[AINews] OpenAI reports median internal Codex output tokens grew 56x in Research, 32x in Customer Support, 27x in Engineering, and 13x in Legal since November 2025.

2026-06-26 · Source: Latent.Space - Www.latent.space · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Software Development & Engineering · Depth: Expert, extended

Summary

OpenAI's internal economic research indicates a significant surge in Codex token usage across its departments since November 2025, with median output tokens growing 56 times in Research, 32 times in Customer Support, 27 times in Engineering, and 13 times in Legal by June 2026. This trend highlights increasing internal adoption of AI agents for long-running, cross-functional tasks. Concurrently, the AI ecosystem saw rapid advancements in open models, including Z.ai's GLM-5.2 Max reaching 1595 on Code Arena: Frontend, and the launch of MIT-licensed Ornith-1.0 coding models (9B to 397B MoE). Google integrated computer use into Gemini 3.5 Flash, while Hugging Face announced surpassing \$100M annual run-rate, demonstrating the viability of open model platforms. Concerns about benchmark integrity and the growing importance of synthetic data generation and data curation were also prominent.

Key takeaway

For AI/ML Directors evaluating internal AI adoption, OpenAI's 13x-56x internal Codex token growth across departments by June 2026 highlights the rapid, often unexpected, utility of agents beyond coding. You should foster environments for experimentation and persistent workflows, supporting review loops and specialized tooling. Prioritize robust evaluation strategies, potentially partnering with external organizations, to avoid benchmark overfitting and ensure real-world performance.

Key insights

Rapid internal AI agent adoption at OpenAI, alongside advancements in open models and agent infrastructure, signals a significant shift in AI application.

Principles

• Scaling laws continue to drive AI capabilities.
• Separate eval creation from model optimization.
• Research taste and problem-solving are key.

Method

Replicate papers to develop research taste. Generate synthetic data via agent loops for better training/evaluation. Use data curation to induce concision and reduce serving cost.

In practice

• Explore GLM-5.2 for coding and agent tasks.
• Consider Ornith-1.0 for agentic coding.
• Utilize Qwen-AgentWorld for offline agent training.

Topics

• AI Agents
• Large Language Models
• Model Evaluation
• Open-Source AI
• Synthetic Data
• AI Infrastructure
• OpenAI Codex

Code references

• baidu/Unlimited-OCR

Best for: Executive, AI Product Manager, Product Manager, AI Scientist, Machine Learning Engineer, Director of AI/ML

Related on AIssential

• TAI #208: Open Models Find Their Role as Agent Token Bills Rise — Rising agent token costs necessitate a tiered model strategy, using cheaper open models for high-volume, verifiable tasks.
• [AINews] AI vs SaaS: The Unreasonable Effectiveness of Centralizing the AI Heartbeat — OpenAI's Codex 5.3 enhances coding with speed and mid-turn interaction, shifting focus to agent orchestration and user tools.
• The last six months in LLMs in five minutes — LLM capabilities, particularly coding agents and local models, advanced rapidly from November 2025 to April 2026, marked by frequent "best" model shifts.
• OpenAI Codex Can Now Ship Live Shareable Websites — AI tools are rapidly evolving to empower non-developers with direct, integrated, and automated capabilities for complex tasks.
• Claude Opus 4.6 and GPT-5.3-Codex 30 Mins Apart — AI agent teams are rapidly advancing, demonstrating autonomous complex task execution and self-improvement capabilities.
• OpenAI's AI data agent, built by two engineers, now serves 4,000 employees — and the company says anyone can replicate it — OpenAI's internal AI data agent demonstrates the transformative power of accessible data intelligence for enterprise operations.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Latent.Space - Www.latent.space.