Opus 5.5 vs GPT-6 Sol and Luna: the AI price war just rewrote your automation budget

Two labs, one hour apart, same message: a little more capability for a lot less money.

What happened

Anthropic released Claude Opus 5.5. Input and output tokens are now $4 and $20 per million, 20% less than Opus 5. Cache reads dropped to $0.20 per million, 60% less, and Anthropic says output is more than 30% faster. Because the model also uses fewer tokens per task, Anthropic puts the saving at about 40% versus Opus 5 on typical workloads at default settings.

About an hour later, as Simon Willison noted, OpenAI released GPT-6 Sol ($2 input, $10 output per million) and GPT-6 Luna ($0.10 and $0.50). Willison reports both are half the price of their GPT-5.6 equivalents. Ars Technica framed the pair as efficiency releases rather than capability leaps, aimed at enterprises that have started using model routers to lean less on pricey frontier models.

One caution from Willison's testing: Opus 5.5 at the "max" thinking level hit its 128,000 output token limit twice on a simple SVG prompt without returning an answer.

My take

For the automations I build, the line that matters is cache reads. Willison points out that in longer agentic conversations, more than 90% of input tokens are billed at cached prices. If your agent reuses a big system prompt, a CRM schema or a knowledge base on every call, a 60% cut there moves your monthly bill far more than the headline rate.

Three things I would do this week:

1. Re-price your running workflows. A lead scoring or ticket triage flow that was too expensive on a top model last month may now pencil out.
2. Route by job. Cheap models like Luna handle classification and extraction; save the bigger models for steps that need judgment.
3. Do not default to maximum effort. Set a sensible effort level and a token budget, and alert when a run blows past it.

Anthropic's own AutomationBench table, a business workflow benchmark, shows Opus 5.5 at 40.0% and GPT-6 Astra at 41.4%. Useful gains, but nowhere near "set and forget". Keep the human checkpoints.

This is the kind of cost-aware agent setup I build for teams. More at romielwillautomate.dev.