Sol and Luna resolve the teased Tuesday release
OpenAI launched GPT-6 Sol and GPT-6 Luna on 22 September, resolving the unnamed Tuesday release tracked in this story. Both models are available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users and through the API as `gpt-6-sol` and `gpt-6-luna`. Free and Go users can access Luna in the desktop app. OpenAI says the rollout through ChatGPT surfaces is gradual and that the models are not yet available in the ordinary Chat surface.
The launch follows GPT-Live-1 and ChatGPT Images 2.5, which OpenAI shipped earlier in the same release cycle. Those releases remain part of the provenance of this story, but the material change on 22 September is the arrival of the cheaper GPT-6 tiers with named model IDs, published pricing and defined product availability rather than another schedule signal.
API prices fall while benchmark claims remain vendor evidence
OpenAI lists GPT-6 Sol at $2 per million input tokens and $10 per million output tokens, down from the GPT-5.6 Sol promotional rates of $4 and $20. GPT-6 Luna is listed at $0.10 per million input tokens and $0.50 per million output tokens, compared with $0.20 and $1.20 for GPT-5.6 Luna. The company describes the new family as 50% cheaper relative to the earlier promotional pricing and keeps GPT-6 Astra as its highest-capability tier.
OpenAI also publishes gains for Sol and Luna on professional-work, factuality, coding and computer-use evaluations. Those comparisons are useful launch measurements, but most were run by OpenAI or taken from provider reports under specific harness and effort settings. They do not establish an independent cross-provider ranking, and production cost will still depend on prompt length, reasoning effort, tool use and cache reuse.
GPT-6 prompt caching becomes more explicit
The same release adds more control over repeated context. OpenAI says eligible shared prompt prefixes on GPT-5.6 and later can remain reusable for at least 30 minutes after the latest cache write or reuse. Cached reads are billed at 0.1 times the standard input-token rate and cache writes at 1.25 times that rate. Developers can now set explicit cache breakpoints, inspect a Prompt Caching Dashboard and cache-miss diagnostics, and issue prewarm requests that populate the cache without generating a model response.
GPT-6 also adds a configuration-update path that can change reasoning effort without rewriting the earlier cached prefix, while OpenAI recommends keeping tool definitions stable and changing allowed tools or tool choice instead of rebuilding the tool list. These controls can reduce latency and cost for long-running agents with stable prefixes, but the savings are workload-dependent: cache writes carry a premium, prewarming itself is billed, and frequently changing prompts may produce little reusable context.
A banked Codex reset is rolling out on narrower plan terms
OpenAI product leader Tibo Thsottiaux separately said on 22 September that a banked reset was being loaded into Plus, Pro and Business accounts. His post establishes that rollout was under way, not that every eligible account had already received it. The GPT-6 launch page does not mention the reset, and the statement does not extend the grant to Enterprise, Edu, Free or Go users even though some of those plans receive Sol or Luna.
OpenAI's Help Center describes a banked reset as a one-time Codex usage-limit reset that can refresh eligible five-hour and weekly windows when the user applies it. The documentation also says eligibility, delivery timing and expiration vary by offer, plan, workspace and region. It documents earlier Astra-rollout grants rather than final terms for the 22 September grant, so the new reset should not be treated as universally delivered or assigned an expiry date without offer-specific evidence. The story remains Confirmed on the model release and published product changes, with the reset kept as a narrower first-party rollout claim.