Anthropic Releases Opus 4.7, Not as ‘Broadly Capable’ as Mythos AI
Anthropic has launched a new version of its Opus AI model, which it says has notable improvements in software engineering and performing complex, long-running tasks.
The announcement comes a week after Anthropic previewed a new model, Mythos, to a select few customers, which is set to supersede Opus as the most powerful AI model in the company’s portfolio. Anthropic made mention of this new model in the Opus announcement, saying Opus 4.7 would not be as “broadly capable” as Mythos and that it had experimented with ways to tone down the model’s cyber capabilities before release.
For at least the next few weeks, Opus will remain the company’s most powerful publicly available model. Alongside more than 10 percent improvements in agentic coding, according to benchmarks published by Anthropic, the company also said the AI model has far better visual reasoning and is more creative when designing presentations and slides.
It is not the first time Anthropic has reduced the power of its AI models.
In late 2025, it launched Claude Haiku, which it said was its fastest and most efficient model yet. It also happened to be significantly cheaper for the company to build. In this new environment, it appears to be more about damage control, which is forcing Anthropic to reduce its capabilities, although it is also reportedly struggling to manage the huge demand for its AI models from businesses.
Could Mythos ever reach public availability?
The launch of a toned-down version of Mythos suggests Anthropic doesn’t want to release its most powerful model to the general public any time soon. Over the past week, it has provided access to banking and government institutions in controlled pilots, aimed at familiarising them with the AI model and demonstrating its capabilities.
The US government, which just last month banned federal agencies from using Claude and designated Anthropic a “supply chain risk”, is reportedly pushing for access to Mythos. It marks a sharp turnaround only a month after the Pentagon spat, which Anthropic is planning to fight in court, although the Pentagon may not have stopped using the model even with the federal ban.
Beyond the high risk of cybercriminals using the AI model to exploit system vulnerabilities, Anthropic is also concerned about the model being distilled by Chinese operators. It has previously accused firms like DeepSeek of distilling its model, which can then be reconfigured and used as a tool for cyberattacks, particularly as DeepSeek and other Chinese AI models are open-source.
It is not the only firm that has looked at restricting access to its AI models. OpenAI has made GPT-5.4-Cyber available to a similarly select group of institutions due to cybersecurity risks. Google and Meta have also previously held back video models from public release, citing concerns about misinformation and deepfakes.
For more on Anthropic’s enterprise push, check out how Claude now connects Excel and PowerPoint workflows to turn data into presentations without switching apps.
About Author
