Anthropic has introduced Claude Sonnet 5.5, the latest version of its language model, designed to meet the needs of most users. This update follows the release of Opus 5.5 by just a few days. The company claims that Sonnet 5.5 offers improved intelligence and lower usage costs. The price per token remains the same as Sonnet 5: $2 per million input tokens and $10 per million output tokens. However, Sonnet 5.5 uses fewer tokens to complete tasks, which should reduce operating costs by up to 30% and increase speed by 30%. According to performance tests, Opus may no longer be as necessary, allowing users to save costs by switching to Sonnet. While Sonnet 5.5 is generally second to Opus 5.5 in performance, the difference between the two is small. On the Terminal-Bench 4.0 test, which evaluates AI efficiency in code-related tasks, Sonnet 5.5 scores 70%, compared to 10% for its predecessor. Anthropic compares its models to GPT-5.6 Sol, the previous version from OpenAI, which scored a maximum of 37% on the same test. These results are supported by independent evaluations, including those from Artificial Analysis, which confirm Sonnet 5.5 as the top performer, ahead of Opus 5.5. The best large language model from OpenAI, GPT-6 Astra, scores 56% on the same test. OpenAI has postponed the release of GPT-6.1 Astra, an updated version of its top model, originally expected during the OpenAI DevDay 2026 event, which begins at 7 p.m. in the company's time zone. OpenAI, led by Sam Altman, has told the press, including the Associated Press, that the release is delayed because the model has not met the company's standards, particularly in terms of safety. This issue has sparked concern among experts, some of whom warn of potential catastrophic consequences if the development of powerful AI continues unchecked. While the extent of the danger and the motives of critics are debatable, it is clear that AI systems are not adequately regulated. These systems have been found to perform actions that should be prohibited, leading to incidents such as website and online service breaches. OpenAI has been involved in some of these uncontrolled events and had announced a pause in training its most advanced models over the weekend. The delay of GPT-6.1 Astra is likely a direct result of this decision, though no new release date has been set. OpenAI promises new features during its conference, but it remains unclear whether the event will preview consumer-facing AI products, such as those being developed with Jony Ive. Thibaut Sottiaux announced on X that the 200 € per month plan for the largest users of OpenAI models will return on September 30, 2026. The plan had been suspended two weeks ago after the success of GPT-6 Astra forced the company to halt new registrations. However, this reopening comes with a drawback: the plan now offers a quota that is half the size of the previous one. In exchange, OpenAI will not restore the 5-hour limit on this plan, allowing all tokens to be used freely within the weekly quota. Additionally, lighter versions of GPT-6, Sol and Luna, are priced 50% less than the previous 5.6 versions. This could allow users to maintain their previous level of usage by switching to these models, according to the company's manager. However, the changes are complex and still unclear, especially since OpenAI promises new features for these subscribers, which will be revealed during tonight's conference. For now, the update seems to offer a reduction in quotas at the same price without clear advantages. The changes to OpenAI's plans and the delayed release of GPT-6.1 Astra highlight the ongoing challenges in balancing innovation with safety and regulation in the AI field. As companies like Anthropic and OpenAI continue to push the boundaries of what AI can do, the industry faces increasing pressure to ensure these technologies are developed responsibly. Tonight's OpenAI conference may provide more clarity on the company's roadmap and the future of AI for both developers and the general public.