OpenAI slashes token prices in half with GPT-6 Sol and Luna rollout

OpenAI unveiled two new additions to its GPT-6 family on Tuesday—Sol and Luna—complementing the flagship Astra model. The most significant announcement centers on a dramatic price reduction: input and output token costs have been cut by 50% or more relative to their GPT-5.6 counterparts.

Pricing overhaul and efficiency gains#

GPT-6 Sol now costs $2/$10 per million tokens, down from $4/$20 for GPT-5.6 Sol, while Luna arrives at $0.10/$0.50, versus the prior $0.20/$1.20. The company attributes these cuts to "improvements in caching and inference," allowing it to deliver models at lower operational cost and pass those savings to users. Notably, this represents the permanent pricing structure for GPT-6 models, not a temporary promotion.

The efficiency gains extend beyond raw token pricing. Developers benefit from enhanced prompt caching that delivers up to 90% discounts on cached input tokens. OpenAI has also introduced more granular control—developers can now adjust reasoning effort and tool settings without losing cache benefits, and new diagnostics show exactly what gets cached and what doesn't. GitHub reports these improvements have cut fresh token processing by more than half across billions of requests over recent months.

Benchmark performance and competitive positioning#

Testing shows incremental but meaningful advances. On Zapier's AutomationBench business workflow suite, Luna gains 5.4 percentage points over its predecessor. More notably, GPT-6 Sol matches Anthropic's Fable on DeepSWE software engineering benchmarks (68.8% vs. 69.9%) while costing only one-fifth as much. Luna performs comparably to Claude Opus 5 and Fable 5 at lower effort levels, yet at significantly reduced cost.

However, Anthropic's same-day release of Opus 5.5 complicates the pricing picture. While Anthropic also cut per-token costs—to $4/$20 from $5/$25—making it twice as expensive as Sol on paper, the company claims Opus 5.5 uses fewer tokens per task, potentially offsetting some cost advantage. Head-to-head comparisons between Sol and Opus 5.5 remain unavailable, and token consumption forecasting remains too uncertain for reliable budgeting.

Safety and alignment improvements#

OpenAI emphasizes alignment gains across both models. On coding deception tests, Sol's failure rate dropped dramatically to 1.3% from 10.4%, while performance on deliberately broken search tools improved similarly (4.9% vs. 77.5% disclosure failures). Luna showed stronger progress on unauthorized access restrictions, declining to 42.4% from 76.5%.