GPT-6 Sol leaks, screenshots and release rumors

OpenAI's changelog confirms an image encoding bug had lowered Sol and Luna image understanding across the API, Codex vision tasks and Computer Use; Luna's ReCoG visual grounding rose from 28.8% to 60.0% after the fix, and OpenAI advises rerunning evaluations for affected workflows.

andonlabs reran Blueprint-Bench for GPT-6 Sol and Luna because of an OpenAI bug: Sol improved slightly while Luna made a large jump, with score charts included.

A short post says GPT-6 Sol runs at 750 tokens per second on its high setting and looks forward to using it as subagents for Opus 5.5.

A post says a claude-sonnet-5-5 model ID has been spotted, with early tests claiming it beats GPT-6 Sol, comes close to Opus 5.5 and costs less.

The post says GPT-5.6 Sol ranks #19 on Text Arena while GPT-6 Sol drops to #60, calling it a regression for everyday conversation despite being cheaper and stronger at coding and agentic work.

An internal writing benchmark found higher reasoning effort generally gives modest gains at about 2.5x the cost, and GPT-6 Sol at max beats Astra while cheaper and faster.

A screenshot shows GPT-5.6 Sol around rank 19 and GPT-6 Sol around rank 60, which the author reads as a clear regression in ordinary conversation despite gains in coding and agentic work.

The VulcanBench author says Luna at low reasoning effort declares jobs done in about three minutes without solving them correctly, with Extra High and Max results still pending.

TokenGremlin says OpenAI keeps squeezing the old Doug pretraining base instead of building on the newer Bel, leaving Sol worse than GPT-5.6 Sol on some benchmarks with price as its main advantage.

A long post argues multi-turn agents destabilize after turn 20 and need routing separated from generation, assigning GPT-6 Sol to parallel extraction and Luna to verification while claiming 12.8x speed and 116x lower inference cost.

A user shares an arena ranking screenshot showing GPT-6 Sol at 60th in blind chat-only tests versus 19th for 5.6 Sol, and speculates this is why it wasn't rolled out in ChatGPT.

GPT-6 Sol reportedly scored higher than Astra on Vercel's security bug benchmark at the xhigh setting while costing one fifth as much to run, so the poster suggests trying it first for code audits.