GPT 6 Sol & Luna Are Here (And 50% CHEAPER!)

Summary
Chase from Chase AI reviews OpenAI’s launch announcement of GPT-6 Sol and GPT-6 Luna, discussing their positioning relative to GPT-6 Astra and prior GPT-5.6 models. He examines official benchmarks, pricing reductions, factual error rate improvements, and prompt-caching enhancements.

What is shown

• OpenAI announcement page: “Introducing GPT-6 Sol and Luna” [00:00].
• FrontierCode benchmark chart comparing GPT-6 Astra, GPT-6 Sol, GPT-6 Luna, Claude Opus 5.5, and Claude Fable 5.1 against cost per task [00:20].
• Comparison graph on FrontierCode showing Claude Opus 5.5 scoring 54.4% at $6.19 per task [01:30].
• Agent’s Last Exam benchmark chart comparing GPT-6 models against GPT-5.6 predecessors and Claude models [02:20].
• API pricing table detailing token costs and 50% price reductions across GPT-6 Sol and Luna [03:05].
• Factuality chart showing error rates on difficult prompts across effort levels [03:20].
• Overview of prompt caching improvements, caching dashboards, and cache retention rules [04:11].
• Model availability and rollout details across ChatGPT tiers and API model strings (gpt-6-sol, gpt-6-luna) [05:14].

Claims & numbers

• Anthropic released Claude Opus 5.5 about an hour prior to OpenAI’s announcement, according to the presenter [00:04].
• On FrontierCode at max effort, the presenter notes GPT-6 Astra scores 53.3% at $4.59 per task, GPT-6 Sol scores 49.3% at $2.14 per task (medium effort: 45.9% at $0.80), and GPT-6 Luna scores 42.4% at $0.11 per task [00:55 - 01:27].
• On FrontierCode, Claude Opus 5.5 scores 54.4% at $6.19 per task [01:34].
• On Agent’s Last Exam, GPT-6 Luna achieves 50.9% at $0.15 per task compared to GPT-5.6 Luna’s 50.4% at $2.57 per task [02:46].
• API pricing: GPT-6 Sol costs $2 input / $10 output per million tokens (50% cheaper than GPT-5.6 Sol’s $4 / $20) [03:08].
• API pricing: GPT-6 Luna costs $0.10 input / $0.50 output per million tokens (50% cheaper than GPT-5.6 Luna’s $0.20 / $1.20) [03:16].
• Factual error rate on difficult prompts at max effort: GPT-6 Luna drops to 7.6% (from 12.0% on GPT-5.6 Luna); at low effort, GPT-6 Luna drops to 27.7% (from 36.8%) [03:31 - 03:51].
• Factual error rate for GPT-6 Sol drops to 4.6% at max effort (versus 8.5% for GPT-5.6 Sol), nearing GPT-6 Astra’s 3.9% [03:52 - 04:03].
• Cached input token reads receive a 90% discount [04:27].
• Adjusting reasoning effort levels or tool availability no longer invalidates prompt cache prefixes [04:45].

Notable quotes

• “So we didn’t just get one new model today, we got three. About an hour ago, Anthropic released Claude Opus 5.5, and now OpenAI is giving us GPT-6 Sol and GPT-6 Luna.” [00:00]
• “It’s about a third of the cost at a 5% drop in score. So this is still an awesome model...” [01:40]
• “Adjusting your reasoning effort keeps you in a warm cache state, which is going to use less usage and it’s going to save you money...” [04:56]

Assessment
This is a third-party commentary and analysis video walking through OpenAI’s official launch blog post and charts. No live hands-on testing or independent prompting is performed; all metrics and benchmarks shown are directly from OpenAI’s documentation.

Described by gemini-3.8-flash on 2026-10-12 from the video’s audio and frames.