OpenAI cut the price of its top model by over 20% mid-quarter and announced it in a three-paragraph forum post, which tells you how routine price drops have become there.
The model is GPT-5.6 Sol, and the new rates hold for three months from August 21. Input now costs $4 per million tokens, down from $5. Output went from $30 per million to $20. That works out to a 20% cut on input and a 33% cut on output.
The discount applies wherever you pay by usage. The API is covered, including Fast mode and long-context requests, and so are Batch and Flex processing. Codex credits and eligible ChatGPT Work plans come along too. Subscriptions didn't change.
If you build or buy AI search tools, file this under supply chain: the raw material those tools run on just got cheaper, and that changes what a vendor can afford to measure.
The numbers and the catch
Of the two cuts, I'd look at output first. An AI search engine reads a compressed context, then spends most of its tokens writing a long answer with citations, so on that work a third off output saves you more than a fifth off input. If you sell answers, summaries or articles at volume, your biggest variable cost just shrank by a third.
Then there are the four words that make it a promotion: "for the next 3 months." The window runs through late November and OpenAI hasn't repriced the model for good, so I'd be wary of any margin that only works at $20 output.
Line up the quarter and the trend is plain. In July the smaller Terra and Luna tiers got cheaper. On August 13 came an Ultrafast preview with a claimed speedup of up to 14x for Sol, and after that the flagship got its timed discount. Prices and speeds keep improving on a steady beat, and this price still expires.
The timing makes sense once you see what shipped next: GPT-6 Astra, two weeks later, on September 3. Discounting the outgoing flagship keeps developers spending on OpenAI during the switch, and it gives OpenAI a lever in its pricing fight with Gemini and Claude.
How often does ChatGPT mention your brand?
Most founders have no idea. The answer might surprise you.
What cheaper tokens do to AI search tools
Underneath, every AI visibility tool is a machine for buying tokens. Citation trackers and brand monitors work this way, as do answer-engine auditors: they send buyer-shaped prompts to AI engines and parse the answers for who got cited. The token price is what one more look at your position costs.
So a cut of 20-33% reaches past vendor margins and into roadmaps. A vendor can now afford to run more prompts for each brand across more engines, and to repeat the same prompt more times. The repeats are the part I'd watch, because churn is what makes these numbers hard to trust right now.
Recent tracking research found citations reshuffling at around 45% per answer regeneration. So one run of your prompts could look different next time. A September study built on 252,407 answers asked outright whether ten prompts are enough to measure a brand reliably. Any answer you can trust takes volume, and volume just got a third cheaper, so my bet is the serious tools spend the savings on statistical confidence.
The same math helps if you've thought about building your own tracker. A thorough weekly run of fifty queries across seven engines, with a few regenerations each, now costs low hundreds of dollars a month in Sol tokens, and less on smaller models. The token bill is also the smallest thing you'd be signing up for.
What you'd really be building is the harness around those calls. Your prompt set has to stay stable, and each engine's quirks need handling. Somebody parses the cited domains and stores the trends. Then the whole thing has to run every week, busy weeks included, and in my experience that's the part that fails first.
We make a tool in this category, so these price moves get our attention too. Cheaper tokens are nice, but the job still comes down to running checks every week, whatever OpenAI charges this quarter. That's also why a three-month window is worth having even though it ends: at the same budget, it pays for three months of denser checking.
The second-order effect: cheaper answers, fewer clicks
If you own a site, tool budgets are the smaller story. Cheap inference powers the whole August shift in search visibility. AI Overviews spread onto more queries and AI Mode got tested as a default, and free ChatGPT access paid for by ads belongs on the same list. Cheaper full answers make each of those easier for Google and OpenAI to sustain, across more of your buyers' questions.
So when fewer searches turn into clicks for you, part of the reason is what an answer costs to produce. Already 68% of Google searches end without a click, and AI Mode sessions come in around 93% zero-click. This price cut just made the economics under both numbers a third better.
I can't see this discount as harmless for organic traffic. When answers cost less to write, engines write more of them, and the visibility you care about keeps moving from links into citations.
There's a quieter version of the same shift. With frontier tokens this cheap, a vertical search tool can build answers into its product. So can a support bot that recommends software, or an agent that draws up vendor shortlists. Each is another place your brand shows up or doesn't. Their number keeps growing, and your analytics will miss nearly all of them.
15 hours a week manually. Or 15 minutes with RankControl.
Track citations, monitor competitors, and fix content gaps across every AI search engine. Automatically.
What to do with a three-month window
The window closes in late November. I'd use it for three things.
Start with measurement if you've been putting it off. Record where your brand gets cited on each engine while checks cost less, using weekly per-engine tracking or a manual pass. The events of August already reshuffled sources once this quarter, and GPT-6 Astra's rollout looks like the next thing likely to shake them up, so a baseline taken now is what you'll compare against.
Shopping for an AI visibility tool? Ask the vendor where the discount went. You want to hear about more regenerations per check, and ideally more prompts and engines too. If the sampling and the price both stayed put, the savings became their margin.
If you run LLM workloads yourself, check your model routing again. Sol calls get the discount without you doing anything, but the July cuts to Terra and Luna mean the cheapest model that still does your job may have changed twice this quarter. An hour of evals could be worth 30% for good.
The discount runs out in late November, and the trend behind it keeps going. Every quarter of 2026 has made intelligence cheaper and AI answers more plentiful, and each one has raised the price of not knowing where your brand stands in those answers. I'd put the savings toward finding out.
AI search traffic grew 835% this year. Is your content ready?
RankControl generates 26 content formats optimized for ChatGPT, Claude, and Perplexity. Published on your domain, matched to your brand.
