One of the main use cases for LLMs has been software development. The 2026 Sonar State of Code Developer Survey report found that 42% of committed code was AI-generated, even though 96% of developers did not fully trust the code. A June 2026 study by New Relic found that “67% of surveyed technology leaders report that AI generates or significantly refactors between 51% and 75% of their organizations’ weekly code output”. Google claims 75% of its production code is AI-generated, while the figure is 30% at Microsoft.
However, coding from scratch is one thing: understanding an existing code base is quite another. Around 70% of enterprise IT budgets go on support and maintenance of existing systems. An estimated 800 billion lines of COBOL remain in production, and that is a language that was released in 1960. By one estimate, 43% of banking systems are still written in COBOL. Corporate systems have arcane database schemas and file structures, often with barely documented code, while much of a system’s history exists only in the memories of experienced support programmers.
One US AI startup is trying to tackle this thorny problem. Reflection AI has an AI agent called Asimov, released in July 2025, that ingests code bases, database schemas, code repository commits and even project notes and documentation. It uses these sources to build a knowledge graph of an existing enterprise’s systems. The approach is similar to Google’s Deep Research, which reconciles hundreds of sources of data to build reports on a given subject.
This sounds almost too good to be true, and it is hard to verify since Reflection has so far only worked with carefully selected customers and conducted private research on early deployments. Asimov is not yet generally available, and nor is the proprietary LLM behind it. Asimov is a self-hosted, secure Virtual Private Cloud (VPC) product, deployed directly within a customer’s own cloud environment on platforms like AWS, Azure, or Google Cloud to ensure enterprise data and code repositories remain completely private. Asimov is available only on application. Reflection AI officially partnered with the US Department of Energy to support its Genesis Mission, a federal scientific research initiative. Under this partnership, Reflection AI serves as the core AI model provider for the US National Laboratories.
Asimov, however, now looks less like Reflection’s ultimate destination than a proving ground for its underlying technology. The company says that the reinforcement-learning techniques it developed for autonomous coding are now being applied to more general agentic reasoning. The company’s ambitions go much further. It is preparing to publish an open-weight large language model that will compete with heavyweight models from China like DeepSeek, Kimi and Qwen.
The company has founders with a serious track record and has attracted a remarkable amount of funding. Chief Technology Officer Ioannis Antonoglou previously served as a founding engineer at Google LLC’s AI research unit DeepMind, and was part of the team that developed AlphaGo. Reflection AI co-founder and Chief Executive Misha Laskin worked alongside him at DeepMind. The credentials of the founders have attracted financial backing from top-tier VCs Sequoia and Lightspeed Venture Partners as well as funding from NVIDIA. There was a seed round of $25 million in March 2025 from investors including Eric Schmidt (former CEO of Google) and a concurrent Series A of $105 million. This was followed by a Series B funding round of $2 billion in October 2025, and a Series C round of $2.5 billion in April 2026. That Series C round was at a valuation of $25 billion, heady stuff for a company without a released commercial LLM product.
The astronomical valuation of a company without a public commercial product reflects not only the track record of the founders, but also the scale of the very real problem that Reflection promises to tackle. If its product genuinely can make sense of complex enterprise code bases and build useful knowledge graphs that translate into something that an LLM can successfully code against, then it may actually make some money. The AI developer tools market alone is estimated to be worth $15 billion. In addition, the underlying technology, which uses complex reinforcement learning (RL) to let AI agents manipulate software tools and reason through long documents, could in principle be applied to the knowledge work automation market, covering areas like financial modelling and corporate document creation, which could be much larger. There may also be potential to licence the LLM to enterprises and governments that are concerned with privacy and do not wish to use models hosted on public clouds.
All this is extremely speculative. There are no benchmarks, no public customers and no published model at this stage. It is perhaps a sign of just how frenzied the AI market is right now that a company with no public customers and no commercial product can be valued at $25 billion. To give a sense of perspective, there are only 40 publicly quoted software companies with a market cap that high. Reflection’s valuation puts it just below established companies like Verisign, Zoom and MongoDB. Zoom, for instance, has a market cap of $27 billion but also has revenues of just under $5 billion. Reflection AI has a lot of work to do to demonstrate that its huge valuation is justified. Reflection may indeed have found a way to unlock decades of accumulated enterprise software knowledge. At a valuation of $25 billion, however, investors are already pricing in rather more than a promising experiment.
