GPT-6 Luna cut output-token pricing by more than half

Anthropic released Claude Opus 5.5, and OpenAI introduced GPT-6 Sol and GPT-6 Luna roughly an hour later. The three models arrive with different capabilities, but their shared message is pricing: vendors are cutting the cost of the tier where companies can run high volumes of agent work.

Luna cut output pricing from $1.20 to $0.50

GPT-6 Luna costs $0.10 per million input tokens and $0.50 per million output tokens. GPT-5.6 Luna previously cost $0.20 and $1.20. GPT-6 Sol is priced at $2 for input and $10 for output, half the price of GPT-5.6 Sol.

Claude Opus 5.5 is 20% cheaper than Opus 5 at $4 and $20. Cache reads fell 60% to $0.20. Anthropic also says typical workloads cost 40% less because the model uses fewer tokens and that output is generated more than 30% faster.

The price sheet is starting to rewrite model routing

For product teams, this changes the default architecture. Luna can serve cheap, frequent steps, Sol can handle harder decisions, and Opus can take long coding or knowledge-work tasks. At this price spread, using one model for everything becomes an expensive convenience.

GPT-5.6 Terra is particularly exposed. It matches GPT-6 Sol's input price but has a higher $12 output price. Existing routing rules therefore need new evals rather than a simple model-name replacement in configuration.

A cheaper token does not guarantee a cheaper completed task

Price comparisons do not reveal how many steps a model needs, how often it retries or how much failure costs. Anthropic mostly compares Opus 5.5 with its own models and older OpenAI models. OpenAI presents its capabilities within a separate evaluation frame. A direct independent Sol versus Opus 5.5 test is still missing.

Workload shape matters too. Long agent sessions benefit from cheap cache reads, while short generation jobs may be dominated by output pricing. One table cannot settle a purchasing decision.

Evals tied to the real invoice will pick the winner

Teams should replay their own task set across all three models and measure success, latency, tokens, retries and human intervention. Only the combined result exposes the cost of a completed outcome.

If these low prices persist, continuous multi-model routing will become more attractive. The useful signal will not be another vendor chart, but a month of production logs in which the cheaper model preserves quality and actually reduces the bill.

Lilith's verdict

Vendors rewrote the price tags within an hour, but the enterprise bill will not rewrite itself. The winner is the model that crosses the finish line with a completed task, not the one advertising the cheapest fuel.