Two Price Resets in One Week: How to Re-Evaluate Your AI Model Costs

Between September 22 and 28, 2026, the two leading model providers reset their pricing: GPT-6 Sol and Luna at half the cost of GPT-5.6, Claude Opus 5.5 about 40% cheaper to run than Opus 5, and Sonnet 5.5 claiming up to 30% less per task at unchanged prices. Any AI feature budgeted before this is running on stale assumptions.

Measure Cost per Task, Not Price per Token

List prices are only half the story. A model that solves a task in fewer tokens, needs fewer retries or uses cached context can be cheaper overall even at a higher token price. The only number that matters is the total cost of a successfully completed task.

A Simple Re-Evaluation Process

• Collect 50–200 real examples from production, with the outputs you consider correct.
• Run them through your current model and two or three candidates, including a cheaper tier.
• Score quality the same way for each — automated checks where possible, human review for the rest.
• Record tokens, retries, latency and cost for each run.
• Pick the cheapest model that meets your quality bar, per use case — not one model for everything.

Route by Difficulty

The spread inside one model family can be twentyfold. Classification, extraction and short summaries often work on the smallest tier; complex reasoning and long documents justify the largest. Routing each request to the right tier usually saves more than any single price cut.

Don't Forget Caching

If your prompts reuse long instructions or documents, caching discounts can dominate the bill. Opus 5.5's cache reads dropped 60% versus Opus 5. Structure prompts so the reusable part comes first and stays identical between calls.

Make It Routine

With releases arriving monthly, keep your evaluation set and scripts ready so re-testing takes an afternoon, not a project. Abstract your model calls behind one internal interface so switching is a configuration change, not a rewrite.