OpenAI has merged the usage limits for its ChatGPT and Codex APIs, standardizing on the more restrictive ChatGPT-tier rate limits across both services. The change, noted in updated documentation and discussed on Hacker News this week, significantly reduces the available throughput for Codex users without warning.
Service: OpenAI API | Limit Type: Requests per minute (RPM) & Tokens per minute (TPM) | Codex New Limit: 10 RPM / 40,000 TPM | ChatGPT Limit: 10 RPM / 40,000 TPM
What Changed for Codex Users
Previously, Codex (the model family behind GitHub Copilot and the code-davinci-002 endpoint) operated under separate, more generous limits. Developers could request up to 20 requests per minute (RPM) and 150,000 tokens per minute (TPM) on its dedicated endpoints.
The new policy caps all code-* and gpt-3.5-turbo-instruct models at 10 RPM and 40,000 TPM—identical to the standard gpt-3.5-turbo chat completion limits. This is a 50% cut in request rate and a 73% reduction in token throughput for high-volume Codex applications.
How the New Limits Work in Practice
The limits are enforced per API key, per model. The 10 RPM cap is a hard ceiling; exceeding it triggers a 429 rate limit error. The 40,000 TPM limit is a rolling window.
For code generation, where prompts and completions can be lengthy, the token limit is the primary constraint. A single complex request can consume 4,000 tokens, allowing only ~10 such operations per minute before hitting the cap.
Bottom line: Codex lost its high-throughput status and is now governed by the same restrictive quotas as consumer-facing ChatGPT API calls.
Developer Reactions and Workarounds
The Hacker News thread, which garnered 63 points and 25 comments, revealed frustration from developers who built workflows around the higher Codex limits. Key feedback highlights the operational impact:
- Sudden breaking changes: Applications hitting rate limits without code alterations.
- Cost inefficiency: The lower TPM limit makes batch processing slower and more expensive due to idle time.
- Lack of communication: No official announcement preceded the documentation update.
Developers report two immediate workarounds: distributing load across multiple API keys or implementing aggressive client-side request queuing and retry logic. Both add complexity.
Pros and Cons of the Unified Policy
Potential Benefits:
- Simplified billing and management: One limit structure for all GPT-3.5-class models.
- Reduced infrastructure strain: May improve stability by capping high-volume Codex usage.
- Steers users to newer models: Aligns with OpenAI's push toward chat-completion formats.
Significant Drawbacks:
- Breaks existing applications: Codex-powered tools built for the old limits now fail under load.
- Hinders batch processing: Automated code review, translation, or documentation tasks are severely throttled.
- Erodes trust: Silent changes to core API terms disrupt developer planning.
Alternatives for High-Volume Code Tasks
For projects requiring sustained code generation or analysis, the new limits make native OpenAI Codex API impractical. Three alternatives now offer competitive or superior throughput.
| Feature | OpenAI Codex (New Limits) | Anthropic Claude 3 Haiku | Google Gemini 1.5 Flash | Open-Source (e.g., DeepSeek-Coder) |
|---|---|---|---|---|
| RPM Limit | 10 | 100+ (varies by tier) | 15 (free tier) | None (self-hosted) |
| TPM Limit | 40,000 | ~500,000 (Team tier) | 1,000,000 | Hardware Dependent |
| Best For | Prototyping, low-volume tasks | Production pipelines, large batches | Cost-effective experimentation | Full control, unlimited scale |
| Cost Factor | $0.12 / 1K tokens (output) | ~$0.25 / 1K tokens | $0.075 / 1K tokens | Compute/Infrastructure cost |
Anthropic Claude 3 Haiku provides a direct cloud alternative with higher limits and strong coding performance. Google's Gemini 1.5 Flash offers an exceptional 1 million TPM quota on its free tier, making it a compelling option for experimentation. For full control, self-hosting a model like DeepSeek-Coder on cloud GPUs eliminates rate limits entirely, trading API simplicity for infrastructure management.
"How to Check Your Current Usage"
Navigate to the OpenAI API usage dashboard. Use the rate limits endpoint programmatically: GET https://api.openai.com/v1/usage/limits. Monitor for 429 errors in your application logs, which signal limit breaches.
Who Should Use This (And Who Shouldn't)
Stick with the OpenAI Codex API if:
- You are in early prototyping or development.
- Your usage is sporadic, with bursts under 10 requests per minute.
- You are tightly integrated with the OpenAI ecosystem and value simplicity over cost/scale.
Switch to an alternative if:
- You run automated, batch-oriented code tasks (documentation, translation, review).
- Your application requires consistent, high-volume throughput.
- The sudden change violated your trust in API stability for core infrastructure.
The Bottom Line: A Strategic Shift, Not Just a Quota Change
This unification is a strategic consolidation, signaling OpenAI's deprioritization of the standalone Codex product line in favor of its chat-based model family. For the developer ecosystem, it's a stark reminder of the risks of building production infrastructure on a third-party API without contractual service-level agreements (SLAs). While the limits remain workable for light use, the silent change in terms erodes the foundational stability required for enterprise adoption.
The move actively pushes volume users toward competitors or self-hosted solutions, potentially fragmenting the toolchain OpenAI once sought to unify.
Top comments (0)