After MiniMax released M3.1 Flash on October 19, which field is the answer? How to separate reasoning_content, content, and the JSON string inside a tool call

As of 20 October 2026: the M3.1 Flash answer is in content and the thinking is in reasoning_content. Turning thinking off returns 400. A tool argument is a JSON string that must be parsed again. JSON Schema is still documented only for the older MiniMax-Text-01.

The answer from M3.1 Flash is in content. The thinking is in reasoning_content. Thinking cannot be turned off: thinking.type set to disabled, or reasoning_effort set to none, returns 400. A tool argument is not an object. It is a JSON string that has to be parsed again. On the official text API page, response_format with a JSON Schema is still listed only for the older MiniMax-Text-01. Do not point that switch at the October 19 model.

This piece is written for 20 October 2026, from the launch note that was still on MiniMax’s research page the day before, and from the OpenAI-compatible API page. The model id is MiniMax-M3.1-Flash, with a 1,000,000-token context. We already wrote about why the first Haiku 5.5 block is not the answer. Same habit, different vendor: select by field name, not by position.

What actually shipped on October 19

The launch note says MiniMax M3.1 Flash rolls out that day. Space Bunny, which processed more than 200 trillion tokens while running anonymously, was an early version of this model. Online inference is nearly three times faster, and agent usage cost is 70% lower than the predecessor. Weights come later. Pro comes later. What you can call today is the API, the M Plan, and MiniMax Code.

The same page gives BankerToolBench 76.2, DeepSWE v1.1 at 73.7%, OSWorld V2 at 61.7, and Agents’ Last Exam at 29.6, and it ranks the last two first among Chinese models. CyberGym is 82%. Everyday output can pass 200 tokens per second, with the median staying above 170. Those are capability numbers. A parser has to read the fields in the next section.

ItemLaunch note or API page
Model idMiniMax-M3.1-Flash
Context1,000,000 tokens
Per million tokensinput $0.15, output $0.60, cached input $0.015
Thinking depthlow medium high xhigh max; omitted means max

Where the answer sits on each protocol

The OpenAI-compatible base URL is https://api.minimax.io/v1. The invocation guide says thinking always comes back on reasoning_content, and content holds only the final answer, so you do not peel it out of a <think> tag. On a stream, thinking fragments arrive first on delta.reasoning_content. The answer follows on delta.content.

The same OpenAI SDK page also lists M3.1-Flash among models whose content contains <think> tags that must be preserved. The next paragraph says those tags stay inside content only when reasoning_split is false, and M3.1-Flash does not allow that setting. Read both sentences together. Send the whole assistant message back. When you parse, the answer is content and the thinking is reasoning_content. If a tag really shows up inside content, do not treat it as your business JSON, and do not delete it before the next turn.

ProtocolDepth fieldThinkingAnswer
OpenAI Chat Completionsreasoning_effortreasoning_contentmessage.content
Anthropic Messagesoutput_config.effortblock with type thinkingblock with type text
OpenAI Responsesreasoning.effortoutput item with type reasoningoutput_text, or output_text inside the message item

The Anthropic-compatible base URL is https://api.minimax.io/anthropic, and the path is /v1/messages. Select blocks by type. The first block is often thinking. That is the same class of mistake as Haiku 5.5.

Why turning thinking off returns 400

If you omit thinking, thinking stays on. {"type": "adaptive"} also leaves it on. M3.1-Flash rejects {"type": "disabled"} with HTTP 400 and error code 2013. The API page prints this text:

model "MiniMax-M3.1-Flash" requires adaptive thinking; thinking.type="disabled" (including reasoning.effort=none) is not allowed (2013)

reasoning_effort applies only to M3.1-Flash, and none is not a value. To spend fewer tokens, lower the level to low. Do not try to switch thinking off. The same page splits the other models: MiniMax-M3 skips thinking and answers directly when it receives disabled; M2.x accepts the field and ignores it, so thinking stays on. A request copied from M3 that turns thinking off will 400 as soon as the model id becomes M3.1-Flash.

