Cognition hits $48B on $900M ARR and puts GPT-6 Astra to work in Devin, as Latent Space's AEO tracker exposes frontier-model recommendation bias

Cognition raised $2B at a $48B valuation led by a16z, Accel, and Founders Fund as annualized revenue reached $900M; OpenAI detailed Devin's use of GPT-6 Astra to test its own work; and Latent Space's new AEO tracker found frontier models softly biased toward…

Three reports from Sept 7-11, 2026 cover the fast-moving AI agent market. Latent Space (Sept 7) launched an Answer Engine Optimization tracker running 6 prompt variations across 7 search-enabled frontier models over 161 product categories, scoring first choices, alternatives, mentions, and anti-recommendations; it found 28 categories with a universally dominant primary choice and soft self-lab biases (e.g., Claude recommending Claude Code), while newer Anthropic generations (Sol→Astra, Opus→Fable) consult fewer sources and change answers less under paraphrased questions; Gemini, GLM, and DeepSeek pairs were excluded due to errors and rate limits. TechCrunch (Sept 8) reported Cognition raised $2B at a $48B valuation, four months after a $26B round, led by Andreessen Horowitz, Accel, Founders Fund, General Catalyst, and Avenir; run-rate revenue grew from $492M to $900M since May, with $4-5B projected by end of 2026 and a leased Nvidia cluster possibly pushing 2026 burn to $800M; customers include Mercedes-Benz, NASA, Goldman Sachs, and Citi, and Cognition is training its own model to cut reliance on OpenAI and Anthropic. OpenAI's case study (Sept 11) details GPT-6 Astra integration across Devin's cloud agent, CLI, and desktop products, including testing the iPhone game Otter Run with a simulator recording and coverage report and fixing screenshot-reported bugs with verification; co-founder Walden Yan says Astra's proof-of-work could reduce manual code review. Source note: 'Astra' refers to Anthropic's Sol→Astra generation in the Latent Space report and to OpenAI's GPT-6 Astra in the OpenAI case study; the reports do not reconcile the shared name.

• Latent Space's AEO tracker ran 6 prompt variations across 7 search-enabled frontier models over 161 product categories, scoring first choices, alternatives, mentions, and anti-recommendations.
• 28 of 161 categories show a universally dominant primary choice across surveyed frontier models.
• Models show soft bias toward their own lab's products, e.g., Claude recommending Claude Code.
• Newer Anthropic generations (Sol→Astra, Opus→Fable) consult fewer sources and are less likely to change answers under paraphrased questions; Gemini, GLM, and DeepSeek pairs were excluded due to errors and rate limits.
• Cognition raised $2B at a $48B valuation, four months after a $26B round, led by Andreessen Horowitz, Accel, Founders Fund, General Catalyst, and Avenir.
• Cognition's annualized run-rate revenue grew from $492M to $900M since May; $4-5B is projected by end of 2026.
• A leased Nvidia cluster could push Cognition's total 2026 cash burn to $800M.
• Cognition, founded by Scott Wu, counts Mercedes-Benz, NASA, Goldman Sachs, and Citi as customers and is training its own model to cut reliance on OpenAI and Anthropic.