What happened
OpenAI released GPT-6 Sol and GPT-6 Luna on September 22, trained the same way as its top-tier GPT-6 Astra, and halved API prices from the GPT-5.6 generation's promotional rates.
Why it matters
Sol and Luna bring GPT-6-class performance to OpenAI's faster, cheaper tier, and the API cut lowers the cost of building on these models for developers.
What to watch
Whether the lower prices hold beyond the promotional framing, and how external evaluators respond, since Sol and Luna carry no external evaluation section.
WHO IT HITSDevelopers and businesses building on OpenAI's API, plus ChatGPT paid-plan users, get access to a cheaper GPT-6-class tier; the halved prices directly lower per-token costs for teams running these models in production.
Summaries like this, in your inbox every morning.
OpenAI is positioning Sol and Luna as faster, cheaper models trained the same way as GPT-6 Astra, its top-tier model, which the company still calls its highest-performance model across all areas. The API price cut applies against the promotional pricing of the GPT-5.6 generation's same-named models, and OpenAI says it achieved the lower cost through caching and inference efficiency and passed that on to prices. A prompt-cache improvement also makes cached input reads 90% cheaper, and caching now survives mid-conversation changes to reasoning effort or available tools, with developers able to choose what gets cached.
The release lands in a busy stretch for OpenAI. CEO Sam Altman had agreed with Anthropic CEO Dario Amodei's September 12 essay calling for pacing adjustments in AI development, and OpenAI said it would give independent evaluators employee-level access. The company published a framework for tracking and disclosing misalignment cases on September 16 and on September 21 called for US-led international standards for frontier AI. The Sol and Luna announcement and system card appendix, however, make no mention of pacing, external evaluators, the misalignment framework, or international standards, and unlike Astra, no external evaluation is described for these models.
On safety, the Preparedness Framework treats Sol and Luna as having "High" capability in cybersecurity and biological/chemical areas, short of Astra's "Critical" level as judged by OpenAI's Safety Advisory Group, while AI self-improvement stays below the "High" threshold. Alignment evaluations improved over the GPT-5.6 generation, yet the appendix also flags traits OpenAI calls undesirable for monitoring: Sol showed signs of noticing it was being evaluated in 4.76% of cases, up from 2.56% for GPT-5.6 Sol. How much the price cut and the Astra-style training translate into real-world adoption will likely depend on whether the promotional framing holds and how external reviewers respond to the thinner disclosure.
For example, today's edition would include:
AI-summarized, only the topics you pick — one digest a day via Email, LINE, or Slack.
Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. Q&As are published on this page for other readers too.
Anthropic released Claude Opus 5.5 today and cut its price 20%, with input at $4 per million tokens and output…
Firecrawl announced it has raised $75 million in a Series B round led by Smash Ventures, with participation fr…
ASRock is shifting its business focus toward AI
Anthropic and OpenAI, which spent early September warning that model capabilities are outrunning the safeguard…
Jessica Wachter of Wharton and co-authors estimate hyperscaler spending will reach nearly 1.1兆ドル by 2027, and…
In September 2026, Amazon Web Services and Salesforce announced that Salesforce CRM data will be usable from A…
