ClawsGO Docs

Choose a model

Three ClawsGO 1.0 tiers, plus frontier models from Anthropic and OpenAI. A reference multiplier is shown next to every model.

Switch models in the "Model" submenu of the composer settings. Every row ends with a reference multiplier: roughly what the model costs on a typical task relative to ClawsGO 1.0 Lithium (x1.0). Actual charges are metered per token at each model's own input, output and cache rates, see Credits.

ClawsGO 1.0

ClawsGO 1.0 is our own model line, post-trained for the ClawsGO Science Harness. The three tiers trade depth against cost:

TierPositioningRef. multiplierAvailable from
LithiumLight tier, lowest latency and lowest consumptionx1.0Free
SiliconWorkhorse tier, balanced capability and consumptionx2.0Plus
TitaniumFlagship tier, performance firstx3.0Pro

New accounts start on Lithium.

Third-party models

The menu also carries Anthropic and OpenAI models. The ClawsGO Science Harness stays the same; only the underlying model changes:

ModelPositioningRef. multiplierAvailable from
Claude Sonnet 5Workhorse, strong at coding and agentic tasksx2.0Plus
Claude Opus 5Flagship, deep reasoning and complex tasksx5.0Pro
GPT 5.6 TerraMid-weight, balanced general capabilityx1.0Plus
GPT 5.6 SolFlagship, strong at maths and complex reasoningx2.5Pro
GPT 6 AstraNew-generation flagship for hard reasoning and researchx5.0Pro
Claude Fable 5.1Anthropic's strongest model, for the hardest research problemsx14.0Team and Enterprise only

"More → View all models" at the bottom of the menu opens a searchable dialog grouped by vendor, including previous-generation models. Switching to the Claude Fable line asks for confirmation: it uses several times the credits of Claude Opus 5.

Models outside your plan are greyed out; clicking one opens the subscription section. Each plan's model range is listed in Plans.

Which to pick

  • Everyday questions, formatting, small code changes: Lithium.
  • Multi-paper reviews, data analysis, drafting: Silicon or Claude Sonnet 5.
  • Long, complex research that needs repeated self-checking: Titanium, Claude Opus 5 or GPT 6 Astra.
  • The hardest problems: Claude Fable 5.1.

Switching mid-conversation

Switching on the new-task page has no side effects. Switching inside an existing conversation asks for confirmation: it breaks the context cache, so the first message after the switch costs significantly more, then costs return to normal. The switch applies to later turns only; the turn already running is unaffected.

If a model declines your content, switching to another model and replying "Continue" is the intended way out.

Context window

The "Context" submenu has two settings:

  • 256K: the standard window, right for most tasks. Earlier content is compacted automatically as the limit approaches.
  • 1M: keeps more detail across very long papers, datasets and codebases. There is no feature fee, but each turn feeds the model far more, so credit consumption rises noticeably. Plus and above.

Every new task starts at 256K. A conversation on 1M shows "1M" on the settings button.

Reasoning

Five levels: Low, Medium, High, Extra High, Max. The default is High. Cluster mode always runs at Extra High and cannot be changed.