Where Claude sits in my systems
SwellChat is my 24-hour assistant for business websites. One engine serves many tenants. Each tenant has its own knowledge base and a guardrail prompt, and the engine picks a model per tenant from an environment setting. Since August 8, 2026, demo and prospect tenants run Claude Sonnet 4.6 as the primary model. A salon tenant runs an open model on Groq first and falls back to Claude Haiku 4.5 when Groq times out or errors. That same tenant uses a Claude vision model to read promo flyers the owner uploads, so the assistant can quote this week's specials without anyone typing them in.
Three client systems run in their own protected spaces on my server, each with its own Anthropic key. One answers product questions for a manufacturer. One is the salon. The third is a meeting copilot.
Smaller builds use Haiku. Every Rep, a fundraising app I built on October 28, 2026, calls claude-haiku-5-5 through the official SDK. Lime, a menu-bar tool that turns my spoken review notes into a clean list, sends transcripts to Haiku. Humanize keeps Haiku as its last fallback.
One demo assistant answers in five languages from a single English knowledge base: English, Korean, Chinese, Japanese, Tagalog. I did not write translations. Sonnet handled it, and it passed 7 of 7 test probes before I sent the link.
What it costs
| Model | Input per 1M tokens | Output per 1M tokens | Cache hit per 1M |
|---|---|---|---|
| Claude Fable 5.2 | $8 | $40 | $0.20 |
| Claude Opus 5.5 | $4 | $20 | $0.20 |
| Claude Sonnet 5.5 | $2 | $10 | $0.10 |
| Claude Sonnet 4.6 | $3 | $15 | $0.30 |
| Claude Haiku 5.5 (prompts up to 100K tokens) | $0.10 | $0.50 | $0.01 |
| Claude Haiku 4.5 | $1 | $5 | $0.10 |
The Batch API takes 50 percent off both directions if you can wait for results. Web search costs $10 per 1,000 searches on top of tokens.
For a grounded chat assistant the input side dominates. My tenant prompts run 8,000 to 12,000 tokens once the guardrail and retrieved knowledge are in. On Sonnet 4.6 at $3 per million, a 10,000-token prompt costs 3 cents before the model writes a word. A cached copy of the same prompt costs 0.3 cents. That gap is why prompt caching matters more than model choice for this kind of work.
An idle demo tenant costs me about $1 to $2 a month in API spend. Real conversations are cheap. Testing is what burns money, and I learned that the hard way.
The outage: August 22, 2026
At 19:06 GMT every assistant I host started answering "[assistant unavailable]". The paying salon went down with the demos. Three sales pitches had dead demo links sitting in prospects' inboxes.
The chain went like this:
- Around August 21, Groq retired the Llama model the salon tenant ran first. Every salon message fell through to Claude.
- I had built 31 demo tenants that month, and each build ran a live test suite against the API.
- Those two together hit a $50 monthly spend limit on my Anthropic organization. The API began returning errors that said access would come back September 1.
- I had given the salon its own API key, thinking that isolated it. It did not. The key lived in the same organization, so it shared the same limit.
I raised the limit to $100 the next morning and everything came back without a restart. The assistants were dark for about five hours.
Three things changed after that:
- Tenant builds and tests run offline against the assembled prompt. One live smoke test of 5 probes runs at deploy, which costs about 30 cents per tenant.
- Each paying client runs in its own space on my server, with its own key.
- A fallback carrying full traffic is an alarm. Mine had done it for a day before anything told me.
The spend limit is a good feature. Set it on purpose for each client. Then wire an alert to the same phone that gets your sales calls.
If you only use the Claude app
The API is billed by the token and is what the notes above cover. The Claude app is billed per seat. On November 14, 2026 Anthropic lists Pro at $20 a month, Max at $100, and Team at $25 per seat monthly with a 2-seat minimum. I run my own work on Max. A team that only needs people chatting with Claude does not need the API or me. You need the API when Claude has to answer your customers or work inside your own software.