Models and effort

Windlass runs on two DeepSeek models. Which one you use, and how hard it thinks, affects both the quality of the result and how quickly you consume your weekly allowance.

Available models

DeepSeek V4 Flash is the default on every plan. It is quick, inexpensive in allowance terms and more than capable for the bulk of day-to-day tasks: editing files, writing scripts, running commands, summarizing documents, answering questions about your code.

DeepSeek V4 Pro is the stronger model. It reasons more deeply, holds complex plans together over longer tasks, and is noticeably better at multi-step agentic coding, where it has to explore a codebase, make coordinated changes across files and verify them. It is available on the Pro and Max plans.

Choosing a model

Start with Flash. Switch to Pro when:

• The task spans many files or steps, and earlier attempts with Flash lost track of the plan.
• You need careful reasoning: subtle bugs, architecture decisions, tricky data transformations.
• Correctness matters more than speed, for example in a migration or a security-sensitive change.

Switch back to Flash for routine follow-ups. Many people plan and implement the hard part with Pro, then use Flash to tidy up, write documentation and handle small edits.

Reasoning effort

Each model can run at one of four reasoning effort levels, chosen from the model menu in the message box:

• Off. No extended thinking. Fastest and cheapest; fine for simple edits and lookups.
• Low. A short think before acting. Good for routine tasks.
• High. The default. Enough reasoning to plan multi-step work and check its own results.
• Max. The model thinks as long as it needs. Reserve it for the hardest problems; it produces the most reasoning tokens.

Use Max for the hardest part of a task and return to High afterwards. High effort with Flash is often a cost-effective middle ground before reaching for Pro.

Context window and compaction

Sessions support a context window of up to 1M tokens, so the agent can keep a large codebase, long documents and a lengthy conversation in view at once.

As a session grows, Windlass compacts it automatically: the earlier conversation is condensed into a summary that preserves decisions, file references and open questions, and the agent carries on from there. You can trigger compaction yourself with /compact whenever you reach a natural checkpoint.

Compaction keeps long sessions workable, but a fresh session is still the cleanest start for a new topic. The agent is at its best when the context contains only what is relevant to the task.

How models and effort affect your allowance

Your weekly allowance is consumed by the tokens the agent processes. Two things change the rate:

• Model. DeepSeek V4 Pro consumes allowance roughly four times faster than Flash for the same work.
• Effort. Higher effort levels produce more reasoning tokens, so they cost more.

Combined, Pro at High effort is the most powerful setting and the most expensive. Reserve it for work that deserves it. For the full picture, see Usage and limits.

Recommendations

• Default to Flash at High. It covers most tasks and stretches your allowance furthest.
• Escalate deliberately. Move up to High effort, then to Pro, when a task proves hard, rather than starting at the top.
• Split the work. Do the difficult design or debugging in a Pro session, then continue routine implementation in a Flash session.
• Attach only relevant files. Every file in context is processed by the model, whichever one you use.
• Compact at checkpoints. A tidy context helps both models reason well and keeps token counts down.