OpenAI launches GPT-6 Sol and Luna at half the price, reigniting the frontier AI cost war

24 Sep 2026 4 min read 1648 views Technology
Technology

On September 22, 2026, OpenAI launched GPT-6 Sol and GPT-6 Luna, two new models whose main selling point is not just capability but price: both cost roughly half as much as GPT-5.6 in the API. Sol now costs $2 per million input tokens and $10 per million output tokens, while Luna, the lightweight model, comes in at $0.10 and $0.50 respectively.

The launch comes just 19 days after GPT-6 Astra and in a particularly busy week: Anthropic released Opus 5.5 with lower prices and Google introduced Gemini 4 Argon, aimed at complex reasoning and cyber defense. The takeaway is clear: competition at the AI frontier is no longer played only on benchmarks, but on how much each solved task costs.

This week's context

According to MarkTechPost, Technology Org and Digital Applied, both models rolled out gradually throughout September 22. Sol is the flagship model, designed for long agentic and coding tasks; Luna is the efficient version for high volumes. Both offer a context window of roughly one million tokens. Bear in mind that the published performance figures come mostly from OpenAI's launch materials and have not yet been widely verified independently.

What has changed

  • Half the price: Sol drops from $4/$20 to $2/$10 per million tokens (input/output) compared with GPT-5.6. Cached input costs $0.20 per million.
  • Luna, even cheaper: $0.10 input and $0.50 output per million tokens, a cut of around 58% on output.
  • Pricing modifiers: Batch and Flex modes cost half, Fast mode costs double, and requests over 272,000 input tokens carry a surcharge (2× input and 1.5× output).
  • Coding: on DeepSWE v1.1, Sol reaches 68.8% and Luna 66.6%, close to the best competing models according to OpenAI.
  • Computer use and automation: Sol scores 60.5% on OSWorld 2.0 and 33.2% on AutomationBench at a far lower cost per task than its rivals, again according to OpenAI's data.

Impact for development and product teams

For anyone building products on top of LLMs, a 50% price cut changes the math for many use cases that did not pay off until now: agents chaining dozens of calls, bulk document processing, or coding assistants working for hours. Luna in particular makes it viable to use a reasonably capable model for high-volume tasks such as classification, extraction or routing. At the same time, with three major launches in the same week, depending on a single provider is riskier than ever: prices and relative rankings shift within days, and the benchmarks published by each lab are not directly comparable.

Practical recommendations

  1. Recalculate the cost per task of your current pipelines with the new prices, including caching, Batch and the long-context surcharge.
  2. Evaluate Sol and Luna on your own test set before migrating: vendor benchmarks are a starting point, not a guarantee.
  3. Split tasks by difficulty and route simple ones to Luna and complex ones to Sol (or to other providers' models).
  4. Keep an abstraction layer over your model provider so you can switch quickly based on price and quality.
  5. Watch requests above 272,000 tokens: the surcharge can wipe out much of the savings if context size is not kept in check.

What to watch next

  • Independent evaluations that confirm (or not) the DeepSWE, OSWorld and AutomationBench figures.
  • How Anthropic and Google respond on pricing after OpenAI's move.
  • The arrival of GPT-6.1, the planned successor to GPT-6 Astra.
  • Changes to usage limits and regional availability of the new models.

Conclusion: GPT-6 Sol and Luna are not just two more models: they signal that frontier AI is entering a phase of aggressive price cuts. For technical teams, the opportunity lies in reviewing which use cases now become profitable and in designing architectures that take advantage of competition between providers instead of being locked into one.

Sources and documentation

Share