Analysis: AI price war reignites as OpenAI, Anthropic, Xiaomi, SpaceXAI release cheaper, more efficient models

Editor's Note: This article was originally published on WeChat public account "Lei Tech AGI" (ID: leitech) by author 三七二十一, and is republished here with authorization from Cyzone.

A new round of price wars appears to be underway in the large language model (LLM) space.

Over the past few days, OpenAI, Anthropic, Xiaomi, and SpaceXAI have all launched new models in quick succession. While GPT-6 Sol/Luna, Claude Opus 5.5, MiMo-V2.6, and Grok 4.7 differ in name and positioning, they share a common theme: making intelligence more affordable.

Price Cuts and Performance Boosts

OpenAI has taken the most direct approach. The price of GPT-6 Sol has been nearly halved compared to its predecessor, while Luna pushes even further into low-cost territory. Anthropic has cut the API unit price of Claude Opus 5.5 by 20%, and claims the actual cost of typical tasks has dropped by approximately 40%.

Not all moves are outright price reductions. Xiaomi's MiMo-V2.6 and SpaceXAI's Grok 4.7 have maintained their previous generation's pricing while pushing performance forward. For developers, however, the net effect is similar: the same amount of money is buying more machine intelligence at an accelerating pace.

Image credit: Leitech @ GPT

Why Models Are Getting Cheaper and Stronger

This trend is not entirely new. LLM API prices have been declining for years, particularly driven by competition among open-source model vendors in China, which has significantly reduced the cost per million tokens. What makes this round different is its timing.

On one hand, OpenAI and Anthropic have begun publicly discussing slowing the pace of frontier model development, citing concerns about safety alignment risks from rapid AI capability growth. On the other hand, several model vendors are simultaneously accelerating efforts to make existing capabilities cheaper and more efficient.

Meanwhile, AI agents are moving from chat interfaces into programming, research, and office work. A single task may now consume not thousands of tokens, but hundreds of thousands or even millions, with continuous model calls lasting hours.

Where Are the Savings Coming From?

According to explanations from various vendors, the direct cause is improved efficiency in both model training and inference. Compared to the past, when scaling pre-training was the primary focus, model vendors now have many more levers to pull for efficiency gains.

OpenAI attributes the price reduction of GPT-6 Sol/Luna directly to improvements in caching and inference efficiency. Both models inherit some training methods and capabilities from GPT-6 Astra but reduce costs further on the inference side, allowing savings to be passed on to API pricing.

Anthropic has gone a step further. While the API unit price of Claude Opus 5.5 is only 20% lower than Opus 5, Anthropic claims the actual operating cost for typical tasks has dropped by 40%. Beyond cheaper individual tokens, the new model requires fewer tokens to complete the same task, and cache read prices have fallen by 60%.

Image credit: Anthropic

Smarter models are themselves becoming a form of price reduction.

Xiaomi has focused more on the training side. MiMo-V2.6 maintains its API price while expanding the scale of agentic reinforcement learning over V2.5. During less than six days of live reinforcement learning training, Flash and Pro models generated approximately 750,000 trajectories. Through larger batches, more complex task environments, and finer reward mechanisms, Xiaomi says the model learned to complete tasks with shorter paths and fewer tokens. As a result, V2.6 Flash now comprehensively surpasses the previous generation V2.5 Pro at the same price.

SpaceXAI has taken a different route. Grok 4.7 uses a larger base model rather than a smaller one, with extended training on more difficult tasks, enhanced reinforcement learning, and improvements in self-verification, long-context handling, and agent capabilities. While maintaining the same API price as Grok 4.6, it has pushed coding and agent task performance further.

These four technical approaches differ, but the outcome converges. Today, model vendors are seeking efficiency across the entire chain—from post-training reinforcement learning and inference optimization to prompt caching and reducing ineffective token and tool usage. This is the most direct technical backdrop for the current price war.

LLMs are still getting stronger, but increasingly, capability improvements no longer require simply trading higher inference costs. For developers, the result is straightforward: the same task can now be completed with less money.

Why Brake and Accelerate at the Same Time?

Technical efficiency gains alone do not fully explain this round of price wars.

Just a month ago, OpenAI announced a temporary slowdown in model scaling, including a two-week pause in reinforcement learning training for newly deployed models and continued suspension of the largest-scale frontier reinforcement learning training. Two direct reasons were cited: internal agent safety incidents and GPT-6 Astra reaching a "critical cybersecurity capability" threshold.

The core issue is that frontier model capabilities are growing too fast for monitoring, alignment, and safety measures to keep pace.

Anthropic CEO Dario Amodei went further, publicly proposing "Pace the Frontier" and calling on the entire industry to slow the rate of capability growth in frontier AI models. Anthropic even specifically mentioned this in the release of Claude Opus 5.5.

Image credit: Wikimedia

At first glance, this seems contradictory to the ongoing price war.

However, a key point is often overlooked: OpenAI and Anthropic are not calling for a slowdown in technological progress across the entire AI industry, but rather a slowdown in the expansion of the most advanced capability frontiers.

According to Amodei, "pacing" does not mean stopping model training or technological advancement. On the contrary, he believes today's models themselves represent a vast "research goldmine" worth investing more time in understanding, aligning, and utilizing.

This provides another lens for understanding the current price war.

OpenAI offers a telling data point. After imposing stricter safety restrictions on Astra in August, GPU allocation for Astra-class models dropped by 59.2%. However, GPU allocation for other model categories increased by 17.2%, offsetting approximately 85% of the decline. Computing power did not sit idle—it simply flowed to other models and experiments.

This does not prove that GPT-6 Sol/Luna's price cuts are a direct result of slowing Astra's development. But it does illustrate that competition among top model vendors is not limited to a single path of "training stronger models."

The same computing power and R&D investment can be directed toward pushing the next capability ceiling, or toward optimizing existing models through reinforcement learning, inference optimization, caching, distillation, and engineering improvements—making already-achieved capabilities cheaper and more reliable.

The Rise of Agents Amplifies the Value of Efficiency

In the chatbot era, a single response might consume only a few thousand tokens. In the agent era, models work continuously for hours, repeatedly reading context, calling tools, self-checking, and even orchestrating other agents. Every reduction in model cost can directly determine whether an agent product can scale.

But agents have another dimension. When explaining why he began supporting a slowdown in frontier models, Amodei cited a core change: Recursive Self-Improvement (RSI)—in simple terms, AI increasingly participates in the development of the next generation of AI.

Anthropic disclosed in August that Claude already plays a "dominant" role in 26% of internal AI R&D tasks, with over 90% of related work involving human-AI collaboration. OpenAI has observed similar trends, with coding agents increasingly involved in writing code, building experiments, analyzing results, and advancing model research.

Cheaper, better models mean more agents can run. More agents can participate in model R&D, improving researcher efficiency and accelerating the training and iteration of next-generation models.

The cheaper intelligence becomes, the easier it is to produce more intelligence. This accelerating diffusion of intelligence may be the most noteworthy aspect of the current price war.

Final Thoughts

Over the past few years, LLM price reductions have become almost taken for granted. With more models, improving computing and algorithmic efficiency, prices have naturally trended downward.

But as OpenAI and Anthropic seriously discuss slowing the growth of frontier capabilities, another competition becomes clearer: top-tier machine intelligence is being optimized, democratized, and diffused at an accelerating pace.

As William Gibson wrote in Neuromancer: "The future is already here—it's just not evenly distributed." Now, that uneven distribution of machine intelligence is rapidly disappearing.