Supported LLM Models

Set a model per agent, or per request with the model field. You can use a pinned model id or a latest-of-a-tier alias.

Default model

When a request doesn’t name a model, HiveGate uses DEFAULT_CHAT_MODEL, and falls back to gemini-3-flash-preview if that isn’t set. The value must be a model id or alias from this page. An unknown value stops the server at startup, instead of failing every chat that omits model.

# .env
DEFAULT_CHAT_MODEL=anthropic:sonnet-latest

Latest-of-a-tier aliases

An alias follows the vendor’s newest model in a tier without a code change, and never moves to a pricier tier. HiveGate resolves an alias by listing the vendor’s models and taking the newest one in the tier. It caches the answer for 24 hours and logs it (Model alias anthropic:sonnet-latest -> ...). If the vendor can’t be reached, it uses the last known answer, or the fallback below, and retries after 5 minutes.

Alias	Fallback
`anthropic:haiku-latest`	`claude-haiku-4-5`
`anthropic:sonnet-latest`	`claude-sonnet-5`
`anthropic:opus-latest`	`claude-opus-5`
`openai:luna-latest`	`gpt-6-luna`
`openai:terra-latest`	`gpt-5.6-terra`
`openai:sol-latest`	`gpt-6-sol`
`openai:astra-latest`	`gpt-6-astra`
`google:flash-lite-latest`	`gemini-3.5-flash-lite`
`google:flash-latest`	`gemini-3.8-flash`
`google:pro-latest`	`gemini-3.1-pro-preview`
`xai:grok-latest`	`grok-4.7`

Prefer aliases for defaults, and pinned ids when you need a model that never changes.

Pinned models

OpenAI: GPT-5.4

Model ID	Enum member
`gpt-5.4`	`gpt_5_4`
`gpt-5.4-mini`	`gpt_5_4_mini`
`gpt-5.4-nano`	`gpt_5_4_nano`

OpenAI: GPT-5.6

Model ID	Enum member
`gpt-5.6-luna`	`gpt_5_6_luna`
`gpt-5.6-terra`	`gpt_5_6_terra`
`gpt-5.6-sol`	`gpt_5_6_sol`

OpenAI: GPT-6 (current flagship)

Model ID	Enum member
`gpt-6-luna`	`gpt_6_luna`
`gpt-6-sol`	`gpt_6_sol`
`gpt-6-astra`	`gpt_6_astra`

GPT-6 runs on the OpenAI Responses API. There is no gpt-6-terra, so openai:terra-latest resolves to gpt-5.6-terra.

Google Gemini 3.x (stable)

Model ID	Enum member
`gemini-3.8-flash`	`gemini_3_8_flash`
`gemini-3.7-flash`	`gemini_3_7_flash`
`gemini-3.6-flash`	`gemini_3_6_flash`
`gemini-3.5-flash`	`gemini_3_5_flash`
`gemini-3.5-flash-lite`	`gemini_3_5_flash_lite`
`gemini-3.1-flash-lite`	`gemini_3_1_flash_lite`

Google Gemini 2.5 (existing Google users only)

Model ID	Enum member
`gemini-2.5-pro`	`gemini_2_5_pro`
`gemini-2.5-flash`	`gemini_2_5_flash`
`gemini-2.5-flash-lite`	`gemini_2_5_flash_lite`

Google now serves Gemini 2.5 only to existing users: new API keys get 404 NOT_FOUND. Prefer Gemini 3.x or a google:*-latest alias.

Google Gemini (preview)

Model ID	Enum member
`gemini-3.1-pro-preview`	`gemini_3_1_pro`
`gemini-3-flash-preview`	`gemini_3_flash`

Anthropic Claude

Model ID	Enum member
`claude-fable-5-1`	`claude_fable_5_1`
`claude-opus-5-5`	`claude_opus_5_5`
`claude-sonnet-5-5`	`claude_sonnet_5_5`
`claude-haiku-4-5-20251001`	`claude_haiku_4_5`
`claude-haiku-4-5`	`claude_haiku_4_5_undated`
`claude-fable-5`	`claude_fable_5`
`claude-opus-5`	`claude_opus_5`
`claude-opus-4-8`	`claude_opus_4_8`
`claude-opus-4-7`	`claude_opus_4_7`
`claude-opus-4-6`	`claude_opus_4_6`
`claude-opus-4-5`	`claude_opus_4_5`
`claude-sonnet-5`	`claude_sonnet_5`
`claude-sonnet-4-6`	`claude_sonnet_4_6`
`claude-sonnet-4-5`	`claude_sonnet_4_5`

Every Claude id here is a pinned snapshot, including the ones without a date. Only the anthropic:*-latest aliases move. The current lineup is claude-fable-5-1, claude-opus-5-5, claude-sonnet-5-5 and claude-haiku-4-5-20251001; the rest are older models that are still served.

xAI Grok

Model ID	Enum member
`grok-4.7`	`grok_4_7`
`grok-4.6`	`grok_4_6`
`grok-4.5`	`grok_4_5`
`grok-4.3`	`grok_4_3`
`grok-4.20-0309-reasoning`	`grok_4_20_reasoning`
`grok-4.20-0309-non-reasoning`	`grok_4_20_non_reasoning`

Z.ai GLM

Model ID	Enum member
`glm-5.3`	`glm_5_3`
`glm-5.3-flashx`	`glm_5_3_flashx`
`glm-5.3-flash`	`glm_5_3_flash`
`glm-5.2`	`glm_5_2`

Served through Z.ai’s OpenAI-compatible endpoint. Z.ai documents no endpoint for listing models, so there is no -latest alias. Only models with documented function calling are listed, because every HiveGate agent has tools.

DeepSeek

Model ID	Enum member
`deepseek-flash`	`deepseek_flash`
`deepseek-v4-pro`	`deepseek_v4_pro`

deepseek-flash is a moving name (currently V4.1-Flash).

Provider keys

Provider	Model ids start with	API key variable
OpenAI	`gpt-`, `openai:`	`OPENAI_API_KEY`
Google Gemini	`gemini-`, `google:`	`GOOGLE_API_KEY`
Anthropic	`claude-`, `anthropic:`	`ANTHROPIC_API_KEY`
xAI	`grok-`, `xai:`	`XAI_API_KEY`
Z.ai	`glm-`	`ZAI_API_KEY`
DeepSeek	`deepseek-`	`DEEPSEEK_API_KEY`

Claude refusals

When a Claude safety classifier declines a request, Anthropic returns no content. HiveGate reports this as a run error with a stable message, stop_reason=refusal, category=..., model=..., which reaches streaming clients as an error event. The message doesn’t include any request content. Previously a refusal showed up as an empty reply.