OpenAI has been shipping at an incredible pace on the coding front, and the latest developments around Codex and the GPT-5 model family are worth paying attention to — especially if you’re evaluating which AI coding tools fit your workflow.
GPT-5.3-Codex: The Latest Coding Model
Released just this month (February 2026), GPT-5.3-Codex is OpenAI’s most capable agentic coding model to date. The headline numbers: 25% faster than GPT-5.2-Codex, new records on SWE-bench Pro and Terminal-Bench, and it accomplishes all of this while using fewer tokens than previous versions.
What makes this release interesting beyond benchmarks is what OpenAI says about its development: GPT-5.3-Codex was instrumental in creating itself. The team used early versions to debug its own training, manage deployment, and diagnose evaluation results. That’s a milestone worth noting.
Key capabilities:
- Mid-turn steering: You can guide the agent while it’s actively working on a task without losing context. Ask questions, discuss approaches, redirect — all in real time.
- Frequent progress updates: The model communicates what it’s doing as it works, making it feel more like pair programming than fire-and-forget.
- Long-horizon execution: GPT-5.1-Codex-Max demonstrated the ability to work independently for more than 7 hours on complex tasks. The 5.3 version builds on this with better context compaction across multiple context windows.
Codex as a Platform, Not Just a Model
The bigger story is how Codex evolved throughout 2025 from a coding model into a full development platform. It now works across:
- Terminal (Codex CLI) — agent-style coding over real repositories
- IDE (VS Code extension) — inline assistance while you code
- Web (Codex app) — cloud-based tasks and code review
- Mobile (ChatGPT iOS) — manage tasks from anywhere
- GitHub — deep research connector for codebase analysis
OpenAI reports that 95% of their engineers use Codex weekly, and those engineers ship roughly 70% more pull requests since adopting it. Whether those numbers translate directly to external teams is debatable, but the internal adoption signal is strong.
The GPT-5 Model Family for Developers
The model lineup has expanded significantly:
- GPT-5.2 — the general-purpose powerhouse with frontier reasoning, tool-calling, and vision
- GPT-5.2-Codex — optimized for long-horizon agentic coding with context compaction
- GPT-5.1 — faster, more token-efficient, with dynamic thinking time based on task complexity
- GPT-5.3-Codex-Spark — research preview optimized for near-instant iteration (1000+ tokens/sec)
- GPT-4.1 — now available directly in ChatGPT, popular for everyday coding as a faster alternative
For developers building on the API, the model choices now come down to cost/latency/quality tradeoffs rather than fundamental capability gaps.
Apps SDK: Third-Party Apps Inside ChatGPT
OpenAI launched an Apps SDK that lets developers build interactive applications that run natively inside ChatGPT conversations. Early apps include Canva, Spotify, Zillow, and Booking.com. The SDK extends MCP (Model Context Protocol) so developers can define both logic and UI for their apps.
For developer tool builders, this opens up interesting possibilities — imagine debugging tools, deployment dashboards, or monitoring interfaces that live right inside your ChatGPT conversation.
What This Means for the LatAm Developer Market
For developers in our region, a few practical takeaways:
- Codex is included with ChatGPT Plus — no separate subscription needed. At $20/month, that’s significant value for individual developers.
- GPT-4.1 is now in ChatGPT directly — a fast, coding-optimized model that’s a great daily driver.
- The CLI is open source on GitHub — you can run it locally over your own repos with full human oversight.
- Windows support is improving — GPT-5.2-Codex was the first model natively trained for Windows environments.
The pace of releases (GPT-5 → 5.1 → 5.2 → 5.3-Codex in roughly 6 months) shows OpenAI is treating developer tools as a top priority.
Are you using Codex CLI or the ChatGPT coding features in your daily work? I’d love to hear how it compares to other tools you’ve tried — especially for those of us working across different tech stacks in Latin America.