In this video, Matthew Berman reviews GPT-5.6, framing it not as a minor point release but as a major leap over GPT-5.5. His thesis is that OpenAI has "squeezed every drop of juice" out of the GPT-5 training run, producing one of the most capable and effective models he has ever used. He backs the claim with two long-running agentic demos, a look at browser and computer use, third-party enterprise benchmarks from box, updated pricing, and a walkthrough of the model's three sizes and reasoning levels. He closes by contrasting 5.6 with a newer model called Fable.
Berman gave 5.6 plus Codex a simple /goal: make an Excel clone and continue until feature parity. The roughly eight-word prompt ran for over five days before he manually stopped it. The resulting single-page HTML app supports sorting (ascending and descending), live formula entry shown in the input bar, data validation, conditional formatting, tables, find and replace, and even pivot tables built from selected data with headers. He notes there are still rough edges and that it was nowhere near done when he stopped it. Notably, the agent used computer use, opening the real Excel on his desktop and going back and forth between the real app and the cloned version.
Berman describes browser use as an area where 5.6 is "phenomenal." He says Codex's browser has increasingly become his main driver because he can get real work done with it, such as having it sort through Gmail or perform complex DNS record changes from a single prompt.
Using /goal again with the prompt "create a clone of Minecraft, feature parity," the agent ran about seven days. It looked like real Minecraft after roughly one day, then kept going deeper, building out world regions, mobs, and biomes. Berman calls it the best AI-generated Minecraft he has made, citing nice shadowing, 3D block-mining animation, farmland and carrots, a tabby cat, breakable glass, seed generation, and an easy-to-use inventory.
Box, the video's sponsor and hosting partner for the demos, ran its "complex work" evaluation, which tests real knowledge work like reading documents, reconciling numbers, doing due diligence, and reviewing expert output for errors. Reported accuracy: GPT-5.5 at 63.3 and Tara at 59% (a drop from Soul). GPT-5.6 Luna scored roughly the same as Tara but much faster and less expensive. Across industry subsets (public sector, life sciences, healthcare), Soul dominated GPT-5.5. All three models are coming to Box AI.
GPT-5.6 is both cheaper and more token-efficient than Fable: $5 per million input tokens versus $10 for Fable (with cheaper cache hits), and $30 versus $50 per million output tokens. Berman says 5.6 feels like it has a more direct line of sight to accomplishing a task than Fable.
Berman frames Fable as feeling like a brand-new model that "sees around corners" better than 5.6. His analogy: 5.6 is a maximally souped-up Honda Civic where every bit of horsepower has been squeezed out, while Fable is a Ferrari fresh off the line, unoptimized but with far higher ceiling. He points to a separate full review with more demos, including an operating system and a Rube Goldberg lab.
Unlike Fable, 5.6 comes in three sizes: Luna (smallest), Tara (medium), and Soul (largest). Each also offers multiple reasoning levels, with Soul going up to "ultra" (described as a quota burner). This enables model routing: plan with Soul, do most implementation with Tara on high reasoning, and offload low-requirement work like deploying to Luna. Berman mentions he wrote a delegation skill (linked on GitHub) so you can route among these models within Codex to save quota while keeping similar quality.