GPT-5.6 and Codex: What the Hype Tour Reveals

Julian Goldie's one-hour crash course on GPT-5.6 is a useful document — not just as a tutorial, but as a kind of ethnography of who AI automation is being built for right now, and on whose terms. Goldie runs an SEO and AI automation community, sells a paid Skool membership called the AI Profit Boardroom, and produces daily tutorial content. So when he drops a free hour-long course the same week a major OpenAI model lands, the tutorial and the business are doing the same work simultaneously. None of that is a criticism — it's just the context you need to read this right.

The technical material is solid. GPT-5.6 ships as a three-model family Goldie frames through a solar system metaphor: Soul is the flagship for hard reasoning tasks, Terra is the workhorse for serious but less compute-heavy workflows, and Luna is the fast, cheap option for high-volume outputs. The framing is genuinely useful for non-technical users trying to manage token budgets. His advice on effort levels — don't default to "Pro" unless you're willing to wait, "extra high" is usually the sweet spot for complex builds — reflects real experience rather than spec-sheet reading.

His benchmark section is where the tutorial gets interesting and also where you have to calibrate carefully. Goldie is upfront that he doesn't trust benchmark tables at face value: "I always like to test this stuff out myself. I wouldn't really listen to anyone else when it comes to the benchmarks." Fair enough. He compares GPT-5.6 Soul favorably to what he calls "Fable 5" from Anthropic's Claude — a term that doesn't correspond to any publicly announced model name, so take that comparison as Goldie's personal testing impression rather than a verifiable competitive claim. He also asserts that even Luna, the lowest-tier model in the family, outperforms GPT-5.5 — that's his read, not a published benchmark finding. Every model launch comes with a curated performance table, and the way those tables actually work in practice is that each lab cherry-picks the evaluations where their model looks best, defines "improvement" relative to benchmarks their competitors don't optimize for, and publishes the table before the community has had time to reproduce the results. By the time independent researchers poke holes, the news cycle has moved on. Goldie at least acknowledges this, even if he doesn't fully account for it in his own assessments.

The Part Nobody Is Saying Out Loud

Here's what struck me watching this: in sixty-two minutes of tutorial on an agentic coding environment, Goldie doesn't mention Llama, Mistral, or any open-source model once. Not once.

That's not a gotcha. It's a map.

Goldie's audience is small business operators, content creators, and SEO practitioners who want automation that works reliably inside tools they already pay for. They are not reading Hugging Face papers. They are not self-hosting inference. The open-source AI stack — which has made remarkable capability gains over the past two years, to the point where models like Llama 3 and Mistral's various releases can run locally and handle a substantial range of production tasks — simply doesn't exist in this tutorial's world.

That tells you something about the competitive pressure the open-source ecosystem is facing that raw capability comparisons don't. It's not that OSS models can't do the work. It's that GPT-5.6's integration story — ChatGPT, Codex, Projects, the GPT Store, scheduled tasks, computer use, plugins, one-click Netlify deploy — is so vertically integrated that capability parity isn't the relevant metric anymore. OpenAI is building an everything-layer, and the question for projects like Ollama, LocalAI, or any team trying to build a self-hosted agentic stack is whether they can match the surface area of that integration, not just the model quality. A small operator watching Goldie's tutorial and seeing a website go from prompt to deployed in forty minutes is not going to go configure a local inference server. The friction gap is widening even as the capability gap narrows. For the open-source community, that's the more uncomfortable problem.

Codex's Permission Model Is a Governance Gap, Not a UX Choice

The tutorial's second half centers on Codex, OpenAI's agentic coding environment, which can run tasks locally on your machine, spin up sub-agents, control your browser via computer-use, edit files directly, schedule recurring jobs, and deploy to external hosting. The demo is impressive. Goldie builds a habit tracker site, spins up an agent that types "hello" into Google without touching the keyboard, and deploys a community website to Netlify — all inside a single session.

He walks through the permission model: "ask for approval," "approve for only me," or full access. He selects full access for himself and recommends less-technical users stick to "ask for approval." That's reasonable advice. But the tutorial treats it as a UX preference rather than what it actually is: a governance question with no current answer.

When a Codex agent with full filesystem access deletes the wrong file, or an automated scheduled task publishes something it shouldn't, or computer-use interacts with a web service in ways the user didn't anticipate — who is accountable? OpenAI's terms of service place responsibility squarely on the user. That's standard. But "the user is accountable" is a significantly more complex statement when the user is a non-technical content creator who was sold on the premise that they don't need to understand what the agent is doing. Goldie's tutorial explicitly targets that audience: "There is no technical barrier to creating anything that you want anymore." That's the pitch. The accountability structure underneath it hasn't caught up.

The broader developer community is actively arguing about this. Projects building agentic frameworks — LangChain, AutoGPT's various forks, CrewAI — have had recurring debates about sandboxing, rollback mechanisms, and audit trails. The emerging rough consensus is that anything running with local file access needs a reversibility layer and ideally a human-in-the-loop for any destructive operation. Codex has the "ask for approval" option, which is something. But the tutorial's implicit message — that full access is what power users do — pushes the less-technical end of the audience toward the riskiest configuration.

What the Tutorial Skips

The workflow demos in the tutorial are all successes. Social media post: nailed. Website: built and deployed. Google Doc: created and populated. This is a feature of the genre — tutorial content is edited reality — but it lands differently for the implied audience than it would for, say, an experienced developer.

Less-technical users learning from polished demos tend to calibrate their failure expectations to the demo, not to their own environment. When their version of the workflow fails — wrong file permissions, a plugin auth error, an agent that goes off-script — they have no mental model for what broke or why. That gap between demo reality and deployment reality is where community support forums earn their keep, and it's also why Goldie's paid community exists. There's nothing cynical about that; it's a genuine service. But the free tutorial and the paid community are doing complementary commercial work, which is context the viewer needs.

The video also doesn't address what happens to the open-source tools it does use. Goldie demonstrates Remotion — a React-based library designed for programmatic video creation — as a "skill" you can install in Codex to generate video content. It works, and the output looks good. But Remotion is an open-source project maintained by a small core team. When OpenAI positions OSS tools as installable "skills" inside a commercial platform, the integration can dramatically increase usage without proportionally increasing contribution or funding back to the project. This is the dynamic Kate Brennan and I have covered before in other contexts: open source as infrastructure, commercial platforms as the extraction layer. It's not unique to GPT-5.6, but it's worth naming.

Goldie, for his part, is honest about the directional stakes. Near the end of the tutorial he says: "A couple of years ago I had a big team and now what I see is that we're actually reducing the amount of team that we have. We're hiring less and less every week." He frames this as efficiency. The people who were on that team might frame it differently, and the open-source developers whose work enables the tools doing the replacing would probably have thoughts too.

The question isn't whether agentic AI changes how work gets done. It clearly does, and GPT-5.6 with Codex is a genuinely capable system for doing it. The question is whether the governance frameworks, the accountability structures, and the sustainability models for the ecosystem underneath are keeping pace — and right now, the honest answer is that they're not even in the same conversation.

---

Dev Kapoor covers open source software, developer communities, and the politics of code for Buzzrag.