Models
Nexrall is multi-provider. Every agent surface — Apps, Docs, Nexrall Code, the Nexrall CLI, and Nexrall Work — lets you pick which underlying LLM powers the agent, across several providers.
The catalogue
Prices are the vendor's published list rates in USD per 1M tokens (input / output), before Nexrall's markup — the same figures the model pickers show next to each model, editable per model in the admin console.
| Provider | Model | Context | Max output | Price in/out (per 1M) |
|---|---|---|---|---|
| DeepSeek | DeepSeek Flash | 1.05M | 128K | $0.3 / $1.2 |
| Anthropic | Claude Sonnet 5.5 | 1M | 128K | $2 / $10 |
| Anthropic | Claude Opus 5.5 | 1M | 128K | $4 / $20 |
| Anthropic | Claude Fable 5.1 | 1M | 128K | $10 / $50 |
| OpenAI | GPT-6 Astra | 922K | 128K | $10 / $50 |
| OpenAI | GPT-6 Sol | 922K | 128K | $2 / $10 |
| OpenAI | GPT-6 Luna | 922K | 128K | $0.1 / $0.5 |
| OpenAI | GPT-5.6 Sol | 922K | 128K | $4 / $20 |
| OpenAI | GPT-5.6 Terra | 922K | 128K | $2 / $12 |
| OpenAI | GPT-5.6 Luna | 922K | 128K | $0.2 / $1.2 |
| OpenAI | GPT-5.4 | 922K | 128K | $5 / $22.5 |
| OpenAI | GPT-5.4 Mini | 272K | 128K | $0.75 / $4.5 |
| OpenAI | GPT-4.1 | 1.05M | 32K | $2 / $8 |
| Z.AI | GLM 5.3 | 1.05M | 131K | $1.4 / $4.4 |
| QwenCloud | Qwen3.7 Max | 992K | 65K | $2.5 / $7.5 |
Not every model is offered everywhere — the three surfaces have their own catalogues:
| Surface | Offered models |
|---|---|
| Apps (agent coding builds) | Claude Sonnet 5.5, Claude Opus 5.5, Claude Fable 5.1, GPT-5.4, GPT-5.4 Mini, GPT-4.1, DeepSeek Flash |
| Docs (document agent) | Claude Sonnet 5.5, Claude Opus 5.5, Claude Fable 5.1, GPT-5.4, GPT-5.4 Mini, GPT-4.1, DeepSeek Flash |
| Nexrall Code / CLI / Work | everything in the table except Qwen3.7 Max, GPT-5.6 Sol, GPT-5.6 Luna, GPT-5.4, GPT-5.4 Mini and GPT-4.1 |
A few notes on the edges of the catalogue:
- Qwen3.7 Max is temporarily not offered. It is still catalogued so a session that already picked it still resolves, but it cannot be selected while the provider's account is not in good standing.
- GPT-4o Mini is catalogued but not offered — its 128K window is too small for this kind of agent loop, which routinely carries a whole file tree. It stays in the catalogue so an old session id still resolves to a real model.
- Claude Haiku 4.5 (200K window) is not in the picker either — a config saved before the rename (the old
fastvalue) still resolves to it; it is not offered for new selection. - New personal accounts start on DeepSeek Flash. Picking any other model (and the effort control) unlocks with your first payment — or automatically if you belong to a team, where the team pool funds the work. Every picker (the Apps and Docs composers, Nexrall Code in VS Code, and the CLI) shows an Upgrade badge in place of the model selector until then; the badge opens a preview of the line-up a first payment unlocks (Claude, GPT-6, GLM and more) with a button that takes you straight to Add funds. If a saved session or config still names a locked model, the turn runs on DeepSeek Flash and says so.
- GPT-5.6 Sol, GPT-5.6 Luna, GPT-5.4, GPT-5.4 Mini and GPT-4.1 are no longer offered in the developer tools (Nexrall Code, the CLI, Nexrall Work) — the newer OpenAI generations supersede them there, and all five were removed from that picker on 2026-09-28. They stay in this catalogue and keep working for any session already running on one; the GPT-5.4 generation is also still offered by the Apps/Docs builder.
- GPT-6 Astra is OpenAI's most capable model and leads this catalogue at the top of the OpenAI price range — reach for it on the hardest end-to-end work. It is also the only OpenAI model that reasons and uses tools at the same time, which is why a long thinking phase can stream before it acts.
Choosing a model
- The default depends on your account. A never-paid account (and not in a team) starts on DeepSeek Flash — a strong coding model at a fraction of the cost of a frontier model. Once unlocked — your first payment, or team access — sessions start on Claude Sonnet 5.5 instead. Naming a model yourself always wins over either default.
- Reach for Claude Opus 5.5 or Claude Fable 5.1 when a task needs deeper reasoning — unusual architecture decisions, tricky debugging, or a plan spanning many steps.
- Pick the GPT-6 family when you prefer OpenAI models — Astra for the hardest work, Sol for frontier reasoning, Luna for the cheapest high-volume work — or GPT-5.6 Terra as a balanced middle option. The older GPT-5.4 / GPT-4.1 generation remains available in the Apps/Docs builder.
- GLM 5.3 is a mid-priced middle option (0.6×) with always-on reasoning.
- GPT-6 Luna and the legacy Haiku alias are the cheap OpenAI / Anthropic choices for lightweight steps.
Model choice is per session, so a long-running task can be started on a frontier model and continued on a cheaper one — or the other way round.
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 — click the model name in the chat composer (next to the
/button): the picker lists models by provider with their per-1M prices, plus the thinking-effort slider. - VS Code — click the model badge in the composer footer for the same model + effort popover.
- CLI —
/modelinside a session, or--model <id>on the command line.
Older model names
Model selection is by name. Some values from older configurations still resolve, so a saved session, a settings file or a sub-agent's model: written before the rename keeps working — the words turbo / pro / ultra / fast (and the old -t / -p / -u flags, since removed and refused with the replacement named) map to claude-sonnet-5-5 / claude-opus-5-5 / claude-fable-5-1 / claude-haiku-4-5-20251001. Nothing offers them anymore; prefer the real model id (--model claude-opus-5-5) so the model running is always unambiguous.
These names were naming shorthand only — never plans or tiers. There is nothing to subscribe to, and no model is gated behind one. Every offered model is available on any account.
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.