# Supported agents and models

> The coding agents and models Firegoat runs evals on, and how to pick one per prompt or run.

Every run pairs a **harness** with a **model**. The harness is the coding agent that does the work, such as Claude Code or Codex: it reads the prompt, runs commands, edits files and calls your product. The model is the LLM inside it. Your product is used through the harness, so the same prompt can go differently on another harness with the same model, or on another model in the same harness.

## Agents and model ids

Pass a harness id and a model id together. Any other pair is rejected with a `400` that lists the models offered for that harness.

Firegoat offers 10 agents. A run that nothing pins runs on Claude Code with Sonnet 5.5, marked *(default)* below.

**Claude Code**. Harness: `claude-code`

- Opus 5.5: `claude-opus-5-5`
- Opus 5: `claude-opus-5`
- Opus 4.8: `claude-opus-4-8`
- Opus 4.6: `claude-opus-4-6`
- Sonnet 4.6: `claude-sonnet-4-6`
- Sonnet 5.5: `claude-sonnet-5-5` *(default)*
- Sonnet 5: `claude-sonnet-5`
- Haiku 5.5: `claude-haiku-5-5`
- Haiku 4.5: `claude-haiku-4-5-20251001`
- Fable 5.2: `claude-fable-5-2`
- Fable 5.1: `claude-fable-5-1`
- Fable 5: `claude-fable-5`

**Codex**. Harness: `codex`

- GPT-5.5: `openai/gpt-5.5`
- GPT-5.6: `openai/gpt-5.6`
- GPT-5.6 Sol: `openai/gpt-5.6-sol`
- GPT-5.6 Terra: `openai/gpt-5.6-terra`
- GPT-5.6 Luna: `openai/gpt-5.6-luna`
- GPT-6.2 Astra: `openai/gpt-6.2-astra`
- GPT-6 Astra: `openai/gpt-6-astra`
- GPT-6.2 Sol: `openai/gpt-6.2-sol`
- GPT-6.1 Sol: `openai/gpt-6.1-sol`
- GPT-6.1 Terra: `openai/gpt-6.1-terra`
- GPT-6 Luna: `openai/gpt-6-luna`

**OpenCode**. Harness: `opencode`

- GPT-5.5: `openai/gpt-5.5`
- GPT-5.6: `openai/gpt-5.6`
- GPT-5.6 Sol: `openai/gpt-5.6-sol`
- GPT-5.6 Terra: `openai/gpt-5.6-terra`
- GPT-5.6 Luna: `openai/gpt-5.6-luna`
- GPT-6.2 Astra: `openai/gpt-6.2-astra`
- GPT-6 Astra: `openai/gpt-6-astra`
- GPT-6.2 Sol: `openai/gpt-6.2-sol`
- GPT-6.1 Sol: `openai/gpt-6.1-sol`
- GPT-6.1 Terra: `openai/gpt-6.1-terra`
- GPT-6 Luna: `openai/gpt-6-luna`
- Opus 5.5: `anthropic/claude-opus-5-5`
- Opus 5: `anthropic/claude-opus-5`
- Opus 4.8: `anthropic/claude-opus-4-8`
- Opus 4.6: `anthropic/claude-opus-4-6`
- Sonnet 4.6: `anthropic/claude-sonnet-4-6`
- Sonnet 5.5: `anthropic/claude-sonnet-5-5`
- Sonnet 5: `anthropic/claude-sonnet-5`
- Haiku 5.5: `anthropic/claude-haiku-5-5`
- Haiku 4.5: `anthropic/claude-haiku-4-5-20251001`
- Fable 5.2: `anthropic/claude-fable-5-2`
- Fable 5.1: `anthropic/claude-fable-5-1`
- Fable 5: `anthropic/claude-fable-5`
- Grok 4.3: `xai/grok-4.3`
- Grok 4.5: `xai/grok-4.5`
- Grok 4.6: `xai/grok-4.6`
- Grok 4.7: `xai/grok-4.7`

