A Jev-as-a-Judge paper reports that accepting TypeSafe AI’s confident JEV verdicts and escalating the rest to GPT-6 Astra kept 99% of GPT-6 accuracy on 510 held-out preference pairs, at about 57% of GPT-6’s fee. The authors say the escalation threshold did not transfer to every fallback model and should be set on one’s own data.
OpenAI says ChatGPT Voice can use plugins such as email, calendar, and Slack, run on GPT-6 Astra, Sol, and Luna, and work in ChatGPT Work on the web and mobile. The update is rolling out globally today in the latest version of the app.
Codex CLI 0.156.1 includes GPT-6 Sol and GPT-6 Luna in the model picker. The rate-limit switch now suggests GPT-6 Luna, and the model catalog is updated.
Artificial Analysis says GPT-6 Sol and Luna score similarly to their predecessors on its Intelligence Index at about half the cost, with the reduction driven by token prices that are roughly 50% lower.
Glean says Claude Opus 5.5 cuts token costs by 50% versus its predecessor on Glean's own evaluations, and that GPT-6 Luna and GPT-6 Sol each cost 50% less than their predecessors.
Luna is available on all plans and the API; the $10 GOAT plan is given as an example with about 30,000 requests and about 1.5 billion tokens. Sol is available on the Pro and Max plans and the API.
Databricks is making OpenAI’s GPT-6 Sol and GPT-6 Luna and AnthropicAI’s Claude Opus 5.5 available to its customers. Databricks says the GPT-6 models perform strongly at lower cost with a higher Pareto frontier on OfficeQA Pro, and that Claude Opus 5.5 improves cost efficiency on enterprise document parsing.
OpenAI Developers says applications can select reusable prompt prefixes with explicit cache breakpoints, change reasoning effort and tool availability while preserving cached context, and prewarm shared context so responses start sooner.
Higher default cache-hit rates let more GPT-6 input tokens receive cached-input discounts of up to 90%. OpenAI Developers says the change helps agents run faster and cost less.
OpenAI’s two models can be tested on Arena, where votes on real-world agentic tasks feed the leaderboard. Arena.ai says petergostev has also compared GPT-6 Sol with GPT-5.6 Sol using the same prompts at max reasoning.
On Arena.ai, both models can be tried in Battle Mode and Agent Mode.
OpenAI’s GPT-6 Sol and GPT-6 Luna can now be tested in Arena.ai’s Agent Arena, with scores still to come, and in Code Arena for WebDev, Text, Vision, Search, and Document. OpenAI says the models build on GPT-6 Astra and that their API prices are 50% lower than GPT-5.6 promotional pricing.
GitHub Copilot now generally offers two additional OpenAIDevs GPT-6 models: GPT-6 Sol for interactive and agentic coding, and GPT-6 Luna, which GitHub calls a lightweight, lowest-cost option for smaller tasks.
Sam Altman says GPT-6 Sol and Luna are big improvements over 5.6-family predecessors in intelligence, alignment, work output, coding and computer use, at half the price per token and even less per task. He says that on per-task pricing he does not think anything in the market is competitive.
According to Polymarket, OpenAI claims task costs are up to 93% cheaper than Claude Opus 5 in some coding evaluations.
He says both models are big improvements over their 5.6-family predecessors in intelligence, alignment, work output, coding, and computer use, and that they cost even less per task.
GPT-6 Sol and GPT-6 Luna build on advances behind GPT-6 Astra as faster, more affordable models for work at scale. OpenAI says more efficient caching and inference are reflected in API prices 50% lower than GPT-5.6 promotional pricing.
OpenRouter says GPT-6 Sol and GPT-6 Luna use Astra's clearer, shorter, lower-jargon answers and OpenAI's improved prompt caching, with 90% off cached input reads.
OpenRouter is offering OpenAI’s GPT-6 Sol at $2 per million input tokens and $10 per million output tokens, and GPT-6 Luna at $0.10 and $0.50. OpenRouter says both are half the price of their GPT-5.6 predecessors, and that each tops its predecessor’s best AutomationBench score at a fraction of the cost per task.
The September 22, 2026 rollout covers Plus, Pro, Business, Enterprise, and Edu users.
