When an LLM API returns a response, it includes a finish_reason field telling you why generation stopped. The common values are:
- **'stop'**: the model produced its end-of-sequence (EOS) token, or it matched one of the stop sequences the caller supplied. This is a clean natural ending.
- **'length'**: the runtime hit the max_tokens limit before any natural stop signal fired. The output was truncated; the model would have kept generating if allowed.
- **'content_filter'** or similar: a safety / moderation layer halted the response.
- **'tool_calls'** (or 'function_call'): the model emitted a structured tool / function-call output and the runtime stopped to let the caller handle it.
So 'length' specifically means: the cap ran out before the model decided to stop. It is the runtime's way of warning you that the output is incomplete. The fix is one of three things: raise max_tokens for that request, prompt the model to be more concise so it finishes within the cap, or make a continuation call (resubmit the prompt plus the truncated output as the assistant prefix) and let generation resume.
The option-by-option mapping: A describes EOS, which is 'stop'. B describes a stop sequence, also 'stop'. C describes hitting max_tokens, which is 'length' (correct). D describes a rate limit, which surfaces as an HTTP error code (typically 429), not as a finish_reason value.
**finish_reason = 'length' means the runtime hit max_tokens before the model emitted EOS or matched any caller-supplied stop sequence.** It is a hard numeric truncation, not a natural completion.
The full enum across major 2026 providers (OpenAI, Anthropic, Google, others) shares a similar shape:
- 'stop' (Anthropic uses 'end_turn' for EOS-driven and 'stop_sequence' for stop-sequence-driven): natural completion via EOS or a stop string.
- 'length' (Anthropic uses 'max_tokens'): truncation because max_tokens ran out.
- 'content_filter' or 'refusal': a moderation or safety policy intervened.
- 'tool_use' or 'tool_calls': the model emitted a tool / function call and the runtime stopped to dispatch it.
A 'length' termination is operationally significant for three reasons. First, the output is incomplete: any downstream consumer that assumes complete responses (JSON parsers, structured-output validators, tool-call dispatchers) will either error out on a malformed payload or, worse, silently accept a truncated one. Second, the model's hidden state at the truncation point is well-defined: the response is exactly the first max_tokens of what the model would have produced. So a continuation call (resubmit the original prompt plus the truncated output as an assistant prefix) will resume generation deterministically from that exact point, modulo sampling randomness.
Third, a rising 'length' rate is diagnostic: either the cap is too tight for the task (raise max_tokens or break the task into chunks), or the model is not naturally terminating where you expect (a missing or misconfigured stop sequence, or the model is hallucinating continuations past the natural boundary). Production stacks monitor finish_reason distribution as a first-class observability signal.
The other options are wrong because they describe different termination causes: A is EOS (-> 'stop'), B is stop-sequence match (-> 'stop'), and D is a rate-limit, which is a transport-level error surfaced as
Quick reference
TL;DR
finish_reason = 'length' means the runtime hit the max_tokens cap before the model emitted EOS or matched a stop sequence. It is a hard numeric truncation, not a content-driven stop.
Imagine asking someone to write you a story but giving them only 30 seconds. If they finish naturally before the timer runs out, they put a period at the end and hand it to you ('stop'). If the timer rings while they are still writing, they have to stop mid-sentence ('length'). The 'length' signal means the timer ran out, not that the story was finished. To get a complete answer you have to either give them more time, ask for a shorter story, or come back and ask them to keep writing where they left off.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example.
Workloads where the response is expected to be exactly N tokens (e.g. fixed-length completions, embedding extraction) If you intentionally cap output length, 'length' is the expected termination and not a failure; treat it accordingly.
Tasks where a continuation call would not preserve coherence (single-shot reasoning, structured outputs requiring full context) Continuation calls have edge cases; for some tasks the right answer is to raise max_tokens upfront rather than resume from a truncation.
Streaming consumers that already display partial output as it arrives Streaming surfaces partial content directly, so the user sees the truncation in real time; the finish_reason then only confirms the cause for logging.
Quick scan · 2–4 min
Everything important, quickly.
2 min: classify the three termination causes (content / length / policy), map each option to its enum, identify C as the max_tokens-cap case (= 'length'), and name the production handling (continuation calls, observability, structured-output gating).
Side-by-side
Cause
OpenAI
Anthropic
Google Gemini
EOS sampled (natural completion)
'stop'
'end_turn'
'STOP'
Stop sequence matched
'stop'
'stop_sequence'
'STOP'
max_tokens cap hit
'length'
'max_tokens'
'MAX_TOKENS'
Safety / content policy
'content_filter'
'refusal' or stop_reason variants
'SAFETY'
Tool / function call emitted
'tool_calls' / 'function_call'
'tool_use'
'TOOL_CALL'
Real-world examples
Real products, models, and research that use this idea.
1OpenAI Chat Completions returns finish_reason = 'length' when max_tokens (or max_completion_tokens for o-series) is hit before EOS.
2Anthropic's Messages API uses stop_reason = 'max_tokens' for the same condition; the field name differs but the semantics match.
3Google Gemini returns finishReason = 'MAX_TOKENS' for truncation by cap, with the partial output in the candidates field.
4Production agent frameworks (LangGraph, LlamaIndex agents) check finish_reason before parsing tool calls; a 'length' termination skips JSON parse and triggers a continuation call.
5Reasoning-model APIs (o-series, Claude extended thinking) expose separate budgets for reasoning_tokens versus visible output; 'length' can fire on either.
Interviewer follow-ups
What an interviewer would ask next. Try answering before peeking at the approach.
QHow do continuation calls work mechanically, and what are the edge cases?
Sign in or go Pro to see the approach for this follow-up.
QWhy do reasoning models often hit 'length' even on short visible answers?
Sign in or go Pro to see the approach for this follow-up.
QWhy is rising 'length' rate diagnostic of stop-sequence misconfiguration?
Sign in or go Pro to see the approach for this follow-up.
Don't say this
Red flags and common mistakes that signal junior thinking. Click to expand.
Most common mistake
Reading finish_reason = 'length' as 'the model finished naturally and the answer is long'. It is the opposite: the answer was cut off because the runtime hit a numeric cap. Treating a 'length' termination as a complete response leads to silently truncated outputs.