continuation token
An engineering solution for extremely long generation with large language models. When a single generation reaches the segment limit, a special continuation credential is returned, allowing the client to issue the next request and seamlessly continue writing. This mechanism lets output be assembled across multiple requests while preserving coherence, and it is used to work around the maximum output length imposed on a single model response.
Timeline (last 90 days)
Through the continuation token returned by the Google Interactions API, the client can carry this token and a higher token budget in the next round of requests, and the model can continue generating verbatim from the breakpoint.
By default, Argon does not run to the full 1 million tokens, and in tests it often stops at 32K or 65K tokens; through the interactions API, when the limit is reached the server returns a continuation token, and the client can carry that token plus a higher budget to continue generation from the breakpoint, or enable background mode to let the server automatically chain 65K segments together.
Gemini 4 Argon supports automatically chaining together multiple segments of about 65,000 tokens each via the continuation token mechanism of the Google Interactions API, achieving continuous output of up to 1 million tokens.
The continuation token mechanism makes "a million tokens in a single response" engineering-feasible by splitting one ultra-long generation into multiple coherent segments, but it requires the model to maintain contextual consistency across segments.
All Facts (4)
Through the continuation token returned by the Google Interactions API, the client can carry this token and a higher token budget in the next round of requests, and the model can continue generating verbatim from the breakpoint.
50%UnverifiedBy default, Argon does not run to the full 1 million tokens, and in tests it often stops at 32K or 65K tokens; through the interactions API, when the limit is reached the server returns a continuation token, and the client can carry that token plus a higher budget to continue generation from the breakpoint, or enable background mode to let the server automatically chain 65K segments together.
50%UnverifiedGemini 4 Argon supports automatically chaining together multiple segments of about 65,000 tokens each via the continuation token mechanism of the Google Interactions API, achieving continuous output of up to 1 million tokens.
50%UnverifiedThe continuation token mechanism makes "a million tokens in a single response" engineering-feasible by splitting one ultra-long generation into multiple coherent segments, but it requires the model to maintain contextual consistency across segments.
50%