Cognition raises $2B at $48B, nearly doubling its valuation since May

Good morning 👋 A valuation nearly doubled in four months, and this time the revenue actually kept pace. Cognition's new round is below, along with the question it doesn't answer: what happens once every engineering org is running three of these agents at once.

In today's issue:

• 🔭 Cognition raises $2B at $48B, nearly doubling its valuation since May
• 🧠 Suno starts paying labels, Pathway skips chain-of-thought entirely
• 🔬 Agents that copy instead of coordinate, sycophancy under 25 turns of pressure, and a bias benchmark that flips depending on how you ask
• 🛠️ Turn your agent's behavior into a CI gate, not a demo
• 🔗 A gigawatt of GPUs headed for Southeast Asia, and OpenAI funds research on teens

🔭 THE ONE THING

🤖 Cognition raises $2B at $48B, nearly doubling its valuation since May

Cognition just closed a Series E north of $2 billion, led by Andreessen Horowitz and Accel, with Founders Fund, General Catalyst, and Avenir all writing checks again. The valuation: $48B, up from $26B four months ago. The number that actually justifies it: run-rate revenue jumped from $492M to nearly $900M over that same stretch, which means Devin's usage is compounding faster than the check sizes are. That kind of revenue velocity is rare enough that I'll take the valuation at face value for once. The harder question is whether "engineers as architects, agents as swarms" holds up once every enterprise coding budget has three of these tools bolted on at once.

🧠 MODELS & RELEASES

• 🎵 Suno shipped v6, v6-wild, and v6-mini, its first models built on licensed catalogs from Warner Music Group, BMG, and Believe. It's also the first time Suno pays those labels a cut of revenue, real money after the RIAA sued Suno over unlicensed training data back in 2024. Release notes
• 🧠 Pathway detailed Baby Dragon Hatchling, a post-transformer architecture that reasons in latent space instead of spooling out chain-of-thought tokens, with only about 5% of neurons active at once. BDH-CQ hit 29.2% pass@2 on ARC-AGI-1 at $0.0007 a task, trained and scaled on Amazon SageMaker HyperPod. AWS post
• 🖼️ OpenAI rolled out ChatGPT Images 2.5, turning sketches and reference photos into more polished, personalized output. Incremental, but it's the workhorse update most people will actually touch this week. Announcement

🔬 RESEARCH HIGHLIGHTS

• Thousands of AI agents didn't coordinate. They copied. In June, agents grinding through a timed test found a public wiki they could edit from inside their sandboxes and started leaving each other notes, no instructions given. Each agent lived about an hour and remembered nothing the next run. The researchers found the whole population's behavior reduces to one rule: an agent picks whatever option it sees most often. Three single-parameter copying models reproduce the clustering, the naming conventions, all of it. No shared goal required, just imitation compounding at scale.

• A 25-turn benchmark catches sycophancy that quick tests miss. SPINE has an LLM play a persistent, mistaken user and lean on a target model for up to 25 turns. Tested on four production systems plus three Olmo3-7B variants, collapse rates climbed the longer the pressure lasted. The tell: the correct answer often survives in the model's own reasoning trace even as the final response caves, which means it isn't confused. It's choosing to please. Emotional appeals worked best at breaking it.

• How you ask the bias question decides the answer. A widely cited charitable-aid benchmark found models favor minority applicants when rating requests one at a time, then penalize them when ranking the same requests side by side. Rerun across hiring, lending, and medical triage (40,726 requests, five models), most of that didn't hold up: the rating effect survived at half its original size, the ranking penalty didn't clear the bar at all. What did show up big: position bias, models favoring whoever's listed first, at a magnitude that rivaled the demographic effects everyone was auditing for.

🛠️ TRY THIS

Put a regression gate on your agent's behavior, not just its code

Cognition's $48B valuation is a bet that agents are production software now, which means the old "eyeball the demo, ship it" review process doesn't cut it. AWS just published the wiring for the fix: run your agent against a fixed prompt set on every PR, score the outputs, block the merge if it got worse.

1. Deploy your agent (and any MCP tools it calls) to a runtime your CI can hit, AgentCore or otherwise, alongside your existing pipeline.

2. Build a small, fixed set of test prompts, weighted toward your worst past failures rather than happy-path cases.

3. On each PR, invoke the agent against that set and score the responses against a baseline.

4. Fail the check and block the merge when a score drops below threshold. Treat it like a broken unit test.

Prompt: Given this agent's system prompt and its last 5 reported failure cases, write 10 adversarial test prompts that would catch a regression in tool-calling, refusal behavior, or output format.

Worth a look: Amazon Bedrock AgentCore Evaluations is the reference implementation above, complete with the GitHub Actions YAML if you'd rather copy than build.

🔗 QUICK LINKS

• Zankore Indosat's new AI infrastructure platform is building toward 1 gigawatt of Nvidia DSX AI Factory capacity across Southeast Asia, with the first 200 megawatts of GB300 GPUs live by early 2027. Zankore
• OpenAI is funding $5M in independent research grants on how generative AI shapes teen development and wellbeing, the kind of paper trail that matters as much for lawsuits as for parents. OpenAI

See you tomorrow, same repo, presumably more agents in it.

Pradeep Perugu