GPT-6 Sol and Luna
CodexOpenAI launched GPT-6 Sol and GPT-6 Luna on September 22, 2026, two new tiers that bring much of flagship GPT-6 Astra's gains to faster, cheaper models. Sol targets complex, agentic coding work in Codex and ChatGPT Work at $2/$10 per million input/output tokens, while Luna handles high-volume, low-latency tasks at $0.10/$0.50 per million tokens and reaches Free and Go users through the desktop app. Both models cost about 50% less than their GPT-5.6 predecessors and make roughly half as many factual errors as GPT-5.6 Sol. The launch landed just 90 minutes after Anthropic shipped Opus 5.5, underscoring how tightly matched the two companies' release cadences have become.
Featured Video
A video we selected to help illustrate this changelog
Key Takeaways
- GPT-6 Sol and Luna cost 50% less than their GPT-5.6 predecessors, with Luna's output tokens down 58.3% and a 90% discount on cached input reads.
- Sol makes roughly half as many factual errors as GPT-5.6 Sol, delivering what OpenAI calls "Astra-level reliability" without Astra's price tag.
- On OSWorld 2.0, Sol at its highest reasoning effort essentially matches Claude Opus 5's mid-tier score (60.5% vs 60.3%) at about 80% lower cost per task.
- Luna is the first GPT-6 model to reach Free and Go users, though only through the ChatGPT desktop app rather than the web or Codex CLI directly.
- The launch arrived 90 minutes after Anthropic shipped Opus 5.5, highlighting how tightly the two labs' coding-model release schedules now overlap.
- Enterprise workspaces must opt in manually: unlike consumer plans, Business and Enterprise admins have to enable Sol and Luna before their teams can select them.
Sources & Mentions
5 external resources covering this update
GPT-6 Sol | First impressions
YouTube
OpenAI launches GPT-6 Sol and Luna, boasting lower cost and fewer mistakes
TechCrunch
OpenAI releases GPT-6 Sol and Luna models, slashing API costs 50% or more
VentureBeat
OpenAI releases GPT-6 Sol and Luna, and cuts token prices in half
The New Stack
OpenAI upgrading ChatGPT and Codex with two more GPT-6 models
9to5Mac
OpenAI Expands the GPT-6 Family with Sol and Luna
On September 22, 2026, OpenAI extended its GPT-6 lineup beyond the flagship Astra model launched three weeks earlier, introducing GPT-6 Sol and GPT-6 Luna. Where Astra remains OpenAI's top-tier model for the most demanding coding, computer-use, and professional work, Sol and Luna are built to carry the same underlying training improvements into faster, far cheaper tiers that developers can run at scale inside Codex, ChatGPT Work, and the OpenAI API.
Two Tiers, Two Jobs
GPT-6 Sol is positioned as the workhorse for complex, recurring developer tasks: agent loops, long-context code review, and batch processing where per-token cost matters as much as reasoning depth. GPT-6 Luna sits a level below Sol and is tuned for high-volume, low-latency work such as document summarization, quick information extraction, and routine question answering, the kind of task Codex agents run constantly in the background.
Pricing Cut in Half
Both models ship with a permanent, not promotional, price cut. Sol costs $2 per million input tokens and $10 per million output tokens, a 50% reduction from GPT-5.6 Sol. Luna costs $0.10 per million input tokens and $0.50 per million output tokens, 50% cheaper on input and 58.3% cheaper on output than GPT-5.6 Luna, with a 90% discount on cached input-token reads. OpenAI attributes the savings to improvements in inference and caching technology rather than to a smaller or less capable model, and frames the cuts as necessary given how much token volume long-running coding agents consume: the company says its own researchers' daily token usage ranges from a median of $600 to a 90th-percentile figure of $7,000.
Fewer Mistakes, Comparable Reliability to Astra
On OpenAI's internal evaluations, GPT-6 Sol makes about half as many factual errors as GPT-5.6 Sol, which OpenAI describes as "Astra-level reliability at a much lower cost." On the DeepSWE 1.1 coding benchmark, Sol reaches 68.8% and Luna reaches 66.6%. On AutomationBench 1.0.6, Sol scores 33.2% at roughly $0.27 per task, a result OpenAI says substantially outperforms comparably configured competitor models. On OSWorld 2.0 (offline), Sol at its highest reasoning effort scores 60.5%, edging out Anthropic's Claude Opus 5 at medium effort (60.3%) at around 80% lower cost per task.
Rollout and Access
Sol and Luna are rolling out to Codex and ChatGPT Work for Plus, Pro, Business, Enterprise, and Edu subscribers, alongside general availability in the OpenAI API under the model IDs gpt-6-sol and gpt-6-luna. Luna is also available to Free and Go users, but only through the ChatGPT desktop app. Enterprise administrators must manually enable the new models in their workspace settings before their teams can use them. Codex CLI 0.156.1, released the following day, added both models to the in-terminal model picker and now recommends Luna by default when a user hits a rate-limit switch prompt.
Competitive Context
The timing drew attention on its own: Anthropic released Opus 5.5 roughly 90 minutes before OpenAI's announcement, a reminder of how closely the two labs' release schedules have started to track each other as both push toward cheaper, faster coding-focused models.