Reasoning-token bloat on DeepSeek and o-series models

Reasoning models spend hidden thinking tokens by default. On a simple prompt the thinking adds cost without accuracy.

1. With auto routing, the reasoning-bloat guard caps reasoning_effort (or output_config.effort) to low on prompts the classifier scores as simple; the header X-TokenZero-Reasoning-Cap shows the change.
2. Prefer a non-reasoning sibling for extraction and classification; the audit and calculate_model_arbitrage price the alternatives.
3. On DeepSeek, run off-peak where you can: outside Mon-Fri 01-04 and 06-10 UTC every token is half price, and the receipt records the tier.

copycurl -s -X POST https://tokenzero.sh/v1/signup -d '{}'   # workspace + key, no form
npx -y @tokenzero/cli init                                # or: audits the repo, writes .env

copyOPENAI_BASE_URL=https://tokenzero.sh/w//v1
ANTHROPIC_BASE_URL=https://tokenzero.sh/w/
# provider keys stay in your code exactly as they are; nothing else changes

Which models count as reasoning models?
OpenAI o-series and gpt-5/6 families, deepseek-v4-pro and the grok-4.x line are treated as heavy-reasoning by default; the guard only acts when routing is set to auto.