The price war of OpenAI's large models has already written its ending in history.

From the last week of June to early July, the entire AI track underwent collective repricing.

Anthropic took the lead by launching Sonnet 5, offering a limited-time discount to $2/10. A little over a week later, xAI, OpenAI, and Meta released products intensively within 48 hours — Grok 4.5, GPT-5.6 in three tiers, Muse Spark 1.1. Four companies, ten natural days, and every price tier was affected. On July 15, Altman added on X: Sol is already half the price of Fable 5, "happy to deliver at a quarter of the price." Another 75% reduction.

Those who lowered their prices incurred a loss of $9.3 billion in the first quarter. The one selling at high prices started to turn a profit.

This is not what a price war should look like. Three nearly identical wars in history have already written the script for this one.

The expensive one is actually making money

Anthropic’s annualized revenue rose from $9 billion at the end of last year to $47 billion by May this year. OpenAI surpassed $30 billion in the same period. In the first quarter, OpenAI's revenue was $5.7 billion, Anthropic’s was $4.8 billion — OpenAI was still ahead. However, Anthropic projected $10.9 billion in revenue for the second quarter, doubling sequentially and expected to achieve its first operating profit of $559 million. OpenAI reported an operating loss of $9.3 billion in the first quarter.

The ones lowering prices are facing massive losses. The one selling at high prices started to make profits.

The reason lies in the revenue structure. About 85% of Anthropic’s revenue comes from enterprise clients — over 1,000 clients are now paying more than $1 million annually. These clients are purchasing stability, safety, and compliance, not the lowest Token quote. A 75% price drop would not lead them to buy one more Token; a 75% price increase wouldn’t cause them to buy less. OpenAI is quite the opposite — over 60% of its revenue comes from consumer subscriptions, and each price drop stimulates price-sensitive end users and small to medium developers.

Anthropic is not facing a price-sensitive market. Altman is. When Altman called for "another 75% cut" on July 15 on X, Anthropic did not reply — it publicly encouraged enterprises not to reduce AI use due to cost concerns. It's not that they didn't respond; they responded in another language.

Thirty years ago, Intel tried this tactic

In 1998, AMD's market share jumped from 8% to 16%, posing a real threat to Intel for the first time. Intel’s response was a price cut unprecedentedly described by semiconductor analysts — not just a round, but an ongoing scorched-earth assault for many years. Whenever AMD launched a competitive chip, Intel would drop prices in a scorched-earth manner. The confidence was simple: 13 wafer fabs compared to AMD’s 2, with a cost structure that competitors could not replicate.

AMD survived. It was forced to sell off fabs, but did something Intel did not anticipate: it redefined its products rather than imitated. The Ryzen series launched in 2017, and AMD turned around with its products, at one point exceeding Intel’s market cap.

Intel used its cost advantage to keep AMD down for years. But cost advantages expire; product advantages do not.

What Altman is doing is essentially no different from what Intel did in 1998. He has listed cost-cutting techniques that compress inference costs by 50% as “trade secrets.” The Information quoted insiders: “They even don’t want other employees in OpenAI to know because if it leaks, it will quickly be adopted by other labs.” The key term is not “secret,” but “adopted” — once competitors master the same efficiency tools, the only weapon he has left becomes burning cash. Intel didn’t have to burn cash back then; it relied on having more factories to support its price cuts.

Flywheel or pump?

From 2006 to 2018, AWS proactively lowered prices over 100 times. S3 storage saw an 85% price drop over 12 years. But during each price cut announcement, AWS was always profitable.

The flywheel can turn for one reason only: every penny saved comes from real efficiency gains — self-developed Graviton processors have a 40% better price-to-performance ratio than x86, and Nitro offloads virtualization costs to dedicated hardware. No matter how low prices are lowered, costs are lowered even more. That’s what the flywheel is.

What Altman wants to convey is this story. “Inference cost reduced by 50%,” “trade secret” — all wording conveys the same signal: this is not burning cash to seize the market, but returning the cost saved through technological advancements to customers.

However, one number punctures the narrative. OpenAI incurred a $9.3 billion operating loss in the first quarter, losing $1.6 for every dollar earned. AWS has never been in such a deep loss during a price cut. The premise of the flywheel is that price curves and cost curves decline in sync, with costs ideally declining faster. If the cost curve does not keep up — if each price cut is accompanied by funding to fill the price gap — then it’s not a flywheel; it’s a pump.

What Altman has at the entrance of the pump is the $122 billion financing completed in March and $73 billion in cash on hand, while the exit is the IPO. If the pump stops, the water stops too. More troubling is that by the end of 2025, OpenAI's power procurement commitment to cloud service providers amounts to $665 billion, extending to 2030 — regardless of whether AI demand grows as expected by then, this money has to be paid.

