as if they were the same measure.
- Uptime percentages are aggregates the providers publish themselves.

## 2. List prices for hosted building blocks

**Usage priced:**
- 100,000 AI requests a month.
- One realtime message per request, delivered to one browser. Pusher, Ably and Supabase bill the publish and the delivery separately, so this is **200,000 billed realtime messages a month**.
- One webhook per request.
- 500 concurrent browser connections at peak.

### Realtime push to the browser

| Product and plan | Price per month | Covers the usage? | Source |
|---|---|---|---|
| Pusher Channels Startup | $49 (500 connections, 1M messages/day) | Yes, but 500 connections is exactly the cap. Pro ($99, 2,000 connections) is the first plan with headroom. | https://pusher.com/channels/pricing/ |
| Ably Free | $0 (500 MAU, 500 connections, 6M messages/month) | Only if there are 500 or fewer monthly active users | https://ably.com/pricing |
| Ably Pro | $29 (annual) / $35 at 1K MAU; $129/$155 at 5K; $249/$300 at 10K | Yes. The price depends on MAU. | https://ably.com/docs/platform/pricing |
| Supabase Pro | from $25 (500 peak connections, 5M messages) | Yes, at exactly the connection cap. Overage is $10 per 1,000 connections. | https://supabase.com/pricing |
| Firebase Realtime Database, Blaze | Pay as you go: connections not billed; $1/GB downloaded after 360 MB/day | Yes. The cost depends on payload size and can't be computed without it. | https://firebase.google.com/pricing |
| Firebase Cloud Messaging | Free | Not equivalent: it is mobile/web push notifications, not a live stream to an open page | https://firebase.google.com/pricing |

### Job queue / Redis

| Product and plan | Price | Covers the usage? | Source |
|---|---|---|---|
| Upstash QStash pay as you go | $1 per 100K messages, so **about $1/month** before retries (retries are billed) | Yes | https://upstash.com/pricing/qstash |
| Upstash Redis pay as you go | $0.20 per 100K commands, so $0.20 × N per month (N = commands per request) | Yes | https://upstash.com/pricing/redis |
| Redis Cloud Essentials 250 MB | $5 (no replication) / $7 / $9 (AWS us-east-1) | Yes | https://redis.io/pricing/ |

### Webhook delivery

| Product and plan | Cost at 100k webhooks/month | Source |
|---|---|---|
| Svix Basic | $20 + 50,000 × $0.0001 = **$25/month** (retries not billed; 99.9% SLA) | https://www.svix.com/pricing/ |
| Hookdeck Outpost Starter | 5 × $0.10 = **$0.50/month** (no SLA) | https://hookdeck.com/outpost/pricing |

**Cheapest plan that covers the usage, by category:**
- Realtime: Supabase Pro $25, or Ably Free $0 at 500 MAU or fewer.
- Queue: about $1.
- Webhooks: $0.50 (no SLA) to $25 (with SLA).

These are rental prices for the pieces only. They don't include the code that wires the pieces together, which is what the inventory in the README measures.

## Not found
- Pusher overage pricing, and what Pusher itself does at its limits.
- Ably per-message prices (Ably now prices by MAU).
- Google uptime percentages and severity levels.
- Per-model uptime for OpenAI or Google.

## 3. Sources added for the reliability and cost numbers

| What | Source | Notes |
|---|---|---|
| Anthropic incident windows | https://status.claude.com/api/v2/incidents.json | `started_at`→`resolved_at` for incidents tagged "Claude API (api.anthropic.com)"; feed covers back to 2026-08-05 |
| OpenAI incident windows | https://status.openai.com/proxy/status.openai.com/component_impacts?start_at=2026-08-09T00:00:00Z&end_at=2026-11-07T23:59:59Z and https://status.openai.com/proxy/status.openai.com (component names) | Chat Completions + Responses for the overlap; the 14 "APIs" components for the all-API figure |
| ModelRiver uptime | https://status.modelriver.com | per-minute health checks of the API and its database, 90-day window |
| ModelRiver prices | https://modelriver.com/pricing | Free $0 (25k requests), Pro $49 (500k), Scale $299 (5M) |
| US developer wage | BLS OEWS API, series OEUN000000000000015125213 (https://api.bls.gov/publicAPI/v2/timeseries/data/OEUN000000000000015125213) | Software Developers (SOC 15-1252), national median annual wage, May 2025: $135,980. bls.gov web pages refuse automated fetches, so the official API was used |
| Effort model | COCOMO II Model Definition Manual v2.1, USC Center for Software Engineering (1995–2000) | A = 2.94, B = 0.91, C = 3.67, D = 0.28; nominal scale factors 3.72, 3.04, 4.24, 3.29, 4.68 (sum 18.97) |