Skip to main content
Model Inference supports models from Anthropic and OpenAI, all hosted within your data plane. This reference lists available models and links to their official documentation.

​
Available models

Different Claude generations use different structured-output paths under the hood. Sonnet 5.0 and Opus 4.7 / 4.8 / 5.0 use the forced-tool-use path (schema conformance is enforced by client-side validation and these models do not accept temperature or top_p). Sonnet 5.5 and Opus 5.5 use the prompted-tool-use path — the reserved answer tool is offered with an auto tool choice, and the platform asks for the answer in the prompt rather than forcing the call. All other models use native structured output. See Structured Output for the differences.

​
Anthropic Claude models

Claude models excel at analysis, summarization, and following detailed instructions.

​
Claude Haiku 4.5

Model ID: anthropic.claude-haiku-4.5 The fastest Claude model, optimized for high-throughput tasks where speed matters more than deep analysis. Best for:
  • Simple classification tasks
  • Data extraction
  • Quick summaries
  • High-volume processing
Official documentation: Claude Haiku

​
Claude Sonnet 4.5

Model ID: anthropic.claude-sonnet-4.5 A balanced model offering strong performance across most tasks with good efficiency. Best for:
  • Content analysis
  • Code generation
  • Multi-step reasoning
  • General-purpose tasks
Official documentation: Claude Sonnet

​
Claude Sonnet 4.6

Model ID: anthropic.claude-sonnet-4.6 A Sonnet generation with improved reasoning and instruction following. Best for:
  • Content analysis
  • Code generation
  • Multi-step reasoning
  • General-purpose tasks
Official documentation: Claude Sonnet

​
Claude Sonnet 5.0

Model ID: anthropic.claude-sonnet-5.0 A Sonnet 5 generation model. Uses the forced-tool-use structured-output path — the platform enforces output_format_schema conformance with client-side validation instead of native grammar constraints. This model does not accept temperature or top_p; those sampling parameters are silently dropped from your request. Best for:
  • Content analysis
  • Code generation
  • Multi-step reasoning where the newest Sonnet capabilities are preferred
Official documentation: Claude Sonnet

​
Claude Sonnet 5.5

Model ID: anthropic.claude-sonnet-5.5 The latest Sonnet generation. Sonnet 5.5 uses the same settings as Opus 5.5. It uses the prompted-tool-use structured-output path: the platform offers the reserved answer tool with an auto tool choice and asks for the answer in the prompt, rather than forcing the call. Schema conformance is enforced by client-side validation. The model does not accept temperature or top_p (both are silently dropped), and it always thinks. A request for thinking-off is sent as low effort, the closest setting the model takes. In agent conversations, this model rejects a forced tool choice. A run that passes tool_choice.any or tool_choice.specific_tool returns ForcedToolChoiceNotSupported; use tool_choice.auto and name the tool in the prompt instead. Currently available on Bedrock data planes only; Snowflake Cortex support will follow. Best for:
  • Content analysis
  • Code generation
  • Multi-step reasoning where the newest Sonnet capabilities are preferred
Official documentation: Claude Sonnet

​
Claude Opus 4.5

Model ID: anthropic.claude-opus-4.5 A highly capable Claude model for complex tasks requiring deep analysis and nuanced understanding. Best for:
  • Complex reasoning chains
  • Nuanced analysis
  • Research tasks
  • High-stakes decisions
Official documentation: Claude Opus

​
Claude Opus 4.6

Model ID: anthropic.claude-opus-4.6 An Opus generation with strong capabilities for demanding inference workloads. Best for:
  • Complex reasoning chains
  • Nuanced analysis
  • Research tasks
  • High-stakes decisions
Official documentation: Claude Opus

​
Claude Opus 4.7

Model ID: anthropic.claude-opus-4.7 A newer Opus generation for the most demanding inference workloads. Uses the forced-tool-use structured-output path — the platform enforces output_format_schema conformance with client-side validation. This model does not accept temperature or top_p; those sampling parameters are silently dropped from your request. Best for:
  • Complex reasoning chains
  • Nuanced analysis
  • Research tasks
Official documentation: Claude Opus

​
Claude Opus 4.8

Model ID: anthropic.claude-opus-4.8 The latest Opus generation, the most capable Claude model available. Uses the forced-tool-use structured-output path with client-side schema validation, and does not accept temperature or top_p. Best for:
  • Complex reasoning chains where the latest capabilities are preferred
  • Nuanced analysis
  • Research tasks
  • High-stakes decisions
Official documentation: Claude Opus

