Cognition raises $2B at $48B and puts GPT-6 Astra to work inside Devin, as Latent Space maps agent setup and answer-engine bias

Cognition, maker of Devin, raised $2 billion at a $48 billion valuation with $900 million annualized revenue, and a new OpenAI case study shows GPT-6 Astra automating testing across Devin's cloud agent, CLI, and desktop products. Separate Latent Space reports…

Cognition, maker of the autonomous software engineer Devin, raised $2 billion at a $48 billion valuation — four months after a $26 billion round — led by Andreessen Horowitz, Accel, Founders Fund, General Catalyst, and Avenir (TechCrunch, 2026-09-08). Annualized run-rate revenue grew from $492 million to $900 million since May, with $4-5 billion projected by end of 2026, while a leased Nvidia cluster could push total 2026 cash burn to $800 million. Customers include Mercedes-Benz, NASA, Goldman Sachs, and Citi, and the Scott Wu-founded startup is training its own model to cut reliance on OpenAI and Anthropic; TechCrunch frames the raise as investors believing AI coding is far from a winner-take-all market. Days later, an OpenAI customer case study (2026-09-11) detailed how Cognition uses GPT-6 Astra across Devin's cloud agent, CLI, and desktop products to automate testing: Devin tested the iPhone game Otter Run, returning a simulator recording plus a report of passed checks and untested areas, and customer-reported bug screenshots are fixed and verified automatically. Co-founder Walden Yan said Astra's ability to test and prove work functions as expected could reduce manual code review and raise shipping velocity over time. The two Cognition reports are complementary rather than conflicting: TechCrunch's note that Cognition is training its own model sits alongside — not against — the documented, ongoing use of OpenAI's GPT-6 Astra in its products. Latent Space's five-day Grok Bot review (2026-09-05) found browser-based sign-in connects X, Freshdesk, and Google Calendar with no MCP server configs or API keys, contrasting Grok Bot's managed 'agent computer' and personified, composable 'Bots' arranged in 'group chats' — an Agentic Engineer Bot routed tasks across Claude Code, Codex, and Grok Build CLI, directing frontend work to Claude Code and debugging to Codex — with OpenClaw 2.0's user-owned Gateway, which reuses Claude Code or Codex logins, ships a native Codex runtime, and narrows the setup gap via Quick Start and a graphical interface. The reviewer used a Cursor Pro+ account. Latent Space's AEO tracker (2026-09-07) 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 showed a universally dominant primary choice, models showed soft biases favoring their own lab's products (e.g., Claude recommending Claude Code), and…

• Cognition raised $2 billion at a $48 billion valuation, four months after a $26 billion round; led by Andreessen Horowitz, Accel, Founders Fund, General Catalyst, and Avenir (TechCrunch, 2026-09-08).
• Cognition's annualized run-rate revenue grew from $492 million to $900 million since May, with $4-5 billion projected by end of 2026; a leased Nvidia cluster could push total 2026 cash burn to $800 million (TechCrunch, 2026-09-08).
• Cognition's customers include Mercedes-Benz, NASA, Goldman Sachs, and Citi; the Scott Wu-founded startup is training its own model to cut reliance on OpenAI and Anthropic, and TechCrunch frames the round as investors seeing AI coding as…
• An OpenAI case study (2026-09-11) says Cognition uses GPT-6 Astra across Devin's cloud agent, CLI, and desktop products to automate testing and provide evidence for code review.
• In one example, Devin used GPT-6 Astra to test the iPhone game Otter Run, returning a simulator recording plus a report of passed checks and untested areas; customer-reported bug screenshots are fixed and verified automatically with…
• Cognition co-founder Walden Yan says Astra's ability to test and prove work functions as expected could reduce manual code review, with expectations of less manual code examination and higher shipping velocity over time (OpenAI News,…
• A five-day Latent Space review (2026-09-05) found Grok Bot's browser-based sign-in connects X, Freshdesk, and Google Calendar with no MCP server configs or API keys, contrasting Grok Bot's managed 'agent computer' with OpenClaw 2.0's…
• OpenClaw 2.0 supports reusing Claude Code or Codex logins, ships a native Codex runtime, and narrows the setup gap with Quick Start and a graphical interface (Latent Space, 2026-09-05).