GPT‑6 lowers costs, improves caching and expands ChatGPT Ads

GPT‑6 changes the cost-per-task discussion

OpenAI introduced GPT‑6 Sol and GPT‑6 Luna on September 22. In the API, Sol costs $2 per million input tokens and $10 per million output tokens; Luna costs $0.10 and $0.50, respectively. OpenAI describes these prices as 50% below comparable GPT‑5.6 promotional pricing. Availability and limits vary by product and plan.

What it means: Assigning the same model to every task is increasingly hard to justify. A simple response, routine classification and complex analysis have different requirements; paying for maximum capability at every step can hide waste.

Next step: Choose one repeated workflow and record acceptable quality, review time and cost per outcome. Test Luna on bounded tasks and Sol when the work needs more capability. Keep human review for commercial, financial or sensitive decisions, and scale only if the correction rate stays within a defined limit.

Caching can reduce spend when prompts share a structure

OpenAI also announced prompt-caching improvements for GPT‑6. Eligible shared prefixes can be reused within a 30-minute window and receive up to a 90% discount on cached input tokens. The company added metrics and diagnostics to distinguish hits, misses and possible causes.

What it means: Savings depend on the usage pattern, not merely enabling a feature. If common instructions change order or every request uses a different structure, less content will be reused.

Next step: Put stable instructions first and variable data afterward. Measure the cached-token share, total cost and latency before and after. Do not promise a 90% reduction in the total bill: the discount applies to eligible input, and the actual result depends on how much is reused.

ChatGPT Ads expands reach but does not replace attribution

On September 23, OpenAI announced that ChatGPT Ads was expanding to Indonesia, Malaysia, the Philippines, Singapore, Thailand, Vietnam and Taiwan, bringing availability to more than 60 countries. Self-service is limited to eligible businesses; ads appear on Free and Go, while Plus, Pro and Enterprise remain ad-free. OpenAI says ads are labeled and separate from answers.

What it means: The expansion is another sign that ChatGPT is becoming an acquisition channel worth monitoring. It does not mean a company in Mexico should shift budget today or that a campaign will perform the same way across markets.

Next step: Confirm eligibility, target market and available controls before budgeting. If you run a test, use an identifiable offer and landing page, tag visits, and agree on one primary metric—such as qualified opportunities—alongside a spending cap.

A seven-day test that ends with a decision

Select one frequent, low-risk task. Define the acceptable result, run a sample with the current setup, and compare it with a GPT‑6 alternative. Record cost per task, human review time, errors and, where relevant, cache use. End the week with an explicit decision: keep, adjust or stop.

Merluca helps turn AI announcements and new channels into measurable processes. If you need to choose where to automate, control cost or connect a test to actual sales, let’s discuss the workflow you want to improve.