OpenAI Codex Hits Capacity Limits as Demand Surges

OpenAI's Codex coding agent is experiencing widespread capacity issues, with users reporting persistent failures and the dreaded 'Selected model is at capacity' error message. The reliability problems have sparked frustration across the developer community, raising critical questions about whether OpenAI's infrastructure can keep pace with explosive demand for its AI-powered coding tools.

The outages appear to be tied directly to surging usage of Codex, OpenAI's cloud-based software engineering agent that launched in mid-2025. Users report that tasks frequently fail to complete, with the system returning capacity errors instead of results — sometimes repeatedly throughout the day.

Key Takeaways

• OpenAI Codex users are experiencing frequent 'model at capacity' errors
• The failures appear linked to overwhelming demand for the coding agent
• Users report that service was noticeably smoother during a period when OpenAI purged team accounts in bulk
• The capacity crunch raises questions about OpenAI's infrastructure scaling strategy
• Developers relying on Codex for production workflows face unpredictable downtime
• The issue highlights a growing tension between rapid feature launches and backend reliability

'Selected Model Is at Capacity' Becomes a Familiar Sight

The error message has become almost a meme among OpenAI power users. When attempting to run Codex tasks — ranging from code generation to debugging and repository-level changes — developers are increasingly met with a blunt notification that the selected model cannot handle their request.

This is not a new phenomenon for OpenAI. The company faced similar capacity crunches during the initial launch of ChatGPT in late 2022 and again when GPT-4 rolled out in March 2023. However, the current Codex bottleneck feels different to many users because it affects a tool specifically designed for professional software engineering workflows, where reliability is not optional.

Unlike casual ChatGPT conversations, Codex tasks often involve complex, multi-step operations across entire codebases. A failure mid-task does not just mean a minor inconvenience — it can disrupt development pipelines and waste significant time.

A Brief Golden Age After the Team Account Purge

One revealing detail has emerged from user reports: the service ran noticeably better during a period when OpenAI conducted a bulk termination of team accounts. During that window, Codex was reportedly fast, reliable, and responsive — a stark contrast to the current experience.

This observation strongly suggests that the capacity issues are fundamentally a supply-and-demand problem. When a significant number of users were temporarily removed from the system, the remaining users enjoyed smooth performance. Now that demand has presumably recovered — or grown even further — the infrastructure is buckling under the load.

The pattern reveals several important dynamics:

• OpenAI's compute infrastructure has hard limits that active users can saturate
• The company may be onboarding users faster than it can provision GPU capacity
• Premium tier users (Pro, Team, Enterprise) are not immune to capacity constraints
• Codex's architecture likely requires significantly more compute per task than standard chat interactions

Why Codex Is Especially Compute-Hungry

Codex is not a simple chatbot interaction. Launched as a cloud-based coding agent, it operates inside a sandboxed environment where it can read entire repositories, write code, run tests, and iterate on solutions autonomously. Each Codex task can involve multiple model calls, tool usage, and extended reasoning chains.

Compared to a standard GPT-4o conversation that might involve a single prompt-response cycle, a Codex session can consume 10x to 50x more compute resources. The agent might make dozens of internal calls to the underlying model as it navigates a codebase, plans changes, implements them, and verifies the results.

This compute intensity means that even a modest increase in concurrent Codex users can have an outsized impact on infrastructure load. OpenAI's systems need to handle not just more users, but exponentially more compute per user when Codex is involved.

The problem is compounded by the fact that OpenAI has been aggressively promoting Codex as a flagship product, encouraging developers to integrate it into their daily workflows. Success in adoption has become a victim of its own momentum.

The Broader Infrastructure Challenge Facing OpenAI

OpenAI's capacity struggles with Codex reflect a larger challenge that the company — and the entire AI industry — is grappling with. The demand for GPU compute continues to outstrip supply, despite massive investments from cloud providers and chip manufacturers.

