Anthropic and OpenAI cut prices on Opus and GPT-6: the race is no longer just about intelligence, it's about price

By Zendoric · September 23, 2026.

Anthropic has unveiled Opus 5.5, the new version of its flagship model for programming and complex knowledge work, and OpenAI has responded with GPT-6 Sol and Luna, the mid-range and budget variants of its GPT-6 family. According to Anthropic's announcement, Opus 5.5 is priced at $4 and $20 per million input and output tokens respectively —20% less than Opus 5— while cache reads (the reuse of already-processed context, which Anthropic itself says accounts for most of the cost in agentic and coding tasks) fall by 60%, to $0.20 per million tokens. Anthropic adds that, factoring in that Opus 5.5 also needs fewer tokens to complete a task, the real saving on typical workloads is around 40% compared with Opus 5.

OpenAI has followed the same script with GPT-6 Sol ($2 input and $10 output per million tokens) and GPT-6 Luna ($0.10 and $0.50), according to the prices published by the company: roughly half those of their predecessors, with improvements of only a few percentage points on certain benchmarks, according to OpenAI. Neither company is selling these launches as a leap in capability. They are selling them as a price cut.

Context matters. In early September OpenAI had already launched GPT-6 Astra, its most powerful and expensive model, which on some benchmarks and in users' perception outperformed Opus 5, as Ars Technica reports. Anthropic now maintains, on the basis of its own benchmarks and those of its partners, that Opus 5.5 performs better than Astra in coding and knowledge work "in some cases," though modestly. It is one more skirmish in a positioning war: OpenAI has renamed its lineup with the labels Astra, Sol, Terra and Luna —from most to least powerful and expensive— which correspond roughly, if imperfectly, to Fable, Opus, Sonnet and Haiku in Anthropic's catalog.

There is one detail worth not overlooking: Anthropic describes Opus 5.5 as especially capable in "high-risk areas" such as cybersecurity and biology, and applies the same safeguards that Fable 5.1 already carried — requests detected as sensitive territory are automatically and transparently rerouted to an earlier model. It is a reminder that gaining efficiency and power is not free of safety friction, and that frontier labs still treat certain capabilities as something to contain, not merely to sell.

The underlying reason for this wave of price cuts is well explained by Ars Technica's own analysis: companies and developers are using model routers —systems that send each request to the cheapest model capable of handling it— in order to stop paying frontier-model prices for tasks that do not require them, while open-weight models (from Alibaba and others) apply pressure from below. This fits something we have been pointing out at Zendoric for some time: the open frontier —GLM, Qwen, DeepSeek, Kimi— has risen fast enough to force closed labs to compete on cost, not just on intelligence. Anthropic's and OpenAI's response is not generosity: it is a defense of market share against an increasingly competitive and cheap alternative.

Our reading is that this "more for less" phase marks a change of conversation in the industry. For two years the scoreboard has been almost exclusively about benchmarks: who posts the highest score. Now the enterprise customer, the one who pays these models' real bills, is starting to call the shots on a different variable: the cost per task solved, not the cost per raw token. That favors whoever solves problems more cheaply even if they are not the smartest in the class, and penalizes labs that only know how to compete by raising the price of frontier intelligence. It is a classic commoditization dynamic, and it is healthy: it squeezes the current leaders' margins, but it accelerates exactly the kind of cheapening of artificial intelligence that, sustained over time, is the necessary condition for the abundance we defend as this technology's horizon. Without price cuts like these, AI remains a corporate luxury; with them, it becomes infrastructure within anyone's reach.

In the short term, however, this is also a sign that the pace of raw capability improvement is slowing relative to the pace of efficiency gains — something Ars Technica rightly notes at the end of its piece: the orchestration layer and the harnesses (the software that wraps and steers the model) now matter as much as the model itself, and companies are less interested in the frontier and more in deploying what already works predictably and cheaply. That is not stagnation: it is the maturing of a market that is beginning to look like any other technology sector, where the price per unit of value delivered ends up weighing more than the laboratory record.