GPT-6 Sol leaks, screenshots and release rumors

Dan Shipper gained nearly 10k followers in a day and pinned his GPT-6 Sol vs Opus 5.5 comparison review plus a piece arguing AI automation creates more work for human experts.

The post asks whether people have actually switched to GPT-6 Sol/Luna or are still using GPT-5.6.

A post with images lists the Sol/Luna split, a 1.05M-token context, 128K output and API prices, noting cached reads get a 90% discount.

A screenshot post claims the Opus 5.5 system prompt is about 500K tokens on its own, while the GPT-6 Sol Codex system prompt is about 70K tokens.

A Japanese user says GPT-6 Sol and Luna were live on waking, citing DeepSWE scores and per-million-token costs, and calls Luna extremely good value.

The post argues Sol is really Terra renamed and distilled or quantized from Astra, using the 5.6 Terra and 6 Sol prices to explain the 50% cut.

Japanese outlet ITmedia reports that OpenAI released GPT-6 Sol and Luna, with API pricing at half the previous generation.

Someone put GPT-6 Sol, Luna, Opus 5.5 and Grok 4.7 into the Roboflow Playground to compare speed, price and accuracy on uploaded data.

Braintrust ran writing and problem-solving evals: Opus 5.5 led with 80.1% on tasks and 84.4% on writing, while Luna reached 78.0% and 83.8% at roughly 2% of the cost.

A Japanese ts-bench comparison found Opus 5.5 fully solved the benchmark, GPT-6 Sol improved steadily, and Luna performed notably worse than the previous generation.

A complaint post lists Claude Opus 5.5, GPT-6 Sol and Luna, MiMo-V2.6, Step 5 Preview and Grok4.7 launching in quick succession, contrary to claims that AI progress is slowing.

Both models are coming to Word, Excel, PowerPoint, Chat, Cowork and Copilot Studio, and admins need to sort out who can use each model and under which policy.