AI Models Need LTS Releases
GitHub quietly announced something interesting in its Copilot change log: GPT-5.3-Codex is now its first long-term support model.
GPT-5.3-Codex is also our first long-term support (LTS) model, in partnership with OpenAI. LTS models are guaranteed to be available for a full 12 months from the model’s launch date, giving enterprises the stability they need for internal security and safety reviews. GPT-5.3-Codex launched on February 5, 2026 and will remain available through February 4, 2027 for Copilot Business and Copilot Enterprise users.
I have seen model deprecation policies before, but LTS support seems like a different framing of it.
I have seen firsthand that once you build workflows, agents, evals, internal docs, security reviews, and operational assumptions around a specific model, you get pretty tied to a model. Yes, sure a good regression eval suite can help you pin down the problems in a new model upgrade, but it still needs a fair amount of time optimizing the prompts to work well with the new models, getting security audits and other approvals. These business processes work slower than the current speed in which new models are announced.
I feel that while we wish for the rest of the organization to keep up with the engineering velocity of the new agentic engineering world, enterprises that owe customers strong guarantees, and industries where predictable and technical accuracies are paramount, like pharma research, etc., will always keep humans in the loop, and humans will always be the slowest piece.
A newer model can be better overall and still break your workflow. It might follow instructions differently. It might call tools in a different order. It might become more verbose, change how cautious it can be, and in general more willing to reinterpret a prompt that used to work. None of that necessarily means the model is worse, but it can absolutely break a stable deployed agent pipeline.
To be clear, I think model support cannot be indefinite. I imagine that providers will still modify things and add escape hatches for safety, abuse, compliance, and security problems.
Code survival rate
The other interesting phrase in GitHub’s post is “code survival rate.” In GitHub’s telling, GPT-5.3-Codex has a significantly high code survival rate among enterprise customers. I do not know exactly how they define that internally, but I assume it means some version of: code generated or assisted by the model remains in the codebase over time instead of being quickly rewritten, reverted, or repaired.
IMO that is a much better metric than lines of code generated. Generating lines of code is cognitively cheap (it definitely costs money, we can just generate it more mindlessly)