OpenAI has expanded its GPT-6 lineup with two models — GPT-6 Sol and GPT-6 Luna, built on the foundation of the flagship GPT-6 Astra, which was released earlier in September 2026. GPT-6 Sol costs $2 per million input tokens and $10 per million output tokens — exactly 50% cheaper than GPT-5.6 Sol, while GPT-6 Luna costs $0.10 and $0.50 per million tokens compared to $0.20 and $1.20 for GPT-5.6 Luna. The company confirmed that the new prices are permanent, not promotional, and the discount on reading cached input tokens has increased to 90%.
What happened
OpenAI released GPT-6 Sol and GPT-6 Luna — two models created on the technological base of the flagship GPT-6 Astra. GPT-6 Sol is priced at $2 per million input tokens and $10 per million output tokens, compared to $4 and $20 for GPT-5.6 Sol, meaning it is exactly 50% cheaper. GPT-6 Luna costs $0.10/$0.50 per million tokens compared to $0.20/$1.20 for GPT-5.6 Luna: input is 50% cheaper, output is approximately 58% cheaper. OpenAI specifically emphasized that this is a permanent price, not a time-limited promotion. The cost reduction was achieved through inference optimization and caching: the discount on reading cached input tokens has increased to 90%, and a Prompt Caching Dashboard and a tool for diagnosing missed cache have appeared. According to the stated benchmarks, GPT-6 Sol (xhigh) scores 33.2% on AutomationBench 1.0.6, compared to 26.9% for Claude Opus 5 (max), at approximately 9% of its cost per task ($0.27), and 56.4% on Agents' Last Exam.
Context
The release fits into the unfolding price competition in the mid-range and mass segments. GPT-6 Sol was released at exactly the price of Claude Sonnet 5 ($2/$10) and twice as cheap as Claude Opus 5.5 ($4/$20), which Anthropic presented on the same morning. GPT-6 Luna, by input price ($0.10 per million tokens), has almost caught up with open MIT models like Xiaomi MiMo-V2.6-Flash ($0.14/$0.28). The background is also the maturity of cache engineering: GitHub reported a reduction in the share of non-cacheable prompt tokens by more than 50% on billions of requests, meaning the pattern with a stable prefix and variable prompt tail has already been proven in production and is no longer an exotic thing.
Why this matters for the industry
For the industry, OpenAI has moved price competition from the top segment to the mid-range and mass segments: a permanent, not promotional, 50% price reduction plus a 90% discount on cached tokens directly changes the unit economics of agentic and automation workloads in production. Sol covers heavy agentic scenarios at a mid-range price, and Luna opens up previously unprofitable mass processing streams. The cache discount changes prompt architecture: it is beneficial to move long, stable system instructions and knowledge bases into an immutable prefix, and the Prompt Caching Dashboard and missed cache diagnosis provide ready-made optimization tools. If the stated figures are confirmed on real tasks, the break-even point for agentic features shifts significantly downward, and competitors and intermediary startups will likely have to respond with new rounds of cost reduction.
Why this matters for users
The models are already available via API, so verification takes a day, not a quarter: you can run the old and new models on your tasks and measure the actual cost and quality. It makes sense to immediately test Sol on code review, debugging, and data analysis, and Luna on mass summarization and extraction. According to the stated figures, Sol matches or exceeds Claude Opus 5 on key benchmarks at 9–20% of its cost per task, and Luna at high effort levels delivers results at the level of GPT-5.6 Sol at approximately 1/100 of its price. Those who count tokens should recalculate their automation budget taking into account the new caching and return multi-step cycles, self-checks, and long system prompts that were previously cut for economy.
What is still unknown / limitations
Benchmarks AutomationBench 1.0.6, Agents' Last Exam V1, and DeepSWE v1.1 are vendor figures on reference tasks, not a guarantee of quality on specific workloads, so unit economics should be rebuilt after running your own scenarios, not before. It is unknown how competitors will respond to the permanent cost reduction and whether the announced price level will be maintained. There is no independent verification of the stated cost and quality figures at the time of release.
Sources
Author
Look at AI, editorial team