TokenMaxxing: How to Get the Most Out of Your AI Models

What it claims

This is a cost-control guide for Claude and OpenAI models, not research. Its central figures are these. Anthropic's top model costs 100× more per token than its smallest, and OpenAI's lineup spans $10 to $0.10 per million input tokens. On both vendors' top models, output tokens cost 5× more than input tokens. In the worked example, routing alone cut a $30 job to $3.90.

Evidence

• Invoice job. The job is 1,000 invoices, each with a 2,000-token input and a 200-token answer. It costs about $30 on Fable 5.1 and about $0.30 on Haiku 5.5. The routed version used the cheapest passing model, then paid Opus 5.5 to redo 300 failures (a 30% failure rate). Total: $3.90.
• Tool loading. In an Anthropic example, an agent's token use dropped from 150,000 to 2,000 when it loaded only the tool definitions it needed.
• Agent teams. Anthropic says agent teams in Claude Code can use about 7× the tokens of a standard session when teammates run in plan mode.
• Prompt length. OpenAI's testing found leaner prompts scored higher while using fewer tokens. No numbers are given.
• Prompt position. A Stanford-led study found models use information best at the start or end of a long prompt and worst in the middle.
• Price trends. Silicon Data's Token Expenditure Index shows per-token price down more than 90% since 2023, while LLM spend has doubled since late 2025. Bain reports token costs halved from December 2024 to 2025 while tokens consumed grew 450%.

How to use it

Billing and configuration details the summary omits:

• Haiku long prompts. Haiku's listed rate applies only to prompts under 100,000 tokens. Longer prompts cost 5×.
• Thinking tokens. Anthropic bills the full thinking process as output, including the parts you never see. On newer Claude models where thinking can't be turned off, lower the effort setting instead.
• Cache check. Writing to the default 5-minute cache costs extra. If cache_read_input_tokens still reads zero after the first call, caching isn't working. The page's shape for marking a cached block: `` "system": [{ "type": "text", "text": "Your invoice instructions...", "cache_control": {"type": "ephemeral"} }] `` Put stable content first and the parts that change last.
• Claude Code escalation. Default to Sonnet. Move to Opus only after a specific failure, such as a bug Sonnet can't fix after two tries. For simple subagents, set model: haiku.
• /compact. Compacting is itself a large request. Use /clear (free) unless you need the history, and pass /compact a keep-note, e.g. /compact keep the database decisions.
• Session overhead. Keep CLAUDE.md under 200 lines. /usage shows where tokens went. CLI tools like gh cost nothing until they run, unlike always-loaded MCP servers.
• Spend limits.
• On OpenAI's API, set a hard spend limit per project. Spend alerts only notify you and don't stop traffic.
• On Claude Pro/Max with usage credits, set a monthly limit under Settings → Usage.
• For scripted Claude Code runs, use --max-budget-usd.
• Shopify uses circuit breakers that cut off a developer's access when daily spend spikes.
• Free and local options. OpenRouter's free router dispatches to about 20 no-cost models. LM Studio or Ollama need an 8GB-VRAM GPU or an Apple M-series Mac.
• Model test. Run five of last week's tasks on both a small and a large model. If the small one passes, make it the default for a week.

Limits

• Missing price table. The article's Anthropic per-million-token price table did not come through in the page text. Check current prices before reusing the $30 / $0.30 figures.
• Example inconsistency. The invoice example names Fable 5.1 as the baseline but Opus 5.5 for redos. It doesn't say which cheap model handled the passing 700 or how failures were detected.
• Free tiers. They can disappear overnight. Read their data policies before sending sensitive input.
• Sourcing. Most numbers are relayed from vendors, consultancies or press reports, and come without methodology. The Meta figures are about 60 trillion tokens in 30 days and an estimated ~$900M at list price, which the page says Meta likely didn't pay.