The ones burning cash have no way out

In 2014, the subsidy war between Didi and Kuaidi lasted less than half a year, burning through $2.4 billion, with losses peaking at tens of millions daily. They announced a merger on Valentine’s Day 2015. In October of the same year, Meituan and Dianping merged — in an internal email sent by Wang Xing on the day of the merger, he stated bluntly: “Yesterday both sides fought fiercely, today we are shaking hands and being friendly.”

The push for the merger didn’t come from the founders but from investors. Sequoia was the A round investor for both Meituan and Dianping — the harder they fought, the more they burned money on both sides. The logic is simple: neither side can kill the other, and if they continue fighting, neither can survive to IPO.

The price wars of Didi and Meituan ended in a merger, based on a common premise: neither side was making money.

OpenAI and Anthropic do not share the same landscape. Anthropic is already profitable; it’s not that it can’t compete in a price war, but it doesn’t need to. OpenAI is more like Didi — large scale, burning a lot of cash, eager to go public, must stop the bleeding before the IPO. The route Didi found was a merger; OpenAI does not have this route. Microsoft and Amazon are the two largest investors, but they truly care about how much compute power their cloud platforms sell — AWS Bedrock distributes Anthropic's model, taking a big chunk of the profit for themselves. A merger will not happen: both clients are too locked in, it would only create antitrust bombs, and Anthropic is already profitable, showing no motivation for a merger.

Which path is Altman on?

Three historical cases point to the same question: did the ones lowering prices ultimately win?

Intel won for ten years but lost to products. AWS won because its efficiency is real. Didi merged because both sides were losing — but Anthropic is already profitable, so that path does not exist.

If Altman's cost advantage in inference is real — not just accounting optimization, not selective disclosure, but systematically compressing unit inference costs below one-third of Anthropic's levels — then this isn’t a price war, it’s an AWS-style reshaping of the landscape. No matter how much Anthropic emphasizes “enterprise customers don’t care about price,” it cannot infinitely maintain a premium under the premise that competitors are three-quarters cheaper. Enterprise clients might not care about price, but CFOs care about costs.

If the “savings” aren’t enough — the price cuts mainly rely on financing — the capital market will ask him the question he can't answer: your price is already half of your competitor's, why are you still losing money?

The most brutal part of July 15 was that on the same day Altman declared “happy to drop another 75%,” Ed Zitron published a 15,000-word piece titled "AI Lehman." Additionally, OpenAI’s IPO has been advised to be postponed until 2027 — the options are: either wait a year to list at a trillion valuation, or lower the valuation now for a quicker listing. Altman’s response: no adjustments to the trillion-dollar valuation will be accepted. The market does not need to wait for financial reports; it just needs to hear this statement to recalculate.

Anthropic is betting that enterprise clients buying AI is different from buying cloud services — the depth of embedding Claude Code into development processes, and Anthropic's refusal to meet the Pentagon’s demands for removing safety barriers accumulate trust dividends, which are not on the same competitive dimension as “a few dollars per million Tokens.” OpenAI is betting on scale — 900 million weekly active users, Microsoft's family of tools being the default model, and the largest developer community. As prices fall to critical points, Token consumption expands exponentially; it bets on capturing the largest incremental share.

Both bets are facing validation.

Time will provide the answers

In the coming months, this gamble will be validated one by one.

August 31 is the first milestone. If Anthropic's Sonnet 5’s limited-time discount expires and the price returns from $2/10 to $3/15, without an extension or further price reduction signals, it will be telling the market: I won’t follow. If the Q3 renewal rates and average order value of enterprise clients remain stable, this would demonstrate with numbers that this question of "to follow or not" itself is invalid.

October is the second. Anthropic's IPO timing window — the first public market yardstick for the AI industry. Whether it’s priced at $965 billion or a trillion directly determines OpenAI’s subsequent valuation negotiation space. OpenAI’s IPO has been advised to be postponed until 2027. The $122 billion financing from March is in place, but if further financing is needed afterward, the implicit IPO timeline pressure and valuation commitments in the terms will reveal just how much "savings" Altman has left.

Then comes the financial report. Anthropic will observe if profitability can be sustained, while OpenAI will look at whether losses shrink.

Finally, it’s about clients. If a Fortune 500 company switches from Anthropic to OpenAI, even if it's just one case, it will be seen by the market as the first signal that the price war is breaching premium logic.

Altman's true opponent is not Anthropic; it is time.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。