Cut your AI bill: the price-cut cheat sheet

Claude Opus 5.5, GPT-6 Sol and GPT-6 Luna launched on 22 September 2026 with lower prices. Here's how to stack more savings on top.

The new prices (per million tokens)

Move 1: switch to the new model

Change the model name. Opus 5.5 costs 20% less per token than Opus 5. Anthropic says it costs about 40% less to run on typical workloads, because it also uses fewer tokens per task.

Move 2: cache the part you repeat

System prompts, long documents and chat history get re-sent on every call. Cache them.

• The easy way on Claude: add one cache_control field at the top level of your request (automatic caching).
• A 5-minute cache write costs 1.25x normal input, so it pays off after one cache read.
• On Opus 5.5, a cache hit costs 5% of normal input: $0.20 per million tokens.

Move 3: batch anything that can wait

• Claude's Batch API takes 50% off input and output. Opus 5.5 in batch: $2 / $10.
• OpenAI prices Batch and Flex for GPT-6 Sol and Luna at 50% of standard rates.

Move 4: send easy jobs to a small model

Anthropic's own advice: "Choose Haiku for simple tasks, Sonnet for most production workloads, and Opus for the most complex reasoning." On OpenAI's side, GPT-6 Luna is $0.10 / $0.50.

Move 5 (bonus): don't max out thinking

Simon Willison tested Opus 5.5 at the highest thinking level. It hit the 128,000 output-token limit while it was still reasoning. Every one of those tokens is billed as output. Start lower and only raise it if the answers need it.

Stack them

• Caching and the Batch API combine on Claude. The multipliers stack.
• Watch out for fast mode: on Opus 5.5 it costs $8 / $40 and does not work with the Batch API.

Checklist

1. Update the model name.
2. Turn on caching.
3. Move non-urgent jobs to batch.
4. Route simple tasks to a small model.
5. Check your thinking level.

Sources: simonwillison.net, platform.claude.com/docs pricing, techwireasia.com.