**OpenCode v2**. Harness: `opencode-v2`

- GPT-5.5: `openai/gpt-5.5`
- GPT-5.6: `openai/gpt-5.6`
- GPT-5.6 Sol: `openai/gpt-5.6-sol`
- GPT-5.6 Terra: `openai/gpt-5.6-terra`
- GPT-5.6 Luna: `openai/gpt-5.6-luna`
- GPT-6.2 Astra: `openai/gpt-6.2-astra`
- GPT-6 Astra: `openai/gpt-6-astra`
- GPT-6.2 Sol: `openai/gpt-6.2-sol`
- GPT-6.1 Sol: `openai/gpt-6.1-sol`
- GPT-6.1 Terra: `openai/gpt-6.1-terra`
- GPT-6 Luna: `openai/gpt-6-luna`
- Opus 5.5: `anthropic/claude-opus-5-5`
- Opus 5: `anthropic/claude-opus-5`
- Opus 4.8: `anthropic/claude-opus-4-8`
- Opus 4.6: `anthropic/claude-opus-4-6`
- Sonnet 4.6: `anthropic/claude-sonnet-4-6`
- Sonnet 5.5: `anthropic/claude-sonnet-5-5`
- Sonnet 5: `anthropic/claude-sonnet-5`
- Haiku 5.5: `anthropic/claude-haiku-5-5`
- Haiku 4.5: `anthropic/claude-haiku-4-5-20251001`
- Fable 5.2: `anthropic/claude-fable-5-2`
- Fable 5.1: `anthropic/claude-fable-5-1`
- Fable 5: `anthropic/claude-fable-5`
- Grok 4.3: `xai/grok-4.3`
- Grok 4.5: `xai/grok-4.5`
- Grok 4.6: `xai/grok-4.6`
- Grok 4.7: `xai/grok-4.7`

**OpenClaw**. Harness: `openclaw`

- GPT-5.5: `openai/gpt-5.5`
- GPT-5.6: `openai/gpt-5.6`

**Cursor CLI**. Harness: `cursor-cli`

- Auto: `cursor/auto`

**fx**. Harness: `fx`

- Sonnet 4.5: `vercel_ai_gateway/anthropic/claude-sonnet-4.5`
- Opus 4.6: `vercel_ai_gateway/anthropic/claude-opus-4.6`

**v0**. Harness: `v0`

- v0 Mini: `v0/v0-mini`
- v0 Pro: `v0/v0-pro`
- v0 Max: `v0/v0-max`

**Grok Build**. Harness: `grok-build`

- Grok Build 0.1: `grok-build-0.1`

**Mastra Code**. Harness: `mastracode`

- DeepSeek V4.1 Flash: `mastra/deepseek/deepseek-v4.1-flash`
- GLM 5.3 Flash: `mastra/z-ai/glm-5.3-flash`
- Gemini 3.8 Flash: `mastra/google/gemini-3.8-flash`
- Kimi K3: `mastra/moonshotai/kimi-k3`
- Qwen3.8 Flash: `mastra/qwen/qwen3.8-flash`

The same list is in the `harnessModel` field description of the [OpenAPI spec](https://firegoat.ai/evals/api/openapi), as `harness: model, model, ...` pairs.

A prompt pinned to `v0` must use the `blank` scaffold. Firegoat refuses to save it with any other.

<Note>
This page is about cloud runs. The [Firegoat CLI](https://firegoat.ai/docs/firegoat-cli) runs evals on your own machine with your own coding agent, chosen by `harness` in `firegoat.yaml`.
</Note>

## The browser agent

When the coding agent stops and asks for something (a sign-up, an OAuth screen, an API key), a browser agent answers as the simulated user. It follows the prompt's persona instructions (see [Prompts](https://firegoat.ai/docs/concepts#prompts)).

