Codex for the Individual, GitHub Copilot for the Enterprise

A few weeks ago I wrote about why I had moved most of my coding work from Claude Code to OpenAI Codex.

Since then I have spent more time looking at GitHub Copilot, and I think there is an important distinction between what works best for an individual developer and what makes sense for a large engineering organisation.

For my own work, I still prefer Codex.

The new desktop application is excellent. The agentic workflow is very well thought through, particularly the ability to run multiple agents independently, each in its own worktree. I can give an agent a substantial task, leave it running and come back later to review what it has done.

GPT-5.6 Sol is also at least competitive with Anthropic’s models for most coding tasks I do. My own experience is that it is less opinionated, follows instructions more closely and is easier to trust with long-running tasks without constantly having to steer it.

For an individual or small business, the economics are also very attractive. Codex is included within the ChatGPT subscription, normal ChatGPT usage does not consume the Codex allowance, and OpenAI has been fairly generous with additional usage resets recently.

For a large enterprise, though, I think GitHub Copilot is becoming the more interesting proposition.

The main reason is that GitHub is increasingly separating the coding platform from the underlying model. An organisation can use approved models from OpenAI, Anthropic and others through the same environment, and can also support BYOK or local/private models where required.

That means the enterprise does not have to make a long-term bet on one model provider.

At the same time, GitHub already has the things large engineering organisations need around the models: identity, repository permissions, model controls, usage policies, budgets, auditability, PR workflows, reviews, CI/CD integration and enterprise privacy controls.

Importantly, GitHub states that Copilot Business and Enterprise customer data is not used to train AI models.

Yes, Copilot is more expensive than a personal Codex subscription, and heavy agent use can still be metered. But I don’t think that is really the right comparison. The flat-rate economics available to an individual are not the same economics available when deploying coding agents across a large organisation.

The conclusion I am coming to is that enterprises probably should not standardise on a particular LLM at all.

Models are changing too quickly.

It makes more sense to standardise the engineering platform, security controls, agent permissions, repository instructions, testing and governance, while keeping the underlying model replaceable.

Right now, GitHub looks to be in a very strong position to become that platform.