GPT-5.6 Sol Turns Blender Into an AI Speedrun

Why it matters

The GPT-5.6 Sol clip shows why agent performance is shifting from raw benchmark scores to latency, tool control, and the economics of repeated retries.

Chubby (@kimmonismus) pushed a Blender clip on X Friday showing GPT-5.6 Sol running at a claimed 750 tokens per second, a user demo that landed less than a day after OpenAI moved GPT-5.6 from limited preview into broader availability across ChatGPT, Codex, and the API.

The post matters because it frames GPT-5.6 around latency in a live tool loop rather than another table of scores. Chubby wrote that the clip was "not sped up" and later told a commenter it was running in "Very fast mode." The attached video came from [@evayzh]

https://x.com/kimmonismus/status/2075482486901969066

(https://x.com/evayzh/status/2075374512401711203/video/1), and Chubby later posted a correction after a reply from Sir Mr Meow Meow (@SirMrMeowmeow). The text available in the thread does not independently establish the full setup, the prompt, the machine, the connector used to control Blender, or whether the observed throughput was measured by OpenAI's tooling, a third-party interface, or the user.

OpenAI has supplied the part of the claim that makes the clip plausible. In its June 26th GPT-5.6 preview post, OpenAI said it was launching GPT-5.6 Sol on Cerebras in July at up to 750 tokens per second, with access initially limited to select customers. In its July 9th release post, OpenAI said GPT-5.6 was available starting that day across ChatGPT, Codex, and the API, with the global rollout continuing over the next 24 hours.

The Blender demo is useful because it shows where that speed starts to matter. Benchmarks compress work into a score. A modeling tool exposes the waits: generate code, run it, inspect the scene, fix geometry, retry. Blender is also unusually friendly territory for model-driven control because Blender's Python API gives scripts access to scene data, objects, meshes, and UI-exposed settings through the bpy module. A language model does not need to "see" every click if it can write and revise Python against that surface.

That makes the 750-token figure less like a vanity stat and more like agent infrastructure. A slow frontier model can be competent and still feel unusable when every tool call creates another pause. A faster model can attempt more edits, recover from more errors, and keep the user in the loop without turning the session into batch work. For design tools, CAD-like environments, spreadsheets, IDEs, and browser automation, throughput determines whether the product feels interactive or delegated.

OpenAI's own release page points to that direction. GPT-5.6 spans Sol, Terra, and Luna, with Sol as the flagship, Terra as the lower-cost tier, and Luna as the fastest and most affordable model. In ChatGPT Work and Codex, OpenAI says eligible paid users can choose among Sol, Terra, and Luna and set effort levels. In the API, developers get Sol, Terra, and Luna, plus Programmatic Tool Calling in the Responses API and a multi-agent beta that lets GPT-5.6 run concurrent subagents and synthesize their work in one request.

The economics are explicit. OpenAI priced Sol at $5 per 1 million input tokens and $30 per 1 million output tokens, Terra at $2.50 input and $15 output, and Luna at $1 input and $6 output. Cache writes are billed at 1.25 times the uncached input rate, while cache reads retain a 90% cached-input discount. Those numbers put pressure on developers to decide which jobs deserve Sol and which can be routed to cheaper models or cached context.

The same OpenAI page also gives the benchmark context that the viral clip sidesteps. OpenAI reports GPT-5.6 Sol Ultra at 91.9% on Terminal-Bench 2.1, Sol at 70.6% on BenchCAD and 83.4% on BenchCAD with a Python tool, and Sol at 62.6% on OSWorld 2.0. Those are OpenAI-published figures, and several are internal or tool-conditioned measures. They help explain why a Blender clip travels farther than a chart: it converts model performance into visible iteration.

There are still limits to what can be concluded from Friday's thread. The source post is a social video, not a reproducible benchmark. It does not show a full transcript, a timing log, the exact model selector, or the token meter behind the 750-token claim. It also does not prove that GPT-5.6 Sol can reliably perform professional animation or CAD work across arbitrary assets. It shows that, under one demonstrated setup, the model appeared to operate Blender quickly enough for the speed itself to become the news.

That is the real competitive pressure OpenAI is putting into the market with GPT-5.6. Frontier models are moving from answer engines into operators of software. Once the model is in the tool loop, latency, routing, cache behavior, and per-token price become product features. Friday's Blender clip made that legible in a way another benchmark row would not.