What you will be able to do
- Tell hybrid reasoning models apart from reasoning-only models on Databricks Foundation Model APIs, and say which reasoning modes each one can actually run
- Map standard, extended and reduced thinking to the parameter that controls them: budget_tokens, reasoning_effort or output_config.effort
- Predict what a model does when the reasoning parameter is left out, and why that default is itself a choice that affects tokens
Key concept
Reasoning depth as a token-and-latency trade — Every reasoning setting trades accuracy on hard tasks against extra latency and tokens. Picking a reasoning mode for an agent means deciding how many of those tokens the task justifies, within the limits of the model you have.
1.Three exam labels, one underlying dial
The exam describes three reasoning modes: standard, extended thinking and reduced thinking. Only one of these labels comes from the Databricks documentation, and that one is "extended thinking". The docs describe a hybrid Claude model as offering "near-instant responses" or "extended thinking". Other models expose a reasoning_effort level that runs from none or minimal up to high, xhigh or max. The docs never use the term "reduced thinking". It is simply a useful name for picking a lower effort level than the model's default.
In this lesson, standard means a fast reply with no reasoning step, or as little as the model allows. Extended means you give the model room to think: a large budget_tokens, or effort set to high, xhigh or max. Reduced means you keep reasoning switched on but lower it, for example to low or minimal. The same trade sits behind all three, as the documentation says for model after model.
Checkpoint 1 of 4· Check yourself
A team raises an agent step from a low to a high reasoning effort. Which outcome does the documentation describe?
Raising effort buys accuracy on complex tasks and costs latency and tokens. That is the single trade behind standard, extended and reduced thinking.
“Higher reasoning effort can improve accuracy on complex tasks at the cost of increased latency and token usage.”Source: docs.databricks.com
2.Hybrid vs reasoning-only: which modes a model can run
Before you can pick a mode, you need to know which modes the model offers. Databricks sorts reasoning models into two types: "There are two types of models, reasoning-only and hybrid." Hybrid models, such as the Claude Sonnet 4.x and Opus 4.x/5 family, Gemini 3 and Gemini 2.5, are described as models that "support both fast, instant replies and deeper reasoning when needed." On these models, all three modes are available.
Reasoning-only models are different. Claude Opus 5.5 "uses adaptive thinking for every request, and reasoning cannot be disabled." Claude Fable 5.1 works the same way. The GPT-5 family and GPT OSS models always reason too. Some of them accept none anyway, and what that value does varies. For GPT OSS, "none" is treated as unset because GPT OSS models always reason. For Inkling, "none" maps to the lowest reasoning effort instead of disabling reasoning. On these models, standard mode doesn't exist. The cheapest option is reduced thinking.
| Model (examples) | Type | Control parameter | Standard (no reasoning) possible? |
|---|---|---|---|
| databricks-claude-sonnet-4-5, databricks-claude-opus-4-8 | Hybrid reasoning | thinking budget_tokens | Yes: fast, instant replies |
| Gemini 3 models (e.g. databricks-gemini-3-1-pro) | Hybrid reasoning | reasoning_effort (minimal, low, medium, high) | Yes: "minimal" returns no reasoning tokens |
| databricks-claude-opus-5-5, databricks-claude-fable-5-1 | Reasoning only | thinking type adaptive + output_config.effort | No: none and disabled are rejected |
| databricks-gpt-oss-120b, databricks-gpt-oss-20b | Reasoning only | reasoning_effort (low, medium, high) | No: "none" is treated as unset |
| databricks-gpt-5-1 | Reasoning only | reasoning_effort | Effort defaults to none |
Checkpoint 2 of 4· Match them up
Match each model to how it handles reasoning.
Tap a term, then the definition that fits it.
The model type decides which modes exist. Sonnet 4.5 and Gemini 3 can skip reasoning, while Opus 5.5 and GPT OSS always reason.
“Use "minimal" to skip reasoning for the fastest, lowest-latency responses.”Source: docs.databricks.com
Sources2
3.Setting the mode, and what happens if you don't
There are three different controls. Hybrid Claude models take a thinking object with budget_tokens. Most other models take reasoning_effort. Claude Opus 5.5 and Fable 5.1 use the Anthropic Messages API, with thinking.type set to adaptive and the depth set in output_config.effort. The documented request for Opus 5.5 looks like this:
Checkpoint 3 of 4· Fill the gap
This Claude Opus 5.5 request asks for high effort. Which value must thinking.type have?
