Right after Google dropped Gemini 3 Pro and its agent-first developer tools, OpenAI fired back on the coding front, releasing GPT-5.1-Codex-Max, its most capable agentic coding model to date.
What’s new
• Stronger coding performance: Codex-Max beats both Codex-High and Google’s Gemini 3 Pro on coding benchmarks.
• 24-hour sessions: A new compaction system lets the model prune history while keeping context, enabling multi-million-token tasks that run for over a full day.
• More efficient: Uses ~30% fewer tokens while completing complex dev tasks much faster than previous models.
• Built for real workflows: Trained on practical developer tasks across Windows and web environments: writing, debugging, testing, and research.
• Available now: Live in the Codex CLI and IDE extensions for Plus, Pro, Business, Edu, and Enterprise. API access is coming soon.
How this fits the bigger picture
Google used Gemini 3 to push ahead on multimodal reasoning, UI generation, and agentic workflows.
OpenAI is pushing deeper into long-form software automation, optimizing for complex coding sessions instead of one-shot code generation.
Both companies are building agents but they’re aiming at different layers of the stack:
• Google: fast prototyping, interfaces, orchestration
• OpenAI: long-horizon execution, deep code reasoning
Why it matters
Codex-Max signals that OpenAI isn’t conceding the engineering space to Google’s Antigravity. If Gemini 3 showed what agents see and design, Codex-Max shows what agents can actually build over time.
The model wars just entered another round…fast.
Leave a Reply