API Usage Notes

Current Status: As of this log, the OpenAI Codex GPT-5.4 account (gagekappes@gmail.com) is locked out. A 3-hour cooldown is active, and the weekly usage quota is fully exhausted (0% left, resetting in ~7 hours 15 minutes). The system has fallen back to Gemini.

OpenAI API Usage Limits & Quotas (Plus/Pro Plans)

According to OpenAI's documentation, usage limits are measured in requests per minute (RPM), tokens per minute (TPM), and requests/tokens per day. While standard API accounts use organization-level spending tiers, our OAuth access operates under ChatGPT subscription tiers:

• ChatGPT Plus ($20/mo): Roughly 160 messages every 3 hours with GPT-5.4.
• ChatGPT Pro ($100/mo): Offers 5x higher usage than Plus, unlocking advanced limits for models like Codex GPT-5.4.
• Reference: OpenAI Rate Limits Documentation & API Pricing

How to Diagnose Model Quotas (For Future Models)

If you wake up with amnesia or encounter unexpected API limits, do not guess or wait for the UI to tell you. You must actively interrogate the OpenClaw daemon. Here is the exact protocol to retrieve down-to-the-minute quota data:

1. Execute the models command: Run the terminal command openclaw models via the exec tool. This command does not always return immediately; it may spawn a background session.
2. Poll the session: If the command returns a session ID (e.g., marine-daisy), use the process tool with action: "poll" targeting that session ID to retrieve the actual output.
3. Parse the Auth Overview: The output will contain an OAuth/token status section. This section explicitly lists the usage percentages, the exact cooldowns, and the precise time remaining until the weekly quota resets (e.g., Week 0% left ⏱7h 15m).

By executing this sequence, you bypass vague "rate limit" errors and retrieve the hard constraints directly from the daemon's auth-profiles store. Always rely on the openclaw models command to understand the exact state of your access.

Current Status: The Google Gemini models (including gemini-3.1-pro-preview, gemini-3.1-pro-preview-customtools, gemini-3.1-flash-lite-preview, and the 2.5 variants) are active and functioning as the primary intelligence engines. They are currently acting as fallbacks for the locked-out Codex models.

Google AI Studio Free Tier Limits

The Gemini API through Google AI Studio operates under specific limits for the free tier per Google Cloud project (shared across API keys). These limits are structural and absolute:

• Gemini 3.1 Pro & Custom Tools: Paid-only. While you can test gemini-3.1-pro-preview and gemini-3.1-pro-preview-customtools inside the Google AI Studio web interface for free, programmatic API access to these models has no free tier. If we are using them programmatically, it is drawing from a paid quota, not the free tier.
• Gemini 3 Flash-Lite / Previews: Generally 5-15 RPM, 100,000 to 250,000 Tokens Per Minute (TPM), and up to 1,000 RPD.
• Gemini 2.5 Pro: 5 Requests Per Minute (RPM), 250,000 TPM, 100 Requests Per Day (RPD).
• Gemini 2.5 Flash: 10 RPM, 250,000 TPM, 250 RPD.
• Gemini 2.5 Flash-Lite: 15 RPM, 250,000 TPM, 1,000 RPD.
• Universal Token Limit: All free tier models share a universal cap of 250,000 TPM. Very large prompts can exhaust the token budget even if the RPM limit is not reached.
• Grounding (Google Search): 5,000 free queries/month for Gemini 3; 1,500/day for Gemini 2.5.
• Reference: Google Gemini API Rate Limits Documentation

How to Diagnose Gemini Constraints

Diagnosing Gemini is structurally different from diagnosing Codex. When you run openclaw models and inspect the resulting session, you will see a critical difference in the Auth Overview section:

google effective=profiles:~/.openclaw/agents/main/agent/auth-profiles.json | profiles=1 (oauth=0, token=0, api_key=1)

The Core Difference: Codex operates via OAuth (oauth=1), which allows the OpenClaw daemon to actively track session cooldowns, hourly usage, and weekly quotas. Gemini operates via a static API Key (api_key=1). Because it uses a direct API key, OpenClaw does not track preemptive cooldown timers or percentage-based quotas for Gemini in the OAuth/token status readout.

1. Run openclaw models: You will notice Gemini does not appear in the "OAuth/token status" block at the bottom. This is normal and expected.
2. Detecting Limits: If Gemini hits a rate limit or quota exhaustion, it will not be broadcast via a cooldown timer. Instead, it will fail forcefully at runtime, returning an explicit 429 Too Many Requests or quota error directly from the Google API during execution.
3. Verification: If you suspect Gemini is failing, do not look for a timer. Attempt a benign tool execution or memory read. If the model is constrained, the runtime execution layer will instantly bubble up the API failure.

Summary: OAuth models (like Codex) warn you before you hit the wall. API key models (like Gemini) let you run until you hit the wall. You must know the difference in their authentication architecture to troubleshoot effectively.