Only Claude Code runs with the browser agent. On every other harness, a question to the user goes unanswered. Give the agent what it needs up front, as [environment variables](https://firegoat.ai/docs/concepts#environment-variables-and-logged-in-clis), or compare harnesses only on tasks that need no answer from a user.

## Choosing the agent and model

A run uses the first of these that is set:

1. The harness and model passed when the run is requested.
2. The prompt's own pin.
3. Its parent prompt's pin, for a scenario or competitor lane.
4. Your org default, under **Scheduled evals** in [Settings](https://firegoat.ai/go/settings).
5. The platform default, marked in the table above.

A pin is an object with both halves, for example `{ "harness": "claude-code", "model": "claude-opus-5-5" }`. Changing a pin or the org default affects future runs only. Past runs keep the model they ran on.

| Where | Per prompt | Per run |
| --- | --- | --- |
| Dashboard | On the [Prompts](https://firegoat.ai/go/prompts) page: **Harness / model** on the **Run setup** step of **+ New Prompt**, or **Model** in the prompt's settings (the button next to **Request eval**). | The caret on **Request eval**. Tick one combo, or several to request one eval on each. **Default (queue decides)** uses the order above. |
| [REST API](https://firegoat.ai/docs/api) | `harnessModel` on `POST /api/v1/prompts` and `PATCH /api/v1/prompts/{promptId}/config`. `null` clears the pin. | `harnessModel` on [`POST /api/v1/prompts/{promptId}/run`](https://firegoat.ai/docs/api#triggering-a-run). |
| [MCP server](https://firegoat.ai/docs/mcp) | `harnessModel` on `create_prompt` and `update_prompt_config`. | `harnessModel` on `request_run`. |
| [evals CLI](https://firegoat.ai/docs/evals-cli) | `--harness-model.harness` and `--harness-model.model` on `evals prompts create` and `evals prompts update-config`. | The same flags on `evals runs request`. |

```bash
curl -X POST \
  -H "Authorization: Bearer $EVALS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{ "harnessModel": { "harness": "claude-code", "model": "claude-opus-5-5" } }' \
  https://firegoat.ai/evals/api/v1/prompts/<promptId>/run
```

## Sandboxes

Each run gets a fresh sandbox. Unless you choose otherwise, it is a Linux sandbox on E2B.

On a paid plan you can pick the sandbox backend: per run with `backend` (`e2b` or `daytona`) on `POST /api/v1/prompts/{promptId}/run` or `request_run`, or as your org default under **Scheduled evals** in [Settings](https://firegoat.ai/go/settings). On other plans a request that names a backend is refused.

| Backend | Sandbox |
| --- | --- |
| E2B | Linux. The default. |
| Daytona | Linux |

macOS sandboxes are available on request. [Contact us](https://firegoat.ai/request-access) to enable them for your org.

## Comparing agents and models

To compare, request the same prompt on several combos: tick them under **Request eval** on the prompt's page in [Prompts](https://firegoat.ai/go/prompts), or send one run request per combo. Each run records the harness and model it ran on. The [Runs](https://firegoat.ai/go/runs) page shows them, and `GET /api/v1/runs/{runId}` and the `get_run` and `list_runs` MCP tools return them as `harness` and `model`.

- **Compare like with like.** Time, cost and token counts shift with the harness, and Firegoat does not normalize them. When two runs used different combos, say so before reading a difference as a win.
- **Use several runs per combo.** One run is an anecdote. Set **Runs per eval** on the prompt, or request it more than once.
- **Codex counts MCP calls as `exec`.** Codex can call MCP tools from inside its code-execution tool, so those calls appear in tool counts as `exec`, not under the MCP tool's name. Check the `exec` inputs in the trace before concluding it never used your MCP server.

To compare your product with another product on the same task, add a [competitor](https://firegoat.ai/docs/concepts#competitors) and read the [Benchmark](https://firegoat.ai/go/benchmark) page.
