That distribution story carries the cost claim: OpenAI attributes the 50% price cut to improvements in caching and inference rather than to a lossier model. The internal factuality claim - 'about half as many mistakes as GPT-5.6 Sol' - lines up with the Astra benchmark position.
Day-one rollout and access
The September 22, 2026 launch followed a tier-by-tier rollout:
- GPT-6 Sol: ChatGPT Work, Codex, and the API on launch day for Plus, Pro, Business, and Enterprise; Sol Pro to Pro and Enterprise only.
- GPT-6 Luna: rolling out throughout the week to Free and Go users with unlimited text chats and a new Think button for harder questions; rate limits remain on file uploads, image generation, and other tools.
- Work and Codex: the version of GPT-6 Sol that powers Work and Codex is unchanged from the prior release.
The new thought slider on ChatGPT Sol lets users pick the reasoning effort per message. The settings run from none through low, medium, high, xhigh, and a new max effort for the deepest reasoning - the same effort taxonomy as GPT-6 Astra.
How GPT-6 Sol compares to the frontier
OpenAI published three benchmark numbers for GPT-6 Sol in the launch post. On FrontierCode 1.1 Main - a 'mergeability' grading rubric for agent-written code changes - GPT-6 Sol improves substantially over GPT-5.6 Sol and matches Claude Fable 5.1 xhigh at much lower cost.
On DeepSWE v1.1, GPT-6 Sol max effort scores 68.8% on complex software-engineering tasks in real codebases, within 1.1 percentage points of Claude Fable 5's best published score of 69.9% at xhigh effort - at roughly 80% lower cost per task.
On Terminal-Bench 2.1, GPT-6 Sol sets a new state of the art for command-line workflows. OpenAI notes these are agentic-coding benchmarks rather than chat benchmarks; for general knowledge work, GPT-6 Astra remains the higher-scoring OpenAI model.
How GPT-6 Luna compares to GPT-5.6 Luna
GPT-6 Luna replaces GPT-5.6 Luna in the everyday tier. OpenAI positions Luna for 'easy, lower-stakes everyday tasks' - the same workload slot GPT-5.6 Luna filled - and says the new model outperforms GPT-5.6 Luna at one-fourth the API price.
Luna's pricing makes it OpenAI's cheapest API model currently sold. The cache price of $0.01 per million tokens is roughly one cent per million - cheap enough that cached-prompt workflows can run at near-zero input cost.
What to do today
Three workflows are worth a same-day test. First, swap GPT-5.6 Sol for GPT-6 Sol on coding-agent pipelines. At the same input price as GPT-5.6 was a year ago and roughly half of GPT-5.6, the cost-performance math has changed since the May pricing reset.
Second, route Free-tier or high-volume Q&A traffic to GPT-6 Luna once it becomes your default. The $0.01 per million cache price makes cached-heavy RAG workflows viable at scale.
Third, audit factual-error behaviour on long-context recall tasks before retraining evaluation suites. OpenAI's internal factuality claim is bold; external evaluators typically take weeks to publish corroborating results.
GPT-6 Sol API changes are live in the OpenAI dashboard today. Luna's API tier is already published; the Free/Go default swap completes inside the launch week.
What was not announced
Three things stayed off the launch post. There is no GPT-6 Terra - the openai.com catalog lists gpt-6-astra, gpt-6-sol, gpt-6-luna only; GPT-5.6 Terra remains a separate generation. There is no separate 'reasoning' SKU for GPT-6 Sol; users pick effort via the slider.
There is no public training-compute number for either model. OpenAI has not disclosed token-count, GPU-hours, or dataset size for the GPT-6 series; third-party model-card publications remain the only public data on training scale (OpenAI model catalog, September 22, 2026).
Frequently Asked Questions
Reviewers summarised the most common September 23 reader questions - the same questions above plus their follow-ups.
Is GPT-6 Sol better than GPT-6 Luna for coding?
Yes, on published benchmarks. GPT-6 Sol leads on Terminal-Bench 2.1, FrontierCode, and DeepSWE; GPT-6 Luna is positioned for high-volume, lower-difficulty work where cost dominates capability. For agentic coding or multi-step refactor tasks, route to Sol.
For quick code-completion or one-shot utilities, Luna is the cost-effective choice (OpenAI, September 22, 2026; TechCrunch, September 22, 2026).
Does GPT-6 Sol replace GPT-6 Astra?
No. Astra remains the premium reasoning tier with the highest published benchmark scores on GPT-6.
Sol is positioned for cost-effective everyday agentic coding where pricing matters more than peak accuracy. Customers should run a small benchmark suite against their own workload to decide when to use which model.
When does GPT-6 Luna become the default for Free and Go users?
OpenAI said GPT-6 Luna 'will become the default model for Free and Go users this week' - inside the September 22 to 28 launch week.
Existing ChatGPT Free and Go accounts get Luna with unlimited text chats on rollout. Image uploads and tool-based workflows keep their existing limits.
What about the GPT-6 Sol thought slider?
The thought slider lets Plus and Pro users set reasoning effort per message: none, low, medium, high, xhigh, or max. Higher effort uses more tokens and costs more per answer but produces fewer step-by-step errors.
Is GPT-6 Sol the same price as GPT-6 Astra?
Yes - both list at $5 input and $30 output per million tokens.
The difference is that GPT-6 Sol has cached input pricing ($0.20 per million), while Astra does not. For workflows with significant prompt-caching, Sol becomes cheaper in practice; for one-off requests, the two models are equivalent in price (OpenAI API pricing, September 22, 2026).
Where can I read the GPT-6 Sol system card?
OpenAI has not yet published a separate system card for GPT-6 Sol. The September 22 announcement references GPT-6 Astra's system card for safety and preparedness details; expect a Sol-specific follow-up over the coming weeks.
Fazlur Rahman
