Models

This page explains which model IDs a request can name, how the gateway knows what each model accepts, and how to add a model. The live answer to "what can I call" is always the model list:

curl -s http://localhost:8080/api/v1/models -H "Authorization: Bearer $API_KEY" | jq -r '.data[].id'

It lists the foundation models and inference profiles your account can call in the gateway's region, read from Bedrock at start and again every MODEL_CATALOG_REFRESH_SECS, plus the short names below. A model that Bedrock launches while the gateway runs appears there without a restart.

Model IDs

A request names its model with any ID Bedrock accepts:

• a foundation model ID, such as amazon.nova-pro-v1:0, for a model your region serves on demand;
• a cross-region inference profile ID, such as us.anthropic.claude-sonnet-4-5-20250929-v1:0 or global.anthropic.claude-sonnet-5-5, which Bedrock routes across regions. Most recent models are available only this way;
• an application inference profile ARN from your account;
• one of the gateway's short names, such as gpt-5.5, listed below.

The gateway sends the ID to Bedrock as you wrote it, so a profile keeps its routing. The prefixes us., eu., apac., jp., au., ca. and global. all work.

Claude

Claude Sonnet, Opus, Haiku and Fable are served through Bedrock's Converse API, on both Chat Completions and Responses. Recent versions are reached through inference profiles, for example global.anthropic.claude-sonnet-5-5, us.anthropic.claude-opus-5-5 or global.anthropic.claude-fable-5-1.

The gateway handles the differences between versions for you. On the versions that reject temperature and top_p, it drops them. On the versions that think adaptively, reasoning_effort becomes the effort the model expects. When a version does not accept a conversation that ends with an assistant message, the gateway adds a short user turn that asks the model to continue.

response_format and the Responses text.format are honoured on Claude Sonnet 4.5 and 4.6, Haiku 4.5 and Opus 4.5 and 4.6, the versions on which Bedrock supports structured output. On the other versions the gateway answers with HTTP 400 instead of sending a request that Bedrock would refuse.

OpenAI GPT

OpenAI's models on Bedrock are called by short names:

Name	Served through	Chat Completions	Responses
gpt-5.4, gpt-5.5	Bedrock's OpenAI-compatible endpoint	Yes	Yes
gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna	Bedrock's OpenAI-compatible endpoint	Yes	Yes
gpt-6.1-sol, gpt-6-sol, gpt-6-luna, gpt-6-astra	Bedrock Converse, global profile	Yes	Yes
gpt-oss-120b, gpt-oss-20b	Bedrock's OpenAI-compatible endpoint	Yes	No

• GPT-5.x and gpt-oss go to Bedrock's OpenAI-compatible endpoint, which needs a Bedrock API key (AWS_BEARER_TOKEN_BEDROCK) and is available in a few regions only: us-east-1, us-east-2 and us-west-2, depending on the model. In another region the gateway answers with HTTP 400. On Chat Completions, GPT-5.x is served by translating to and from the Responses API, and its reasoning summary appears in <think> tags.
• GPT-6.x goes through Bedrock Converse with the global.openai.* inference profile, so it needs no Bedrock API key and works wherever that profile does. The us.openai.* and global.openai.* IDs work as well. reasoning_effort becomes reasoning.effort, temperature and top_p are dropped because these models reject them, and structured output uses the strict JSON schema the models support.

Other models

Amazon Nova, DeepSeek and every other model in your catalog are served through Bedrock Converse with no extra setup. A model's limits are Bedrock's: when a model does not support images, tools or streaming, Bedrock's error comes back to the client.

Embeddings use their own registry. The API reference lists the supported families.

How the gateway knows a model

What each model accepts lives in a registry, config/models.toml, not in the code. An entry matches a fragment of the model ID and sets, for example:

• whether the model rejects temperature and top_p;
• which form of reasoning it takes;
• the smallest prompt prefix worth caching, and whether it supports a 1-hour cache;
• whether it supports structured output;
• which backend serves it, and in which regions.

A model with no entry is served with the defaults: no reasoning parameters, no structured output, and no prompt caching, except for a Claude ID, which gets a conservative cache threshold.

Add or change a model

1. Copy config/models.toml from the release you run into a directory, for example /etc/bedrock-gateway/config/.

2. Add or change an entry:

   [[model]]
   match = "provider.model-name"
   capabilities = ["drop_sampling_params", "structured_output"]
   [model.params]
   cache_min_tokens = 1024
   reasoning_path = "adaptive_thinking"

3. Start the gateway with CONFIG_DIR pointing at that directory.

When two entries match one ID, the first one in the file supplies the parameters, so put a more specific entry, such as claude-sonnet-5-5, above a shorter one, such as claude-sonnet-5. The capability flags of every matching entry add up. The comments at the top of models.toml describe every flag and parameter.

A short name is an [[alias]] entry, which must come before the first [[model]]:

[[alias]]
from = "my-model"
to = "global.provider.model-name"