Tool arguments are a string, not an object

In the OpenAI shape, tool_calls[].function.arguments is a JSON string. The name is function.name. The call id is id. A simple call can skip thinking, so that turn has no reasoning_content. Check that the field exists before you read it. Whether or not thinking is present, the next turn must append the whole assistant message, including reasoning_content and tool_calls. On the Anthropic path, append the full content block list: thinking, text, and tool_use.

Parsing is two steps. The outer value is the chat response. The inner value is the tool argument. Treating arguments as an object that is already parsed yields an empty location. If a later turn sends only content and drops the thinking, interleaved thinking stops. The API page states that as the condition for keeping the reasoning chain continuous.

Empty content, Schema, and the price

The legacy field max_tokens counts thinking tokens toward the limit. Set it too low and finish_reason is length with an empty content. New integrations use max_completion_tokens. The Anthropic-compatible path still calls the limit max_tokens. An empty answer is a length check first, and a field check second.

The text-generation page says response_format set to {"type": "json_schema", "json_schema": {...}} is supported only by MiniMax-Text-01. The M3.1-Flash notes do not offer that forced exit. If you need JSON, ask for it in the prompt and validate content locally. The tool-argument string is JSON the API already handed you. The business answer is not.

The launch note prices the model flat, with no peak and off-peak split. Per million tokens: input $0.15, output $0.60, cached input $0.015. The OpenAI-compatible page also says a priority service_tier costs 1.5 times standard, and that M3.1-Flash does not support the field. Temperature runs from 0 to 2 and defaults to 1. top_p defaults to 0.95 for this model. n accepts only 1.

Keep the fields, not the last sentence

Keep at least these fields from one call, not only the model’s last sentence that says it is done:

{
  "model": "MiniMax-M3.1-Flash",
  "reasoning_effort": "max",
  "finish_reason": "tool_calls",
  "has_reasoning_content": true,
  "content": "I'll check that for you.",
  "tool_name": "get_weather",
  "tool_arguments": "{\"location\": \"San Francisco, US\"}"
}

Leave tool_arguments as a string so the second parse stays visible. Store has_reasoning_content separately so a simple tool turn is not recorded as missing thinking. For the 400, store status 2013 and the thinking.type you sent. Do not store it as a model outage.

  • JSON Validator — validate the outer record, then validate the string inside tool_arguments on its own.
  • JSON Diff — compare two assistant messages and see whether reasoning_content or tool_calls dropped out.
  • JSONPath — pull the responses whose finish_reason is length and whose content is empty.

FAQ

Can I turn thinking off to save tokens?

Not on M3.1-Flash. thinking.type set to disabled, or reasoning_effort set to none, returns 400 with error code 2013. Lower the level to low. MiniMax-M3 does skip thinking when it receives disabled, so the two requests are not interchangeable.

Which field is the answer, content or reasoning_content?

The invocation guide says that on the OpenAI-compatible API the answer is content and the thinking is reasoning_content. On a stream, thinking arrives first on delta.reasoning_content. On the Anthropic-compatible API, select by block type. text is the answer.

Why is finish_reason length while content is empty?

max_tokens counts thinking tokens toward the cap. If the cap is too small, thinking fills it before the answer starts. New integrations should use max_completion_tokens and leave room for the task.

Can json_schema force the output to be JSON?

The text-generation page says response_format with json_schema is supported only by MiniMax-Text-01. M3.1-Flash has no such forced exit. Ask for JSON in the prompt, then validate content. The tool-argument field is already a JSON string.

Do I send reasoning_content back on the next turn?

Yes. A multi-turn tool call must append the whole assistant message, including reasoning_content and tool_calls. A simple call may have no thinking field. If the field is absent, do not read it, and still keep the message intact.

Is function.arguments an object or a string?

In the OpenAI shape it is a JSON string. After you parse the outer response, parse arguments again to get keys such as location. Treating it as an object leaves those keys empty.

Summary

The October 19 model id is MiniMax-M3.1-Flash. Thinking is on by default, and the off switch is rejected. Take the answer by protocol: content on Chat Completions, the text block on Messages, output_text on Responses. Parse tool arguments a second time. The JSON Schema force-exit is still documented on the older model.

Quote scores from the launch note. In code, keep finish_reason, whether the thinking field is present, and the raw arguments string. An empty content is a length check. A 400 is a check for disabled.