Quick answer: ChatGPT has multiple independent constraints. Match the exact plan, model, runtime, invocation mode, or server configuration shown in each row.
Current documented limits
| Constraint | Current value | Scope | Verified source |
|---|---|---|---|
| General per-file hard limitApplies to files uploaded to a GPT or ChatGPT conversation; more restrictive type-specific limits can apply first. | 512 MB | Platform documented | OpenAIAug 22, 2026 |
| Text and document token capDoes not apply to spreadsheets; the byte-size limit still applies. | 2,000,000 tokens per file | Platform documented | OpenAIAug 22, 2026 |
| CSV or spreadsheet practical capApproximate because the supported size depends on row size. | Approximately 50 MB | Platform documented | OpenAIAug 22, 2026 |
| Image limitType-specific image limit. | 20 MB per image | Platform documented | OpenAIAug 22, 2026 |
| Rolling upload allowanceOpenAI says it may lower this limit during peak hours; Free users have a separate daily cap. | Up to 80 files every 3 hours | Platform documented | OpenAIAug 22, 2026 |
| Free plan upload allowanceA plan-specific allowance, separate from storage and per-file constraints. | 3 file uploads per day | Free | OpenAIAug 22, 2026 |
How to apply ChatGPT limits safely
The monitored baseline covers General per-file hard limit, Text and document token cap, CSV or spreadsheet practical cap, Image limit, Rolling upload allowance, Free plan upload allowance. Treat these as separate constraints rather than one platform-wide capacity number: a workload can fit one row and still fail another because the plan, model, endpoint, runtime, region, invocation mode, or account scope differs.
- Match the production workload to the exact scope printed beside each value and confirm it in the active ChatGPT console, configuration, or response headers.
- Measure the serialized request, token volume, duration, concurrency, storage, or connection demand at realistic percentiles, then preserve headroom for bursts and retries.
- Check every adjacent layer—client, SDK, gateway, proxy, queue, database, and downstream service—for a smaller effective limit before changing architecture.
Specific limit pages
These pages exist because the constraint has a distinct implementation or troubleshooting intent. Closely related keyword variations stay consolidated.