Price the 200K-token cliff before your agent bill doubles

You have a repo, a long document, and a few evenings. The agent looks cheap until one request crosses a token threshold. For a near-zero budget, that threshold is the number to price first.

The 200K-token cliff is the price

For prompts of 200,000 tokens or fewer, Grok 4.6 costs $2 per million input tokens, $0.50 per million cached input tokens, and $6 per million output tokens. For requests of 200,000 tokens or more, Grok 4.6 charges $4.00 per million input tokens, $1.00 per million cached input tokens, and $12.00 per million output tokens, with that elevated pricing covering every token in the request.

Treat that boundary as a fixed limit in the budget. A weekend build can stay near zero if the request is priced before it crosses.

Build a token budget before the agent runs

Start with the request you actually expect to send, not the model card. A whole-repo agent can pull in system instructions, file contents, tool results, and prior turns. Each one adds to the prompt.

Keep a running total as you design the agent. If the total is close to the threshold, stop and cut before you send. The estimate does not need to be perfect; it needs to be good enough to show whether the request is far from the limit or close to it. The cleanest cuts remove repeated context, not the parts that make the agent less useful.

• Estimate total prompt tokens by adding the system prompt, repo files, docs, and conversation history, then compare the number to the threshold.
• Price the whole request as if the higher rate has already happened; if the prompt can grow during the task, assume the worst case and set a dollar figure before the first request.
• Cache repeated context by putting stable instructions, file maps, and long documents in the cached portion, where the same context can be sent again at the cached input rate.
• Cut turns and input tokens by asking for smaller changes, fewer tool calls, and shorter summaries, so each request carries a smaller prompt.
• Check coding benchmarks before you rely on it. Grok 4.6 scored 61 on the Artificial Analysis Intelligence Index, tying GPT-5.6 Sol. On Artificial Analysis's agent task set, Grok 4.6 used about 53 turns and 0.5 billion input tokens, while Claude Opus 5 used about 103 turns and 2.0 billion tokens for the same work. Grok 4.6 achieved 26% on Terminal-Bench v3.0 and was last among the four models in xAI's comparison table. xAI did not release a SWE-bench Pro score for Grok 4.6.

Make each turn earn its cost

Turns are where weekend budgets grow. Each new turn can carry the same instructions, the same file list, and the same prior answers. If the agent keeps asking for more context, the prompt grows even when the answer is short.

Design the task so the agent can finish in fewer moves. Give it a clear stopping condition, such as a finished file, a passing check, or a short summary of what changed. Ask for a plan first, then a small change, then a check. That keeps the prompt small while the agent explores.

Output tokens matter, but a short answer can cost more than a long one if the prompt behind it is huge. The budget starts with the request, not the response.

Rehearse the prompt before the full task

Before you point the agent at the whole repo, send a small rehearsal. Include the same instructions and a representative sample of the files. Watch the prompt size and the number of turns. A rehearsal is cheap enough when both stay inside the budget you set. If the rehearsal feels expensive, the real task will be worse.

Keep the rehearsal until the prompt stays under the threshold. If it crosses, trim the context first. Remove files the agent does not need, shorten the instructions, and ask for a narrower answer. A near-zero budget is built by refusing to send context you do not need.

If the rehearsal stays under the threshold, the full task has a budget you can trust.