You’re in the middle of a detailed response from ChatGPT or Claude—maybe a code review, a long report, or a multi-step analysis—and suddenly it stops. No explanation, no warning. Just… done.
This isn’t a bug, and it’s not the model being lazy. Every AI model has hard limits on how much text it can generate in a single response, and hitting that ceiling is more common than you’d think.
What’s Actually Happening
AI models have two separate limits that matter: the context window (how much text the model can read and remember) and the output token limit (how much it can write in one go).
Most people know about context windows—ChatGPT’s ranges from 8,000 to 128,000 tokens depending on the model, Claude can handle up to 200,000, and Gemini 1.5 Pro goes up to 2 million. But output limits are different, smaller, and enforced separately.
For example, GPT-4 typically caps output at around 4,096 tokens per response—roughly 3,000 words. Claude’s limit varies by model but generally sits between 4,096 and 8,192 tokens. Gemini models have similar constraints.
When you hit that wall, the model doesn’t crash or error out. It just stops wherever it happens to be—mid-sentence, mid-code block, mid-paragraph.
Why This Matters More Than You Think
Output limits bite hardest when you’re asking for:
- Long-form content like blog posts, reports, or documentation
- Multi-file code with extensive comments
- Detailed analyses or line-by-line reviews
- Formatted data exports (JSON, CSV, tables)
- Multi-step tutorials or instructions
The frustrating part? The model doesn’t warn you upfront. It starts confidently, writes 80% of what you asked for, then just… stops. You’re left guessing whether it’s done or cut off.
This is especially tricky with code. If the model stops mid-function, you might not notice until you try to run it and everything breaks.
How to Work Around It
The simplest fix: type “continue” or “keep going” in your next message. Most models will pick up approximately where they left off. This works because the partial response is now part of the conversation context, so the model can see what it already wrote and continue from there.
But “continue” isn’t perfect. Sometimes the model restarts the section instead of continuing it, or it loses the thread and changes direction slightly. For critical work, better strategies include:
Break requests into chunks upfront. Instead of “write a 5,000-word guide,” ask for “write the introduction and first two sections.” Then request the next sections separately. You stay in control of the structure.
Ask for outlines first. Get the model to generate a detailed outline or structure, then request one section at a time. This prevents mid-section cutoffs and makes the final output easier to stitch together.
Use explicit section markers. Tell the model to use clear headings or comments to mark sections. If it cuts off, you’ll know exactly where to resume and can prompt specifically: “continue from Section 4.”
Export structured data in smaller batches. Requesting 500 rows of generated CSV data? Ask for 100 at a time. The model is less likely to hit limits, and you can validate each batch before continuing.
Model-Specific Quirks
Claude tends to be more conservative about output length and will sometimes wrap up early even if it hasn’t hit the hard limit—especially if it senses it’s approaching the cap. You can nudge it to keep going by saying “please continue with the full response” upfront.
ChatGPT with GPT-4 is more predictable: it writes until it hits the token ceiling, then stops. The “continue” command works reliably, but expect to use it multiple times for very long outputs.
Gemini models, particularly the 1.5 Pro, have higher output limits than most competitors, making them a solid choice if you routinely need 5,000+ word responses in a single go.
When It’s Not a Limit Issue
Sometimes the model stops for other reasons: it thinks it’s finished the task, it’s uncertain how to continue, or it’s avoiding generating content that might violate its guidelines. If “continue” produces a vague apology or restarts instead of resuming, you’ve probably hit a different issue—not the output cap.
In those cases, rephrase your request or break it down differently. The model might need clearer direction, not just more tokens.
Understanding output limits won’t make them disappear, but it will save you from the frustration of wondering why your carefully crafted prompt keeps getting half-finished responses. Plan for the cap, structure requests accordingly, and you’ll spend less time wrestling with invisible walls.
Want one practical AI insight like this every day? Subscribe to the One Two Three AI newsletter—no fluff, just actionable tips to help you actually use AI better.