Models and Harnesses

Managed Research supports two public harnesses:

codex
opencode_sdk

Model compatibility is harness-specific. Public model IDs are not the same thing as launch entitlement. Backend preflight remains authoritative for the caller’s plan, Beta Access state, provider readiness, and budget limits.

Launch availability

Access	What it allows
Free	`lite`, `directed_effort`, standard models, and one-time Managed Research credits.
Pro / Team	Paid GA model bundle and higher included usage.
Any plan + Beta Access	`open_ended_discovery`, `heavy`, and all-model access.

lite and heavy describe run depth and coordination limits. They do not mean “smaller model” and “bigger model.” Choose the model explicitly when a harness supports model selection.

Codex harness

Model ID	Reasoning effort support	Launch access	Default
`gpt-5.3-codex`	`medium`	Beta Access; worker/engineer roles only	`medium`
`gpt-5.3-codex-spark`	`medium`	Standard	`medium`
`gpt-5.4`	`medium`, `high`	Beta Access	`medium`
`gpt-5.4-mini`	`low`, `medium`, `high`	Standard	`medium`
`gpt-5.4-nano`	`low`	Standard	`low`
`gpt-oss-120b`	`high`	Standard	`high`

Example:

run = client.runs.start(
    "Review the repo and leave evidence for the smallest safe fix.",
    host_kind="daytona",
    work_mode="directed_effort",
    providers=[{"provider": "openrouter"}],
    agent_harness="codex",
    agent_model="gpt-5.4-mini",
    agent_model_params={"reasoning_effort": "medium"},
)

OpenCode SDK harness

Model ID	Reasoning effort support	Launch access
`anthropic/claude-sonnet-4-6`	None in the catalog	Beta Access
`anthropic/claude-haiku-4-5-20251001`	None in the catalog	Standard
`x-ai/grok-4.1-fast`	None in the catalog	Standard
`moonshotai/kimi-k2.6`	None in the catalog	Standard

Example:

run = client.runs.start(
    "Review the repo and propose the smallest high-impact fix.",
    host_kind="daytona",
    work_mode="directed_effort",
    providers=[{"provider": "openrouter"}],
    agent_harness="opencode_sdk",
    agent_model="anthropic/claude-haiku-4-5-20251001",
)

Source of truth

The backend-owned Managed Research model catalog is the source of truth. The synth-ai SDK mirrors public model IDs, but backend preflight is the final authority for launch availability.

Launch Fields Runbooks and Work Modes

⌘I

Getting Started

Managed Research

Stack

Research Factory

Prompt Optimization

Infrastructure

SDK Reference

Research API

Reference

Cookbooks

Models and Harnesses

Launch availability

Codex harness

OpenCode SDK harness

Source of truth

​Launch availability

​Codex harness

​OpenCode SDK harness

​Source of truth

Launch availability

Codex harness

OpenCode SDK harness

Source of truth