Introducing Long-Term Support Model for Enterprises

GitHub, one of the most important platforms in coding, is making GPT-5.3-Codex its new base model for Copilot Business and Enterprise, and introducing GPT-5.3-Codex as the first Long Term Support (LTS) model for Copilot. It means GitHub customers are getting an automatic upgrade to a model built for serious software work, with stronger performance, a 12-month stability guarantee, and the kind of predictability and trust enterprises need. GPT-5.3-Codex is already the most-used OpenAI model among Copilot Business and Enterprise users. This move reflects an ask from developers who want the best model for the job. We’re excited to partner with GitHub to help bring better coding tools to enterprises around the world, and to make sure developers have access to the most capable models for their most important work.

Arguably one of the most challenging parts of the AI Era ™️ for enterprises is that things are changing quickly, making the important processes that vet new technology a bottleneck to productivity and innovation. With this long-term support model from OpenAI, enterprises can validate the model and reliably use it with the reassurance that it will be available and maintained through February of 2027.

This is a very special one 😄 GPT-5.3-Codex is our very first LTS model, you can expect to see it around for at least the next 12mo! This model has been a daily driver of mine since release; it's stable, it's fast, and it just gets work done.

AI is moving fast, and for enterprises, that pace can be as challenging as the opportunity itself. When technology changes faster than organizations can evaluate it, the process becomes the bottleneck. Long-term support helps give enterprises the stability to adopt with more confidence.

Solid move by #OpenAI. This is definitely going to be a great value addition to #Enterprise Adoption of #AIAssistedDevelopment

GitHub analysed 23M+ Copilot requests to answer a question benchmarks can't: Does AI-generated code actually survive into production commits? The findings are striking, frontier models perform far more closely than rankings suggest, and workflow type (exploratory vs. execution-focused) has a greater impact on code survivability than model choice itself. For tech leaders evaluating AI coding tools, this is a signal worth paying attention to; real-world performance in messy codebases tells a very different story than lab benchmarks. The takeaway reinforces GitHub's core principle: give developers the best tools and the freedom to choose, because no single model wins in every context. Thanks to Kyle Daigle for his insightful post.

The AI coding tool market just got genuinely competitive. For most of 2025 and into early 2026, Claude Code had a clear quality lead. Not a small one. Engineers who used it on real codebases knew the gap was wide. Complex reasoning, large context, instruction-following on hard refactors. Nothing else came close. That gap is closing faster than most people expected. GitHub Copilot Coding Agent is generally available now, with a browser, VS Code integration, and a proper async task model. OpenAI Codex CLI is open source and has been improving steadily. Gemini CLI has had a few quiet releases that surprised people paying attention. And Claude Code, at exactly this moment, has a documented quality regression. GitHub issue 42796 has 178 comments from engineers reporting the same failures since February. The community workaround is a single line in CLAUDE.md: behave like you did in January. That it works tells you something about how the model is being served under pressure. What I keep thinking about is the infrastructure math. Hundreds of thousands of developers adopted Claude Code over the last six months. Serving all of them at the quality they bought the product for requires enormous compute. Every major LLM company has gone through this scaling crunch. Anthropic will fix it. But the window where competitors were clearly behind has closed. Teams that assumed one tool stays dominant indefinitely are now rebuilding workflows mid-project. My answer to which tool to build around in Q3 has changed. Has yours? #ClaudeCode #AIEngineering #DeveloperTools

Excited to share what I've been working on at GitHub... Stacked PRs 🥞 Stacked PRs let you break a large code change into a series of smaller, dependent parts. Each one is focused and reviewable on its own, and together they tell a story of what you're building. This is more important than ever in a world of AI coding agents. The bottleneck is no longer writing code — it's reviewing it. Stacks help solve that: smaller, focused PRs mean higher quality reviews and faster time to merge ⚡ What's even cooler is seeing AI coding agents use stacks themselves. They can use the CLI to break up their own work into focused, reviewable PRs as they build. Our goal is to make stacks work seamlessly across all of GitHub. That means making sure branch protection rules, Actions/CI, code scanning, reviews, rebasing, merging, etc. "just work" as you'd expect. More to come soon ✨

"Hey here's 1432 lines of code, please review by today" One of the questions that's been on my mind lately is - how do I break this PR down so the reviewer can actually review it instead of "scroll, scroll, scroll, approve" - because the PR is too large? One option is to keep a set of commits ready - and then raise the PRs commit by commit, so the reviewer is not overwhelmed But then that gives rise to the problem - if I have Commit1 Commit2 Commit3 Now when commit1 gets merged (and say if some other change comes up in between) - I need to rebase - which means I need to get the rebase reviewed -> approved Interesting that the Github team has thought of that as well and added a "cascading rebase" option. Good stuff! #github #stackedPRs

The new Rubber Duck feature in GitHub Copilot CLI automatically uses a second AI model to review the first model’s work. This helps catch mistakes that a single model might miss. I already use this approach manually. I use Claude Opus 4.6 for planning, Codex 5.3 for coding, and GPT-5.4 for final review. This multi model workflow improves code quality because each model brings a fresh perspective and can catch different kinds of bugs. https://lnkd.in/dAr5HJfn