The AI race is driving up the cost of access to its most powerful models

Anthropic says that Claude was leading 26% of its AI R&D work in August, up from less than 1% in February. The company considers a model to be leading a task when it completes most of the work assigned to the AI under human supervision.

If that estimate is accurate, each version could help build the next. However, the figure comes from Anthropic itself and has not been externally audited; nor does it prove that this collaboration shortens timelines, which depend on computing power and safety testing.

For users, the faster pace comes with a visible cost. In the OpenAI and Anthropic chats, the most powerful models are included in individual licences costing $100 or $200 per month, compared to $20 for the basic subscription. Those who prefer not to pay can skip each new release, but will lose track of what the latest models can already do.

The pace is also evident in release dates. Nikkei measured the average interval between versions at nine major companies in Estados Unidos y China: it dropped from 125 days between 2023 and March 2026 to 44 days since April. The comparison covers three years versus six months and depends on what is considered a version, a caveat that does not change the clear trend for those who use these tools.

Competitive pressure became clear in November 2025. After Google presented Gemini 3, OpenAI declared code red to focus efforts on ChatGPT and released GPT-5.2 twenty-three days later. The company said it had been preparing it for months.

Part of this speed is explained by how the models are built. Pretraining an LLM requires months of computation and an investment that is not often repeated. Then comes post-training, a shorter phase in which the already created system is fine-tuned. One of its techniques is reinforcement learning: the model receives a reward for completing a verifiable task, such as a program that passes all its tests, and adjusts its behaviour. This stage lasts weeks.