OpenAI no longer makes money by relying on its most expensive models.
If there are still people today spending the most money to call OpenAI's most powerful model.
OpenAI would instead advise them to switch to another one.
On July 30, OpenAI released a price adjustment announcement.
The GPT-5.6Luna model saw an 80% price cut, and the Terra model got a 20% price reduction.
Seeing this news, it's easy to focus on the price war that has kicked off in Silicon Valley.
However, if you carefully go through the officially published technical documents and API call guidelines, you will realize that the truly noteworthy move is not the price figures themselves.
Essentially, this is the first time OpenAI has started telling users that many tasks do not actually require the most powerful model.
The official gave a very specific suggestion: for a complex task, first use GPT-5.6Sol to complete requirement analysis and solution design, then hand it over to Luna for execution, code writing and test running.
The most expensive model is responsible for thinking, and the cheapest one is responsible for execution. This strategy would have been equivalent to commercial self-denial two years ago.
You know, in the past, the entire AI industry was desperately telling the market that their own model was the smartest.
But today, OpenAI is standing up and saying you don't always have to buy the most expensive one.
This move is far more important than the price cuts.
1. The Tacit Understanding Between Two Silicon Valley Giants
Let's first look at what happened in the last two weeks.
On July 30, OpenAI adjusted its prices. Luna got an 80% price cut, and Terra was reduced by 20%. Sol did not see a price cut, but added a Fast mode, which can be up to 2.5 times faster than the standard mode, with the price doubled, while keeping the intelligence level exactly the same.
Top-tier models will be retained as always, while mid-to-low-end models are the ones that truly start to achieve high sales volume.
One week earlier.
On July 24, Anthropic did almost the exact same thing. Claude Opus 5 was released, priced at $5 for million tokens input and $25 for output, which is exactly half of the price of Fable 5.
Compared with performance breakthroughs, Anthropic emphasized its cost-effectiveness more publicly: with only half the price, users can get cutting-edge reasoning capabilities that are infinitely close to Fable 5.
Just one month ago, Fable 5 was still the flagship product that Anthropic focused on promoting, with the strongest reasoning, the longest context and the highest price. A month later, Anthropic personally found a half-price alternative for its flagship.
If only one company did this, it could be interpreted as a product adjustment. When two companies do it almost at the same time, it is no coincidence.
They all began to actively reduce the importance of their flagship models.
2. Flagship Models for Brand Endorsement, Volume Models for Profit Generation
I've been thinking about why it happened exactly at this time.
The answer is actually not complicated.
In the past, the biggest value of flagship models was not to make money, but to prove technological leadership. After GPT-4 was released, OpenAI's valuation kept rising. Every time Claude updated, Anthropic would redefine its own technological position. Flagship models undertake the brand value.
But that's not where enterprises actually spend money.
An enterprise runs millions of API calls every day. High-frequency tasks such as customer service, search, approval, code generation and Agent execution take up the vast majority of token consumption. What enterprises care most about in procurement is not the first place in Benchmark, but the cost of one task, the stability, and the ROI.
When the call scale expands to tens of millions per day, the slight intelligence advantage of the flagship model will be instantly offset by the huge computing power cost.
OpenAI's moves and statements mark a turning point: Top-tier flagship models no longer shoulder the heavy responsibility of making profits.
3. AI is Entering the Era of "Mass-Market Vehicle Models"
This scenario was experienced long ago in the automotive industry.
Twenty years ago, the 7 Series defined the height of BMW, the S-Class supported the luxury tone of Mercedes-Benz, and the A8 established the flagship image of Audi. Flagship models determined the upper limit of the brand. But the base that truly supports brand sales and generates profits has always been the BMW 3 Series, Mercedes-Benz C-Class and Audi A4.
Later it became more obvious. Model S proved that Tesla could make cars. It was Model 3 and Model Y that truly made it a global automaker.
Flagship models are responsible for proving capabilities, while mass-market models are responsible for scale.
AI is also starting to enter this stage today. Sol and Fable will continue to exist, they are responsible for pushing the upper limit of technology and refreshing Benchmarks. What truly undertakes commercialization will increasingly be Luna, Terra and Opus.
OpenAI even publicly wrote out the recommendation process this time: Sol is responsible for planning, and Luna is responsible for execution.
This is no longer a single model, OpenAI is designing a division of labor system for models. In the future, what enterprises purchase will not be just one model, but a complete model system. What really determines the cost is not the chief designer, but the construction team that works every day.
4. Models Start to Optimize Themselves
There is another detail that I think is more interesting than the price cut itself.
OpenAI mentioned in the technical notes that this price reduction is not only because more GPUs were procured, nor is it simply due to larger scale. The real reason is that models have begun to participate in optimizing models.
Specifically, under the leadership of human engineers, Sol independently rewrote and optimized the underlying production kernel. It designed hundreds of experiments on its own to improve token generation efficiency, and even participated in monitoring the model training process, intervening directly once problems are found.
As a result, the end-to-end operating cost decreased by 20%, and the token generation efficiency increased by 15%.
This piece of information is easy to ignore, but it is of great significance.
In the past, efficiency improvement relied on engineers. After the model was launched, people optimized the reasoning framework, CUDA, caching strategies and scheduling algorithms bit by bit. It was already pretty good to get a dozen percentage points of efficiency improvement in a whole year.