​
Claude Opus 5.0

Model ID: anthropic.claude-opus-5.0 The newest Opus generation. Opus 5.0 supports extended thinking — pass an inference_config.thinking setting to have the model reason before answering. Opus models often withhold the reasoning text and return only a provider signature; either way the reasoning block is returned on the assistant turn and can be replayed verbatim on a continuation. Uses the forced-tool-use structured-output path with client-side schema validation, and does not accept temperature or top_p. Currently available on Bedrock data planes only; Snowflake Cortex support will follow. Best for:
  • Complex reasoning chains where the latest Opus capabilities are preferred
  • Nuanced analysis
  • Research tasks
  • High-stakes decisions
Official documentation: Claude Opus

​
Claude Opus 5.5

Model ID: anthropic.claude-opus-5.5 The latest Opus generation. Opus 5.5 uses the prompted-tool-use structured-output path — the platform offers the reserved answer tool with an auto tool choice and asks for the answer in the prompt, rather than forcing the call. Schema conformance is enforced by client-side validation. The model does not accept temperature or top_p (both are silently dropped), and it always thinks — a request for thinking-off is sent as low effort, the closest setting the model takes. In agent conversations, this model rejects a forced tool choice. A run that passes tool_choice.any or tool_choice.specific_tool returns ForcedToolChoiceNotSupported; use tool_choice.auto and name the tool in the prompt instead. Currently available on Bedrock data planes only; Snowflake Cortex support will follow. Best for:
  • Complex reasoning chains where the latest Opus capabilities are preferred
  • Nuanced analysis
  • Research tasks
  • High-stakes decisions
Official documentation: Claude Opus

​
OpenAI models

OpenAI models offer strong reasoning capabilities and broad knowledge.

​
GPT-4.1

Model ID: openai.gpt-4.1 Advanced reasoning model with strong analytical capabilities. Best for:
  • Complex analysis
  • Technical reasoning
  • Multi-domain tasks
Official documentation: GPT-4 Models

​
o4-mini

Model ID: openai.o4-mini Smaller, faster model optimized for efficiency. Best for:
  • Quick responses
  • Simple tasks
  • Cost-sensitive applications
Official documentation: OpenAI Models

​
GPT-oss-120b

Model ID: openai.gpt-oss-120b Open-source large language model with broad capabilities. Best for:
  • General-purpose tasks
  • Open-source requirements

​
Extended thinking

Some models can reason before answering. Enable it per-request by setting inference_config.thinking on your Model Inference call. The field takes one of three shapes:
  • "disabled" — the bare string. Explicit off, worth sending to a model that thinks by default so you are not billed for reasoning you did not ask for.
  • { "type": "effort", "level": "low" | "medium" | "high" } — for models that take an effort scale.
  • { "type": "budget", "budget_tokens": 4096 } — for models that take a token budget. Floor is 1024 (the provider’s), and the budget must be smaller than max_tokens because thinking is spent from the same output budget as the answer.
Omit thinking to get the model’s own default behavior. A setting the target model does not accept returns 400 rather than being silently downgraded — support is per-model and per-data-plane, so the same setting can be accepted on one data plane and refused on another. Reasoning comes back on the assistant turn as reasoning content blocks (working as words plus a provider signature, or a signature alone when the model withholds the text), or as redacted_reasoning blocks when the provider encrypts the reasoning. Both shapes can be replayed verbatim on a follow-up turn — send them back unchanged, or omit them if you only need to display the answer.

​
Image input

Claude models on AWS data planes can take images as input. OpenAI models don’t accept images, and neither does any model on a Snowflake data plane. To send an image, add an image content block to a user message, or to the content of a tool_result block. The block uses the same shape as the Anthropic Messages API image block, with base64 as the only source:
The API returns a 400 that names the block and the problem when a request:
  • Puts an image in a system or assistant message.
  • Uses a media_type other than image/jpeg, image/png, image/gif, or image/webp.
  • Sends data that isn’t valid base64.
  • Includes an image larger than 512 KB (524,288 bytes) once decoded.
  • Includes more than 20 images.
  • Targets a model that doesn’t accept images on the data plane’s platform.

​
Choosing a model

For detailed guidance on model selection, see Choosing the Right Model.

​
Usage example

​
Related content

Choosing Models

Detailed model selection guide

Running Inference

Submit inference requests

Model Inference API

Complete API reference

Model Inference Overview

How inference works