"thinking": {
"type": " ? "
},
"output_config": {
"effort": "high"
},Opus 5.5 accepts only adaptive thinking. You control depth through output_config.effort, not through enabled thinking with a budget_tokens value.
Source: docs.databricks.comLeaving the parameter out does not get you a neutral "standard" mode. It gets you that model's default, and the defaults are very different. For GPT-5.1 and GPT-5.2, the reasoning_effort parameter is set to none by default. Claude Opus 5.5 defaults to medium. For Fable 5.1, "Databricks does not set a default; Anthropic uses high when you omit this parameter." GLM-5.2, the DeepSeek V4 models and Kimi K3 go to the top: "all of these models default to \"max\" when reasoning_effort is omitted". So the same request without the parameter can produce no reasoning tokens on one model and maximum reasoning on another.
Checkpoint 4 of 4· Check yourself
An agent step sends requests without a reasoning_effort parameter. You switch its endpoint from databricks-gpt-5-1 to databricks-deepseek-v4-pro-0813 and change nothing else. What happens to reasoning?
GPT-5.1 defaults to none, while DeepSeek V4 Pro defaults to max. An unchanged request therefore moves from standard mode to the deepest reasoning setting.
“all of these models default to "max" when reasoning_effort is omitted”Source: docs.databricks.com
Defaults range from none to max depending on the model. If you leave the parameter out, you have let the model choose your token cost and latency, and that choice changes when someone swaps the endpoint.
Sources2
Exam traps
Each one states something that sounds right. Open it to see what is actually true.
1.Any model can be put into a standard, no-thinking mode by disabling reasoning.Why is that wrong?
Reasoning-only models such as Claude Opus 5.5 and Fable 5.1 reject disabled thinking. On those models, reduced effort (low) is the cheapest option.
Covered in Hybrid vs reasoning-only: which modes a model can run
2.Leaving out the reasoning parameter gives you a neutral standard mode.Why is that wrong?
Defaults depend on the model. GPT-5.1 defaults to none, Opus 5.5 to medium, Fable 5.1 to high (set by Anthropic), and several open models to max.
Sources
Every claim above is drawn from one of these pages, quoted as it was written on the date shown.
- 1.https://docs.databricks.com/aws/en/machine-learning/foundation-model-apis/supported-modelsOfficial docs
“It offers two modes: near-instant responses and extended thinking for deeper reasoning based on the complexity of the task.”
↩︎ Three exam labels, one underlying dial - 2.
“Higher reasoning effort can improve accuracy on complex tasks at the cost of increased latency and token usage.”
↩︎ Three exam labels, one underlying dial“There are two types of models, reasoning-only and hybrid.”
↩︎ Hybrid vs reasoning-only: which modes a model can run“These models support both fast, instant replies and deeper reasoning when needed.”
↩︎ Hybrid vs reasoning-only: which modes a model can run“"none" is treated as unset because GPT OSS models always reason.”
↩︎ Hybrid vs reasoning-only: which modes a model can run“"none" maps to the lowest reasoning effort instead of disabling reasoning.”
↩︎ Hybrid vs reasoning-only: which modes a model can run“Databricks does not set a default; Anthropic uses high when you omit this parameter.”
↩︎ Setting the mode, and what happens if you don't“Databricks sets the default to medium when you omit this parameter.”
↩︎ Setting the mode, and what happens if you don't“Higher reasoning effort can improve accuracy on complex tasks at the cost of increased latency and token usage.”
↩︎ Key concept“This model uses adaptive thinking for every request, and reasoning cannot be disabled.”
↩︎ Exam trap 1“For GPT-5.1 and GPT-5.2, the reasoning_effort parameter is set to none by default, but can be overridden in requests.”
↩︎ Exam trap 2“Disabled thinking and manual budget_tokens values are rejected.”
↩︎ Prediction“Use "minimal" to skip reasoning for the fastest, lowest-latency responses.”
↩︎ Checkpoint“all of these models default to "max" when reasoning_effort is omitted”
↩︎ Checkpoint