Claude API

What it is

Claude is an AI model from Anthropic, a competitor to OpenAI's ChatGPT and Google's Gemini. The Claude API lets a website send a prompt and get back a generated response. For chatbots, content generation, structured data extraction, code analysis, anything where you would otherwise hire a human to read text and respond.

The Outerwave use case: a qualification chatbot. A visitor lands on a Corepay solution page, opens the chat, asks "what does this cost?" Claude answers using the page's content as context, qualifies the lead (monthly volume, current processor, chargeback rate), and hands off to sales with a clean summary. Cold form-fill becomes warm intro.

How we use it

Haiku for the chatbot, Sonnet for content

Three Claude models matter in production work:

• Haiku 4.5 ($0.25 per million input tokens, $1.25 output). Fast, cheap, smart enough for qualification chats and quick content tasks. This is the workhorse.
• Sonnet 4.6 ($3 input, $15 output). Quality jumps significantly. We use it for content drafting, page analysis, lead scoring, anywhere a wrong answer would be expensive.
• Opus 4.7 ($15 input, $75 output). Overkill for production. We use it during development for hard reasoning tasks, then prompt-engineer the work down to Sonnet for ship.

Prompt caching is non-negotiable

A useful chatbot needs to know about your business: who you serve, what you sell, how you qualify, your pricing posture. That context goes in front of every message, and without caching you pay full price for it every turn. With caching, the context gets billed once per session and every follow-up is dramatically cheaper. Cache hit rates run high in production, which is what keeps per-conversation cost in the noise.

Streaming the response

The model's response is streamed to the visitor's browser as it's generated rather than waiting for the whole reply to finish. Perceived latency drops from "several seconds of nothing happening" to "starts replying in a fraction of a second," which is the difference between an engaged user and a bounced one.

Structured handoffs, not chat transcripts

Once the model has enough information to qualify a lead, it produces a structured record (name, email, intent, the key qualifying fields) that lands in your admin alongside form submissions. Your sales team reads a one-screen summary, not a wall of chat dialogue they have to parse by hand. The full transcript is still archived for audit, but it is not the deliverable.

Official SDK over raw HTTP

We use Anthropic's official PHP SDK rather than hitting the REST API directly. It handles streaming, retries, validation, and version skew when Anthropic ships new model IDs. The cost is one library dependency, the benefit is not rewriting transport plumbing every six months.

What you get

• A chatbot that actually helps, not a scripted "Sorry, I didn't understand."
• Pre-qualified leads land in your inbox with name, contact, and what they want. Cold contact forms become warm intros.
• Costs that scale with usage, not seat licenses. A typical 1,000-visitor month with 5% chat engagement runs $3-8 in API cost.
• No vendor lock-in to a SaaS chatbot platform (Intercom, Drift, Tidio).
• The full conversation history is logged in your database. Audit any chat, replay any handoff.

Where we don't use it

For deterministic transactional flows (order confirmation, password reset, appointment scheduling), classic form plus email is faster, cheaper, and more reliable. AI is the wrong layer for those.

For very high-volume support (10k+ tickets/month), per-request costs add up. At that scale, we would help a client evaluate a managed support stack (Intercom Fin, Zendesk AI) or fine-tune a smaller model.

For tasks where the answer needs to be exact and verifiable (legal calculations, medical dosing, financial figures), we do not put an LLM in the loop without a deterministic verifier behind it. The cost of being subtly wrong is too high.