The price cut as a market war

In just a few months, OpenAI and Anthropic have executed a series of successive price cuts on their language model APIs, accumulating reductions of up to 80% on certain models. What began as a competitive rivalry between two companies has become the most impactful event in the economics of artificial intelligence in the entire history of the sector.

The chronology is almost brutal. In June 2026, OpenAI launched GPT-5.6 Luna with initial prices of US$1 per million input tokens. On July 30, it announced an 80% cut — dropping to US$0.20/US$1.20/US$0.02 (input/output/cache). Anthropic responded in August by making permanent the US$2/US$10 price of Claude Sonnet 5, which had been launched as an introductory price in June. In September, OpenAI cut GPT-6 Sol cache by half, accumulating a cumulative reduction of 75% over the original prices. Anthropic, meanwhile, launched Claude Opus 5.5 on September 22 at US$4/US$20, a 20% cut from the original Opus 5.

What makes this movement different from any previous price cut in the technology sector is the speed and magnitude. Reductions of 80% in three months are equivalent to what the smartphone market took a decade to achieve — and this in a market that has existed for only two years.

The dynamics behind the cuts

Analyzing these numbers in isolation is insufficient. What really matters is the competitive context driving them. The DeepSeek V4 Flash, launched in April 2026 at US$0.14/US$0.28/US$0.0028, created pressure that no other company in the sector had faced. The Chinese model, sold at a tiny fraction of OpenAI's prices, gained massive adoption — 13 providers on OpenRouter in May, jumping to 29 providers with the 1017 variant. For OpenAI, Luna at $1 would have lost to DeepSeek V4 Pro in cost per task. The 80% cut was not strategy, it was survival.

Market behavior also plays a central role. The analysis by Max Woolf shows that OpenAI and Anthropic are running a "loss leader" dynamic — selling below actual cost to consolidate market share before open-weight competitors gain traction. The comparison with Uber and Lyft in the 2010s is immediate: artificially low prices to eliminate competitors and then control price formation. However, there is a crucial difference. Unlike the ride-hailing market, the LLM inference market is not a duopoly. Providers like DeepSeek, open-weight models, and aggregators like OpenRouter maintain real competitive pressure.

The impact on AI agents

The dramatic drop in prices has consequences that go far beyond the monthly bill. With inference costs falling, the economic viability of autonomous agents changes completely. Companies that previously calculated whether an agent was worth it at $0.50 per interaction can now run them for pennies. This means the number of interactions per agent multiplies — and with it, both the practical utility and the security risks. Each additional agent turn costs less, but it's also easier to create agents that run billions of interactions without human supervision.

Asana has already reported a 30% reduction in task completion latency and up to 2.5x faster inference per agent turn using Claude Haiku 5.5, launched on October 7 at US$0.10. The same price that OpenAI charges for GPT-6 Luna. The price parity between the two main AI platforms in the low-cost segment is a symbolic moment — there is less competitive difference between OpenAI and Anthropic in this segment than in any other price range.

For individual developers, the price drop means it is possible to run agents continuously at a fraction of the previous cost. Woolf herself, who maintained a US$100/month subscription with OpenAI's 5x Pro plan, reported that as costs fell 75%, she completely changed her consumption pattern — moving from moderate use to near-continuous agent execution. The catch: OpenAI simultaneously cut the usage allowance for Pro users by 50%, under the argument of "increasingly efficient models." It's the digital equivalent of reducing the amount of gas in the tank and expecting the car to go further.

The IPO mystery

One of the most curious aspects of the price cuts is the timing. Both OpenAI and Anthropic are in the process of preparing for an IPO — OpenAI, expected in 2027, and Anthropic, within a couple of months. Anthropic's IPO prospectus revealed scaling costs and P&L that, according to the Financial Times, showed an "adjusted operating profit" for the quarter. Reports indicate that profit margins per token can be as healthy as 80%, with significant caveats.

Cutting the main revenue product by 80% before an IPO is the opposite of what a company wants to show to investors. The most plausible explanation is that these cuts are, in fact, a signal of confidence — the companies are willing to sacrifice short-term revenue to consolidate market dominance before going public. If prices remain low (which is likely, given the competitive ecosystem), future revenue will be lower, but customer base and market share will be unquestionable.

Where the market is heading

Three trends emerge from this analysis. First, the price drop is likely irreversible. Just as data storage fell to the point of being practically free, LLM inference tends toward minimum marginal cost. Second, competition will not be won by price — but by perceived quality, tool ecosystems, and integration. Third, the big winner of this price war is the end user, but the potential big loser is the financial sustainability of startups that depend on third-party API margins. If OpenAI and Anthropic cut so deep they cannot sustain R&D costs, who pays for the next generation of models?

The price cut in LLMs is not just a change of price table — it is a sign that the era of inference as a product is transforming into infrastructure. When the ability to generate text or code becomes as cheap as electricity, the real competitive advantage shifts to the ecosystem, not the model itself.

Sources: Max Woolf's Blog, The Next Web, TechRepublic

✓ Independent sources cross-checked and verified before publishing