Today, technological evolution has embarked on a new path. Models have begun to take over the engineering optimization of underlying code and computing power scheduling, iterating operation around the clock.
This is a cycle that keeps accelerating. The smarter the model is, the stronger its ability to participate in optimization. The faster the optimization, the faster the cost drops. The lower the cost, the larger the call volume. The larger the call volume, the more data is generated, which will continue to train the model.
I checked the price changes of OpenAI in the past two and a half years. The input price of the first released GPT-4 was $30 per million tokens, it dropped to $5 for GPT-4o, and GPT-4o mini achieved $0.15. Now the pricing of Luna is close to the low-cost range of the early mini models, but its overall intelligence level has long far exceeded the expensive GPT-4 two years ago.
In more than two years, the price has dropped by nearly two orders of magnitude. If the model continues to participate in optimizing itself, this curve will most likely continue to go down.
The truly terrifying thing is not the 80% price cut today, but that cost reduction has begun to have self-driving capabilities.
5. From "Who is the Smartest" to "Who is the Most Cost-Effective"
I increasingly feel that when people discuss AI competition today, they sometimes still apply the framework of the previous stage.
For example, people often keep asking: Who is the smartest? GPT, Claude, Gemini, DeepSeek. Every time a new model is released, the first thing the media looks at is the ranking list. Whoever ranks first wins.
However, the new moves of OpenAI and Anthropic break this rule, and they convey a new signal to the market: The appeal of the single-performance champion is fading.
OpenAI mentioned a sentence in the announcement that I think is particularly critical: it advises developers to match different models according to the importance of the task, the cost of error, the urgency and the scale.
Note that what it is discussing is no longer the model, but the task.
In the past, when an enterprise deployed AI, there was usually only one choice. From today on, it is more like building an organization. Use Sol for the most important tasks, hand over daily execution to Luna, and there may be more lightweight models in the future responsible for simpler things.
This is very similar to the cloud computing industry in the early days. No one will store all data on the most expensive SSD. Hot data is stored on SSD, ordinary data on HDD, and cold data on object storage. People never discuss which hard drive is the fastest, but how to build the entire system in the most cost-effective way.
DeepSeek actually sensed this direction earlier than Silicon Valley. In the past six months, it has hardly emphasized that it must be the smartest in the world, and the few words it keeps repeating are: cheap, fast, and competent enough.
After the cache is hit, the cost per million tokens is so low that it can almost be ignored. It has been betting on one thing: what enterprises need is not the world champion, but the deployment option with the highest comprehensive ROI.
This is not a short-term price war. The follow-up of the two Silicon Valley giants now verifies the inevitability of this commercial path.
6. What OpenAI Truly Wants to Sell is Not Models
Up to now, OpenAI's strategy is very clear. What it really wants to sell is no longer models, but call volume.
In the past, model manufacturers relied on high technology premium to obtain high gross profit margins. Now the business logic has changed to exchanging extremely low thresholds for a super-large-scale traffic ecosystem.
Microsoft makes real money not because Windows is expensive, but because Windows is installed on all computers. AWS makes money not because the profit of a single server is high, but because countless applications around the world run on it every day.
Platform revenue has always relied on penetration rate.
This explains why the price of Sol has never changed. Sol undertakes the brand mission, proving that OpenAI is still the company with the highest technological upper limit. And Luna is the revenue engine.
OpenAI hopes developers form a new default habit: use Luna to write Agents, use Luna to run workflows, and use Luna for batch execution. Only when encountering really difficult problems will they call Sol once.
Once this default habit is established, the subsequent competition will become very difficult. Migrating the underlying model for an enterprise means re-testing, verifying and adapting the entire workflow. The migration cost will get higher and higher.
The real moat has begun to shift from capability leadership to ecosystem stickiness.
7. Flagship Models No Longer Determine The Direction
Looking at it over a longer time horizon, the AI industry is going through a very typical inflection point of economies of scale.
When the unit computing power cost drops to an extremely low level, the total market demand for tokens does not decrease with the drop in unit price, but explodes exponentially.
Flagship models no longer determine the direction, because the focus of technological evolution has shifted from exploring the upper limit of intelligence to industrialized cost reduction of computing power. When the API call cost is low enough to be negligible, the form and boundary of models will begin to blur.
Enterprise developers will no longer focus on the consumption of each token, but seamlessly embed AI into every business process.
The most profound part of this transformation is that the avalanche of API prices is pushing up the migration cost of the entire software engineering. Once the enterprise's workflows, Agent scheduling networks and automated data pipelines are all built on the combination of low-cost models, the underlying model provider will lock in the computing power pipeline for the next ten years.
In the past few years, large model companies sold intelligence. Starting from today, they begin to sell efficiency.
These are two completely different stories.
Words beyond the layout:
In the past, we used to compare the development of AI to consumer electronics, expecting new flagships that break records from time to time.
But perhaps the true victory form of AI is not to become a sensational product.
When the steam engine was first invented, people marveled at its productivity, but today, electricity flows everywhere, driving the operation of the entire civilization, and no one will specifically discuss it anymore.
When humans no longer frantically discuss which intelligence benchmark a new flagship model has refreshed, AI quietly embeds into every system and every instruction, becoming the default water and electricity called behind all automated processes.
Its real era has just begun.
This article is from the WeChat official account "Beyond the Layout", author: Huahua, published with authorization by 36Kr.