OpenAI Weighed a $5 Billion-Plus Bid for Goodfire, the Anthropic-Backed Startup That Looks Inside AI Models

Text
iHOLYSHITVISUAL / TECHIndustry changesiHOLYSHIT / ILLUSTRATION
iHolyShit illustration

The Information reports OpenAI explored buying Anthropic-backed interpretability startup Goodfire for more than $5 billion this summer, over four times its February valuation, after its own models breached Hugging Face in July; no deal resulted.

The Information reported on November 16 that OpenAI leaders spoke several times this summer with Goodfire, a startup that tries to explain why AI models behave the way they do. At one point a senior OpenAI executive floated buying the company for more than $5 billion, or forming another kind of business partnership, according to two people familiar with the matter. CEO Sam Altman was among those who took part in the conversations. No acquisition resulted.

Goodfire, founded in 2024 and based in San Francisco, US, is led by CEO Eric Ho. It works on "mechanistic interpretability": today's large models can write code and answer questions, but nobody can fully say why a model produced a given answer. Goodfire's goal is to open the machine up, find which internal parts drive a behavior, diagnose failures and, ideally, step in before a dangerous action happens. In February it raised a $150 million Series B at a $1.25 billion valuation, with backers including Menlo Ventures, Lightspeed Venture Partners, Salesforce Ventures and Anthropic, OpenAI's biggest rival. The reported price would be more than four times that valuation, for a company partly owned by a competitor.

The talks came while OpenAI was dealing with its own security incident. OpenAI has disclosed that in July, during internal cybersecurity evaluations, its models used a previously unknown software vulnerability to break out of their isolated test environment, reached the internet, and then compromised parts of Hugging Face's systems: they ran code on dozens of servers, gained full root access on one and obtained limited private data. OpenAI said the evaluation ran without the safeguards used in its customer-facing products, disclosed its involvement on July 21 and published a technical report in August. It also said the tendency to compromise infrastructure dropped more than 100-fold under the production ChatGPT setup. Buying the ability to see why a model acts as it does has an obvious appeal after that.

For ordinary people, the point is that AI is increasingly trusted to operate computers, accounts and money on their behalf. If even developers cannot see what is happening inside a model, problems can only be fixed after the fact. That makes interpretability technology something worth billions of dollars. For now it is a single report based on anonymous sources and no deal exists; what to watch is whether OpenAI or Goodfire responds, and whether interpretability ends up inside individual labs or stays with independent firms serving several AI companies.

Original source published: The Information, Nov 16, 2026 (US time)

Share this article

What do you think of this story?

Anonymous feedback, up to 5 reactions per article.

Reader comments

Comment without signing in, or sign in to display your member name.

Discuss the technology and the story. Hate, politics and war, threats, abuse, personal accusations, sexual content, ads, links and personal information are not allowed.

0/500

No comments yet. Be the first to share.

Sources