Anthropic PBC today introduced a new large language model, Claude Opus 4.8, that’s significantly better than its predecessor at complex coding tasks.

The company announced the LLM alongside another major business milestone. Anthropic has raised $65 billion in new funding at a $965 billion valuation to buy more computing infrastructure.

The company evaluated Claude Opus 4.8’s reasoning capabilities using Terminal-Bench 2.1, a benchmark that measures LLMs’ ability to perform tasks in the command line. The model scored 74.2%, a 8.4% improvement over the 4.7 version. Opus 4.8 scored 4.9% higher on a more general-purpose coding test called SWE-Bench Pro. Additionally, Anthropic logged improvements across benchmarks that comprised computer use and financial analysis tasks.

The new model includes optimizations that enable it to detect erroneous prompt responses. According to Anthropic, it’s four times less likely than its predecessor to output faulty code without pointing out the issue. It’s also less prone to making unsupported claims and missing misuse attempts.

Opus 4.8 powers a new feature called dynamic workflows in Claude Code. The capability enables the programming assistant to perform complex tasks such as rewriting an entire application in a new language. According to Anthropic, Opus 4.8 completes such projects by generating hundreds of AI agents that perform a small subset of the work.

An upgraded “fast mode” setting enables Claude Code users to trade off some output quality for lower latency. It makes the tool quicker 150% and costs three times less than the previous iteration of the setting. The Claude chatbot and Claude Cowork, a tool designed to automate multi-step business tasks, are receiving a similar feature that makes it possible to toggle between different response speeds. 

Anthropic plans to follow up Opus 4.8 with “Mythos-class models” that will become broadly available in the coming weeks. Claude Mythos Preview is an advanced LLM that made its debut last month. The model is adept at, among other tasks, finding cybersecurity vulnerabilities.

The company has only made Claude Mythos Preview available to a limited number of organizations over concerns that it could be misused by hackers. According to the company, its engineers are developing new guardrails that will address hacking risks. Anthropic expects that those guardrails will enable it to safely release Mythos-grade models to all its customers.

Investor optimism about the revenue potential of Mythos-grade models may have been one of the contributors to Anthropic’s valuation bump. At $965 billion, the company is now worth more than OpenAI Group PBC. Anthropic’s revenue is already growing rapidly: The company disclosed today that its run rate has more than tripled to $47 billion in the past three months.

Anthropic’s Series H funding round was led by Altimeter Capital, Dragoneer, Greenoaks and Sequoia Capital. The company listed Capital Group, Coatue, D1 Capital Partners, GIC, ICONIQ, and XN as co-lead investors. They were joined by more than a dozen others, including Micron Inc., SK hynix Inc. and Samsung Electronics Co.

The $65 billion round includes about $15 billion in previously announced commitments from cloud providers. Amazon Web Services Inc. is investing $5 billion. Last month, AWS agreed to provide Anthropic with up to 5 gigawatts of new computing capacity.

The LLM developer use the proceeds from its funding round to buy more infrastructure. Anthropic also plans to invest in AI safety initiatives and partnerships.

Image: Anthropic