OpenAI reportedly spends billions of dollars annually on compute infrastructure, primarily through its partnership with Microsoft Azure. The company has also explored deals with other cloud providers and is investing in custom hardware initiatives. Yet these efforts take time to materialize, and demand keeps accelerating.

Several factors are intensifying the compute crunch:

• Agentic AI workloads like Codex require sustained, multi-step compute rather than single inference calls
• OpenAI's user base continues to grow rapidly, with over 400 million weekly active users reported in early 2025
• Enterprise customers are deploying AI at scale, consuming large blocks of capacity
• New features and models (like o3 and GPT-4.1) attract usage spikes upon launch
• Competing priorities between consumer products (ChatGPT) and developer tools (API, Codex) create internal resource allocation tensions

How This Compares to Competitors

OpenAI is not the only company facing infrastructure scaling challenges, but its competitors have taken different approaches that may offer more consistent availability.

Anthropic's Claude, for instance, offers a similar coding experience through Claude Code, its CLI-based coding agent. While Claude has also experienced rate limiting, its architecture as a terminal-based tool means individual sessions tend to be more lightweight than Codex's cloud-sandboxed approach.

Google's Gemini benefits from the company's massive internal infrastructure and TPU clusters, giving it a potential advantage in raw compute availability. Google's Jules coding agent, while newer, has not reported the same level of capacity issues.

GitHub Copilot, powered by a combination of OpenAI and Anthropic models, distributes load across multiple providers, which can provide more resilience against single-provider capacity limits.

For developers who need reliable AI coding assistance, the capacity issues with Codex may push some to evaluate alternatives — at least as backup options for when OpenAI's systems are overloaded.

What This Means for Developers and Businesses

The practical implications of Codex's capacity problems extend beyond mere inconvenience. For teams that have built Codex into their development workflows, unpredictable availability creates real business risk.

Developers should consider several strategies to mitigate the impact:

• Diversify AI coding tools: Maintain familiarity with at least 1 alternative (Claude Code, Copilot, Cursor) as a fallback
• Schedule intensive tasks strategically: Capacity tends to be more available during off-peak hours (late night US time, weekends)
• Use the API directly: In some cases, direct API calls to GPT-4.1 or o3 may be more available than the Codex agent interface
• Cache and checkpoint work: Design workflows that can resume after failures rather than starting from scratch
• Monitor OpenAI's status page: Stay informed about known outages and planned maintenance windows

For enterprise customers paying $200 per month for ChatGPT Pro or significant sums for team and enterprise plans, the capacity limitations are particularly frustrating. These users expect premium reliability in exchange for premium pricing.

Looking Ahead: Can OpenAI Scale Fast Enough?

OpenAI is undoubtedly aware of the capacity problem and is likely working to address it. The company has a track record of eventually scaling through infrastructure challenges — the ChatGPT launch bottleneck in 2022-2023 was eventually resolved, and GPT-4 availability improved significantly within months of launch.

However, the Codex situation is more complex because agentic workloads represent a fundamentally different scaling challenge. Each new user does not just add 1 unit of demand — they add a variable, potentially large amount of compute consumption depending on the complexity of their tasks.

OpenAI's reported $6.6 billion funding round and its partnership expansions with Microsoft suggest the company has the financial resources to invest in infrastructure. The question is whether hardware procurement and deployment can keep pace with the exponential growth in agentic AI usage.

In the near term, users should expect continued capacity fluctuations. OpenAI may implement more aggressive rate limiting, tiered access based on subscription level, or queue-based systems to manage demand. The company might also optimize Codex's underlying architecture to reduce per-task compute requirements.

The Codex capacity crunch is ultimately a sign of success — the product is so useful that demand exceeds supply. But for the developers stuck staring at 'Selected model is at capacity' error messages, that is cold comfort. OpenAI's ability to resolve this bottleneck will be a key test of whether it can deliver on the promise of AI-powered software engineering at scale.