Argon 1M output tokens vs Sume Format 120,000-character schema cap
A 1M-token output limit does not lift Sume's output_schema limits. The 120,000 characters bound the schema, and the fallback projection reads 8,000 characters.
No: a model that can write 1 million output tokens does not let a Sume Format run return a bigger structured object. Google says Gemini 4 Argon has an output limit of 1 million tokens, up from 64,000 (read 2026-10-04). Sume's 120,000-character limit is not an output cap at all. It bounds the output_schema document you send, summed over every property name, key and string value, and the other schema limits are separate.
The two numbers measure different things, and mixing them up leads to schemas that are either too timid or rejected at submit.
What do Sume's limits apply to?
Sume's Structured output page lists four size limits. Each applies to the schema you bind to a run, and a violation is a 400 output_schema_invalid before anything runs, so nothing is charged.
| Limit | Value | Violation code |
|---|---|---|
| Nesting depth | 10 levels | max_depth |
| Total properties | 5,000 across the whole document | max_properties |
| Enum values | 1,000 per enum | max_enum_values |
| Total string length | 120,000 characters over names, keys and string values | max_string_length |
Where does a long model reply go in a Format run?
A Format run's result is mostly media: durable files on media.sume.com listed in artifacts[]. The orchestrator's closing text is part of output.text. When the run does not submit your object itself, Sume builds one afterwards from the generated media and the run's closing text, truncated to its first 8,000 characters. That fallback is the filled_by: "projection" path.
So a very long closing message is not a way to carry more data. If a field must be in output, require only what the Format actually makes, and keep long source material in input, which is written whole to a file in the run's workspace and is never truncated.
How should I design a schema for a long-output model?
Keep the schema small and the artifacts large.
- Put media in
SumeMediaFile#fields rather than inlining long text. - Make optional fields a type union with
null, since an empty projection field reads asnull. - Keep long
descriptionannotations short; they count against the 120,000-character budget. - Send the long brief in
input; the first ~4,000 characters ofinstructionare carried as prompt text.
Does this change if the orchestrator changes?
The model field on a run selects the orchestrator only, per Create a run. The schema subset, the projection rule and the limits above are properties of the Format API, so they hold whichever model plans the run.
Sources
Related posts
More in Formats
- Argon targets long-horizon work; a Sume Format run ends at 90 minutes
Google positions Gemini 4 Argon for long tasks. A Sume Format run is force-finalized as failed 90 minutes after creation, so split long jobs into runs.
- GPT-6.1 Sol errs less, but still pass prices as Format input
OpenAI says GPT-6.1 Sol cut factual errors from 11.4% to 7.7%. For ad video, give the run your prices and claims as input data instead of trusting recall.
- A gift portrait for your store homepage: model product portrait Format
The sume-model-product-portrait Format returns one finished still of a person with your product. Use it for a homepage hero, not for a motion ad.
- November 1 switch: restyle Halloween product stills into a fall set
After Halloween, run your stills through the sume-restyle Format in one bulk queue of up to 100 items and keep the product while changing the mood.
Written by Sume