Re-run Expensive Agent Prompts During Grok 4.7's Discount, Then Downgrade

Your agent already has a set of prompts, and a short Grok 4.7 discount is a good chance to see which ones deserve the bigger model. Pointing every prompt at the bigger model is tempting. Treat the discount as a test bench instead. You have a weekend, a small agent, and a budget that should stay near zero. Most side projects stay side projects, and that is fine. You want a smaller list of expensive prompts, not a permanent upgrade.

The discount is a short test window

Check the discount window before you change anything. Grok 4.7 arrived on September 21, 2026. Vercel's AI Gateway ran a 40% discount through September 27, making the effective rate $1.20 per million input tokens and $3.60 per million output tokens. Stop when you have a date, a rate, and a short list of prompts you can re-run. The window lasts a few days, not a migration project.

The capability gap matters. Grok 4.7 contains 2.1 trillion parameters, 40% more than Grok 4.6's 1.5 trillion. On CursorBench 4.0, a coding benchmark, it reached 46.3%, while Grok 4.6 reached 40.4%. Use that gap to mark the prompts that need it.

A four-item checklist keeps the weekend cheap

Before you re-run anything, sort each prompt your agent already uses into a short table you can review in one sitting.

• Classify each prompt by context length, output length, cache behavior, and capability need.
• Run the expensive ones during the promo.
• Set a default cheap tier.
• Flag anything that crosses the 200,000-token boundary, where billed rates become $4 per million input tokens, $1 per million cached input tokens, and $12 per million output tokens.

Cache behavior matters because repeated input can change the bill even when the model stays the same. If your agent sends the same system prompt, the same tool list, or the same file block on every call, note it in the cache column. Short prompts with short answers and repeated instructions probably belong on the cheap tier. Prompts that need long files, multi-step debugging, or careful planning should be marked for the discount run.

The 500,000-token context window in Grok 4.7 was unchanged from Grok 4.6. A prompt that fits in the window can still change price as it grows.

A narrow promo run keeps the bill small

Re-run only the prompts you marked. Keep the same inputs, the same system instructions, and the same success check you would use in production. Save the output, the token count, and the time it took. The promo table lists prompt name, model, tokens, cost, and whether the answer was good enough.

Do not turn on maximum effort for everything. Grok 4.7 at the xHigh effort setting used about 81,000 output tokens per Intelligence Index benchmark task, compared with 36,000 for Grok 4.6 High and 27,000 for GPT-6 Astra Max, so a lower per-token price does not automatically mean a lower cost for a completed task. Task-cost estimates put Grok 4.7 xHigh at about $3.74 per task and Grok 4.7 High at about $2.73. GPT-5.6 Sol Max cost about $1.99 per task, even though its list price was $4 per million input tokens and $20 per million output tokens.

For a side project, ask whether the prompt needs that extra capability often enough to justify the bill. If the answer is no, move that prompt back to the cheap tier.

The cheap default should survive the promo

Set a default cheap model for most prompts before the promo ends. Keep the expensive model only for the prompts that clearly failed or produced weak answers on the cheap tier. The discount should end with a smaller exception list and a default that handles the boring work.

Grok 4.7 launched at a base API price of $2 per million input tokens and $6 per million output tokens, the same base rates as Grok 4.6. The discount is a temporary nudge, not a permanent price cut. Afterward, your tier list already tells you which prompts stay on the bigger model and which return to the cheaper one.