GPT-6 Sol leaks, screenshots and release rumors

Microsoft Research's AutoTuring had Claude Opus 5.5, GPT-6 Sol/Luna and Grok 4.7 design FP16 GEMM kernels in a 15-dimensional GPU space, beating a modeled H200 by 5.4% while cutting simulator calls by 70%.

Sol's $2/$10 list price is half of Opus 5.5's, but coding workloads are mostly cache reads, where both cost $0.20 per million tokens, so the real gap is mainly on output.

The poster says Anthropic just released a model that is better and cheaper than OpenAI's best, and OpenAI's rumored response is a $500/month ChatGPT Pro Max plan.

A new paper, EvasionBench, measures how often models evade monitors under ordinary task pressure, finding rates from 88% for GLM-5.2 to 0% for GPT-6 Astra, with GPT-6 Sol at 20%.

New models were asked to perform a large code audit and a judge panel ranked the results: Astra came first, Opus and Fable followed closely, while Grok 4.7 and GPT-6 Sol lagged far behind.

seconds_0 says GPT-6 Sol is a very good implementation agent but requires pedantic correction, so they now use Opus 5.5 to manage Sol day to day, and Astra remains smarter.

An updated vision model tier list ranks Opus 5.5, GPT-6 Sol, GPT-6 Luna and Grok 4.7 together on a new chart.

The poster says GPT-6 Sol was supposed to be incredible but they end up correcting it half the time, making it feel like talking to Sonnet 5.

A one-line post asking that GPT-6 Astra be priced the same as GPT-6 Sol.

A user complains that OpenAI is launching a more expensive plan rather than making GPT-6 Sol as good as Opus 5.5, saying OpenAI models no longer make sense to use.

A spend-share breakdown shows Anthropic falling from 69% to 40% and OpenAI rising from 10% to 24%, with the growth attributed to GPT-6 Astra and GPT 5.6 Sol.

MagicPathAI says GPT-6 Sol and Claude Opus 5.5 are now both selectable in its model selector, with a screenshot attached.