Models
Nexrall is multi-provider. Every agent surface — Fullstack Agent Coding, Nexrall Canvas, Nexrall Code, the Nexrall CLI, and Nexrall Work — lets you pick which underlying LLM powers the agent, across several providers.
The catalogue
| Provider | Model | Context window | Relative cost |
|---|---|---|---|
| Anthropic | Claude Sonnet 5 (default) | 1M tokens | 1.5× |
| Anthropic | Claude Opus 5 | 1M tokens | 3× |
| Anthropic | Claude Fable 5 | 1M tokens | 5× |
| Anthropic | Claude Haiku 4.5 | 200K tokens | 0.1× |
| OpenAI | GPT-5.4 | 922K tokens | 2× |
| OpenAI | GPT-5.4 Mini | 272K tokens | 0.5× |
| OpenAI | GPT-4.1 | 1M tokens | 1× |
| DeepSeek | DeepSeek V4 Pro | 1M tokens | 0.1× |
| DeepSeek | DeepSeek V4 Flash | 1M tokens | 0.05× |
| Qwen | Qwen3.7 Max | 1M tokens | 1× |
Relative cost is a signal only, not a live price feed — it's priced per 1M tokens and compared against GPT-4.1 / Qwen3.7 Max (= 1×).
Choosing a model
- Start with Claude Sonnet 5, the default — fast and a strong all-round choice for most coding and automation work.
- Reach for Claude Opus 5 or Claude Fable 5 when a task needs deeper reasoning — unusual architecture decisions, tricky debugging, or a plan spanning many steps.
- Pick GPT-5.4 / GPT-4.1 when you prefer OpenAI models for a given task.
- Reach for DeepSeek V4 Pro / V4 Flash or Qwen3.7 Max when you want far lower per-token cost and don't need the very deepest reasoning — good for high-volume, mechanical work.
- Pick Claude Haiku 4.5 for lightweight, high-volume steps (bulk search, simple classification) where neither deep reasoning nor a huge context window is needed — it's the cheapest Anthropic model in the catalogue.
Effort level
Model choice and thinking effort are separate. Effort controls how much reasoning the model does before acting — low, medium (default), high, or extra. The available levels depend on the model's provider (some models expose only a subset).
Where to switch models
- Web app — the model picker in the chat composer.
- VS Code — the model menu in the chat panel header.
- CLI —
/modelinside a session, or--model <id>on the command line.
Legacy tier aliases
The older turbo / pro / ultra / fast shorthand still works where it always did (including in sub-agent and skill frontmatter's model: field), but it's now just an alias for the Claude models — turbo → claude-sonnet-5, pro → claude-opus-5, ultra → claude-fable-5, fast → claude-haiku-4-5-20251001. Prefer a real model id (--model claude-opus-5) so the model running is always unambiguous.
Context window
Every model supports a large context window, which matters for two things:
- Long agent sessions — many turns of back-and-forth without losing earlier context.
- Large codebases — reading enough of a project at once to make a correctly-scoped change instead of a locally-correct-but-globally-wrong one.
When a session's context approaches the limit, Nexrall automatically compacts older history — summarizing it while pinning your original request and a structured progress ledger verbatim, so the agent doesn't drift from the original goal even after several rounds of compaction.