OpenAI's GPT-6 Sol and Luna halve GPT-5.6 API prices; ChatGPT Voice gains email, calendar, and Slack on the new models

On 22 September 2026 OpenAI launched GPT-6 Sol ($2/$10 per 1M tokens) and Luna ($0.10/$0.50) at 50% below GPT-5.6 pricing, roughly an hour to 90 minutes after Anthropic's Claude Opus 5.5; OpenAI benchmarks show agentic and cost gains but Artificial Analysis…

On 22 September 2026, OpenAI extended its GPT-6 generation, introduced earlier in September with flagship GPT-6 Astra, with GPT-6 Sol and GPT-6 Luna, trained with methods similar to Astra. Sources disagree on the gap to Anthropic's same-day Claude Opus 5.5 launch: TechCrunch reports about 90 minutes, Simon Willison roughly an hour; Willison frames the launches, which followed the prior day's Grok 4.7 and MiMo v2.6 Flash/Pro, as a new price war. Pricing is 50% below GPT-5.6 equivalents (Hacker News specifies GPT-5.6 promotional rates): Sol at $2/$10 per 1M input/output tokens (down from $4/$20) and Luna at $0.10/$0.50 (down from $0.20/$1.20). The Decoder reports the Terra tier was retired, though no other source corroborates this. Caching improvements include up to 90% discounts on cached input reads, a caching dashboard, and new cache-control developer features. API ids are gpt-6-sol and gpt-6-luna; availability is described inconsistently: TechCrunch says ChatGPT Work, Codex, and the API with a gradual ChatGPT rollout, while MarkTechPost calls it an API-only release rolling out to ChatGPT Work, Codex, and a desktop app. On OpenAI benchmarks, Sol scores 33.2% on AutomationBench 1.0.6 at $0.27 per task (xhigh effort) versus Claude Opus 5's 26.9% at max effort at 11x the cost (roughly 91% cheaper per task), 56.4% on Agents' Last Exam, 68.8% on DeepSWE v1.1, 60.5% on OSWorld 2.0, and 49.3% on FrontierCode 1.1 at $2.14 per task versus Fable 5.1's 50.3%. Framing of the DeepSWE result differs: The Decoder notes it trails Opus 5's 73.7% and GPT-5.6 Sol's 72.7%, while Hacker News calls it near Claude Fable 5 at roughly 80% lower cost. Luna scores 66.6% on DeepSWE v1.1 at 93% lower cost per task than Opus 5, and Help Net Security cites a 5.4% accuracy gain with 58% cost reduction for Luna. OpenAI claims Sol makes about half as many factual mistakes as its predecessor and beats Anthropic's Fable and Opus models on several tasks. Independent assessments are cooler: Artificial Analysis (via The Decoder) moves Sol's intelligence score only from 47 to 48 with Luna flat at 37, plus GDPval-AA Elo drops of roughly 100 and 75; Ars Technica calls the gains a few percentage points and positions Sol as the efficient daily driver and Luna as the fast, cheap tier. On 23 September, OpenAI updated the globally available ChatGPT Voice to run on GPT-6 Astra, Sol, and Luna, adding plugins for email, calendar, and Slack so users can look up events, send email, review finance-app…

• OpenAI released GPT-6 Sol and GPT-6 Luna on 22 September 2026, following flagship GPT-6 Astra earlier in September; both were trained with methods similar to Astra.
• Pricing is 50% below GPT-5.6 equivalents (Hacker News: versus GPT-5.6 promotional rates): Sol at $2 input/$10 output and Luna at $0.10/$0.50 per 1M tokens, down from $4/$20 and $0.20/$1.20.
• The Decoder reports the Terra tier was retired; no other source corroborates this.
• API ids are gpt-6-sol and gpt-6-luna.
• Availability conflicts: TechCrunch says ChatGPT Work, Codex, and the API with a gradual ChatGPT rollout; MarkTechPost says an API-only release rolling out to ChatGPT Work, Codex, and a desktop app.
• Timing vs Anthropic's same-day Claude Opus 5.5: TechCrunch says about 90 minutes; Simon Willison says roughly an hour.
• AutomationBench 1.0.6: Sol scored 33.2% at $0.27 per task (xhigh effort) versus Claude Opus 5's 26.9% at max effort at 11x the cost (roughly 91% cheaper per task).
• DeepSWE v1.1: Sol scored 68.8% (The Decoder: below Opus 5 at 73.7% and GPT-5.6 Sol at 72.7%; Hacker News: near Claude Fable 5 at about 80% lower cost); Luna scored 66.6% at 93% lower cost per task than Opus 5.