Keep your Copilot project near zero with three model switches

A weekend project can stay cheap if the model picker is treated like a thermostat: long agentic jobs go to the capable slot, routine jobs go to the cheaper slot, and long prompts get cut before the 200,000-token cliff.

Most side projects do not need the most expensive model all night. They need a repeatable rule for when to spend, when to save, and when to stop.

Grok 4.6 is now available as a selectable model in GitHub Copilot

Its long-task economics make that rule easy to follow.

Three switches keep a weekend bill small

Keep this list next to your editor. Each item is a decision you can make in a minute, before the task starts or before the bill grows.

• Long agentic tasks go to the cheap frontier slot. Use the stronger model for many turns, many files, or sustained judgment.
• High-volume edits go to the cheaper model. Use the lower-priced model for renames, formatting, small diffs, and repetitive changes across many files.
• Prompts split or stop before 200,000 tokens. Check the token estimate before sending, and summarize or split the work if it is close.

The list works because the model picker changes the bill, not only the answer. Start with the long task, because it is the one that can run all night.

Long tasks earn the frontier slot

SpaceXAI launched Grok 4.6 on Wednesday, and it earned a 61-point Artificial Analysis Intelligence Index score, the same score as GPT-5.6 Sol and one point lower than Claude Fable 5

The SpaceXAI console prices Grok 4.6 at $2 per million input tokens and $6 per million output tokens

Claude Opus 5 is listed at $5 per million input tokens and $25 per million output tokens, while GPT-5.6 Sol is listed at $5 per million input tokens and $30 per million output tokens

The cheaper price is only half the reason. Grok 4.6 used about 53 turns and 0.5 billion input tokens on average for knowledge-work tasks, compared with 103 turns and 2.0 billion input tokens for Claude Opus 5, so its real per-task cost is lower than a per-token comparison implies

For a weekend builder, a cheap frontier coding model can keep a long migration, a multi-file refactor, or a test-suite repair from becoming a nightly expense.

If the task is a one-off hard problem, the frontier slot is the right place. If the task is a batch of similar changes, move it to the volume lane before it starts.

Volume belongs to the cheaper model

Not every edit needs the strongest model. Grok 4.3 sits at $1.25 per million input tokens and $2.50 per million output tokens

Grok 4.3 scored 38 on the Artificial Analysis Intelligence Index, 23 points below Grok 4.6

That gap matters when the task is judgment-heavy, but it shrinks when the task is mechanical. If you are renaming variables, updating imports, or generating many small patches, the cheaper model can do the work while the stronger model waits for the hard part.

Model pickers also move. xAI ended the original Grok 4 on May 15, 2026, and traffic to that model now goes to Grok 4.3

Check what the picker actually sends, and keep the volume lane on the cheaper option.

The 200K cliff is where the bill jumps

The third switch is the one people forget until the invoice arrives. Once a prompt reaches 200,000 tokens, Grok API rates double

Cached input tokens for Grok 4.6 are billed at $0.50 per million

That discount helps when the same context is reused, but it does not remove the cliff. A long prompt that keeps growing can turn a routine session into a high-rate session.

Before sending a large prompt, look at the token estimate. If it is close to the line, stop and split the task before the session starts. Summarize the earlier context, move the next step into a fresh prompt, or break the work into smaller chunks.