Eighty percent. That’s how far OpenAI cut the price of GPT-5.6 Luna in 2026, and it’s the kind of number that doesn’t just trim a line item on an invoice. It changes the questions you’re willing to ask a model in the first place.
The July 30 announcements covered the GPT-5.6 family broadly. Terra took a 20% reduction. Luna, the entry tier, took the 80% haircut that put it meaningfully below competing options. Sol sits at the other end of the tradeoff, offering faster processing for a higher price. Taken together, the price-performance ratio across the family improved substantially.
I run an SEO practice that leans on language models for the unglamorous middle of the work: classifying intent across thousands of queries, extracting entities from competitor pages, mapping internal link opportunities, auditing metadata at scale. Every one of those jobs is a volume problem. And volume problems live or die on per-token cost.
What an 80% cut actually unlocks
Here is the mental math that governs most AI-assisted SEO work. You have a task. You estimate token cost per unit. You multiply by the number of units. Then you compare that number to what a human would charge, or to what the insight is worth, and you either run it or you don’t.
At the old price, plenty of jobs sat just on the wrong side of that line. A full semantic audit of a 40,000-URL enterprise site was technically possible and financially irritating. So you sampled. You picked 500 representative pages, ran the analysis, and extrapolated. That’s how consultants have always handled cost ceilings, and it’s fine right up until the pages you didn’t sample are the ones with the problem.
Cut the input cost by 80% and sampling stops being the default. You run the whole site. You run it monthly instead of quarterly. You stop treating model calls as a scarce resource to be rationed and start treating them the way you treat a database query, which is to say, casually.
Jobs that move from “maybe” to “obviously”
- Page-level intent classification across an entire domain rather than a sampled subset
- Continuous monitoring of competitor content changes instead of periodic manual checks
- Entity and topic extraction on every URL, feeding a real internal linking graph
- Rewriting and testing multiple metadata variants per page instead of one best guess
- Log file and query data summarization at a cadence that catches problems while they’re small
None of these are new ideas. They were all just slightly too expensive to do properly, which meant most teams did a smaller version and called it done.
Luna and Sol are not competing, they’re a routing decision
The temptation with a tiered family is to pick one model and standardize. That’s the wrong instinct here. Luna and Sol solve different constraints, and the money is in matching the model to the job.
Sol’s advantage is speed at a higher cost. That matters for anything with a human waiting on the other end. Interactive tools, real-time content assistance, anything where a delay breaks the workflow. Latency has a real price, and it’s often paid in a person’s attention rather than in dollars.
Luna’s advantage is cost at volume, which suits batch work where nobody is watching the clock. Overnight audits. Weekly crawls. Bulk classification jobs that need to finish by morning, not by the next keystroke.
Building your pipeline so the routing decision is a config value rather than a rewrite is the practical move. Most teams I talk to have hardcoded a single model name in a dozen places. That was survivable when there was one sensible choice. With a family this spread out on price and speed, it’s now a tax you pay every time pricing shifts, and pricing has clearly started shifting.
The competitive squeeze nobody’s talking about
Cheap inference is a leveling force, and that cuts in a direction some agencies won’t enjoy. A lot of SEO retainers have been quietly priced around the labor of doing analysis that a model can now do for cents. When the underlying cost of that analysis drops by 80%, the client eventually notices.
The defensible work moves upstream and downstream of the model. Upstream: knowing which questions are worth asking, which data to feed in, how to structure the problem. Downstream: judgment about what to act on, what to ignore, and how a change interacts with everything else on the site. The middle layer, the mechanical extraction and summarization, is becoming a commodity input.
That’s not a complaint. It’s the same pattern that hit every tool-assisted profession. The people who adapted fastest got to do more interesting work with better inputs. The ones who priced their value on the mechanical part had a harder few years.
My advice for the next quarter is unglamorous. Audit which of your AI-assisted workflows are still running on sampled data because of old cost assumptions. Rerun them at full scale. Make your model selection configurable. Then look hard at what you’re actually charging for, because the price floor just moved, and the floor tends to pull everything above it down eventually.
đź•’ Published: