Claude in Whittl¶

Anthropic's Claude is Whittl's highest-quality backend overall — consistently strong on complex Python code, excellent at following multi-step instructions, native vision across all tiers, and the most reliable tool-use implementation.

The tradeoff: most expensive per token of any backend. For serious work where iteration count matters more than per-call cost, it's the default pick.

Getting a key¶

Go to console.anthropic.com.
Sign up / log in.
API Keys in the left sidebar → Create Key.
Name it "Whittl", copy the key (starts with sk-ant-).

Add it in Whittl via Edit → Preferences → API Keys, paste into the Claude API Key field.

Picking a model¶

Three tiers, each with a clear role:

Opus (top-tier)¶

The quality ceiling. Use Opus for:

Hard architectural problems (designing a new system, complex refactors)
Long-context tasks where the input + output together exceed 50K tokens
Agent Mode on difficult tasks — Opus is the most reliable at autonomously working through a multi-round problem

Cost: roughly $15/M input + $75/M output tokens (check anthropic.com/pricing for current).

When to NOT use Opus: simple edits, quick iterations, anything where Sonnet would obviously handle it. You're overpaying 10× for no benefit.

Sonnet (default pick)¶

The "just use Claude" model. 95% of Opus's quality at ~20% of the cost. For most Whittl work, Sonnet is the right choice.

Generating new apps from natural language
Modifying existing projects
Screenshot to App
Most Agent Mode tasks

Cost: roughly $3/M input + $15/M output.

Haiku (iteration tier)¶

Fast, cheap, surprisingly capable. Use Haiku for:

Small surgical edits
Quick iterations on an established project
Tasks where you'd previously have used a cheaper provider

Cost: roughly $1/M input + $5/M output. About 15× cheaper than Sonnet per token.

Quality gap: Haiku is noticeably behind Sonnet on complex architecture and visual-design tasks. For pure "change this line" edits, indistinguishable.

Switching models¶

The Claude model dropdown next to the backend selector chooses which Claude variant to use. Settings persist per project.

You can also switch mid-session — conversation history carries forward. Common pattern: start a project on Sonnet, iterate most changes with Haiku, escalate one hard problem back to Sonnet or Opus.

Prompt caching¶

Claude supports prompt caching: the first request pays full price for the system prompt + tool definitions + code context; subsequent requests in the same session pay ~10% of that. The savings compound across multi-round sessions.

In practice this means:

A fresh project on Sonnet: first request ~$0.15, later requests ~$0.03
Long Agent Mode session (10+ rounds): ~87% cheaper than without caching

Whittl enables caching automatically. You don't need to configure anything — just use Claude and reap the savings.

Vision¶

Every Claude tier (Opus, Sonnet, Haiku) supports vision natively via Anthropic's image content blocks. Drop a screenshot into chat and the AI sees it.

Compared to OpenRouter → Claude via OpenAI-format multimodal:

Direct Claude API: native Anthropic image block format. Slightly richer image reasoning.
OpenRouter → Claude: OpenAI-format image_url content parts. Close but not identical.

If you care about the highest image-reasoning fidelity on a specific project, use the Claude API backend directly rather than Claude via OpenRouter.

Tool use¶

Claude's tool-use API is the most reliable in practice. Whittl's edit_code, syntax_check, create_file, run_code, and (in Agent Mode) bash tools all fire cleanly on Claude. The native text_editor_20250728 tool is used on Claude specifically for multi-round surgical editing.

Other backends work fine — Claude just produces fewer tool-call format errors or hallucinated tool arguments on edge cases.

Rate limits¶

Anthropic's rate limits scale with your account tier:

Free tier: very restrictive, usually maxes out quickly
Usage tier 1 ($5 spent): 50 requests/min, 50K tokens/min
Usage tier 2, 3, 4: higher limits as you spend more

If you hit rate limits, Whittl surfaces the error and you can either wait or switch to OpenRouter → Claude (which pools requests across OpenRouter's user base).

Claude pricing and cost notes¶

Use Haiku by default¶

Switch to Haiku for most work, escalate to Sonnet when quality matters. Most sessions can finish on Haiku alone.

Let prompt caching do its job¶

Keep sessions long. Starting a fresh project for every task defeats caching.

Multi-file projects benefit disproportionately¶

Smart routing + caching means your 30-file project costs roughly the same per edit as a single-file one, because the cached context covers the common ~80% and each new prompt only pays for the delta.

When NOT to use Claude¶

Budget-only work. Claude Haiku is cheap but DeepSeek or OpenRouter free-tier models are cheaper.
Offline / privacy-critical work. Use Ollama.
Exploratory model comparison. Use OpenRouter to swap between many models.

Troubleshooting¶

Invalid API key when the key looks right

Common causes: you copied whitespace into the key, or your Anthropic account is new and hasn't had its first credit purchase yet. Re-copy the key with a trimmed edit, and check your account has a valid payment method.

Model consistently slow

Anthropic's capacity fluctuates by time of day. US peak hours (9am-5pm Pacific) are slower than off-peak. Not a Whittl issue — happens on every tool using the Claude API.

Seeing 529 Overloaded errors

Anthropic's load shedding. Whittl retries automatically with backoff; if it keeps failing, wait 5-10 minutes or switch to a different backend for that session.

What's next¶

Agent Mode — where Claude's tool-use and long-context strengths pay off most
OpenRouter — alternative path to Claude if you prefer consolidated billing
Choosing a Backend — comparison with other backends