Legal AI startup Legora raises $550M at $8.8B, up from $5.6B

Good morning 👋 A legal AI startup just priced itself like a late-stage fintech, and the number tells you more about where the money's flowing than the product does.

In today's issue:

• 🔭 Legora raises $550M at $8.8B, nearly tripling its old valuation
• 🧠 Models & Releases: Claude Duet's proactive agent, OpenAI on Google Cloud, an open model that reads DNA
• 🔬 Research Highlights: PDF ingestion quietly breaking agents, a tougher SWE-bench, self-taught long-context RL
• 🏛️ Policy & Regulation: OpenAI's pitch to Washington on cyber defense
• 🛠️ Try This: audit your agents' access before someone else does it for you
• 🔗 Quick Links: DeepSeek's reported $15B raise, OpenAI buys into biopharma, NVIDIA on the power bottleneck

🔭 THE ONE THING

⚖️ Legal AI startup Legora raises $550M at $8.8B, up from $5.6B

Legora closed a $550M Series E led by Atreides Management, pushing its valuation to $8.8B. Fidelity, J.P. Morgan, and Capital Group came in as new backers alongside the usual venture crowd (Accel, Benchmark, ICONIQ), which tells you this round got priced for public-market money as much as VCs. The underlying numbers hold up: $200M+ ARR, 140,000+ legal professionals using it across 2,400 firms. That's not a thin model wrapper riding a hype cycle. It's a real wedge into legal work, and every incumbent legal-tech vendor just got a loud reminder of how fast the ground shifted under them.

🧠 MODELS & RELEASES

• 🤖 Anthropic put out a research preview of Claude Duet, a proactive agent app built on Opus 5.5 that watches your inbox and Slack, drafts the dashboard before you ask, and holds off on anything that touches money or sends a message until you say yes. US Pro and Max subscribers only, waitlist for now. link
• ☁️ OpenAI models and Codex just showed up in Google Cloud's Marketplace and Model Garden. If your billing and infra already live on GCP, you can now run frontier OpenAI workloads without opening a second account. link
• 🧬 Hugging Face, working with Zhongguancun Academy and the Telethon Institute, open-sourced Carbon-X: 14.8B parameters, Apache 2.0, and it reads or writes raw DNA sequences inside a normal chat turn. Runs on a single 80GB GPU and gets 66.7% average across five genomics benchmarks, closing in on proprietary models at a fraction of the cost. link

🔬 RESEARCH HIGHLIGHTS

• Ingestion is quietly lying to your agent. Feed an enterprise agent 76 real policy documents and strip formatting on the way in, the way most PDF-to-text pipelines do, and accuracy craters: micro F1 drops from 0.690 down to 0.150 with Docling and 0.075 on plain text. The culprit is color and highlighting, the two cues that moved the score most, and 32.8% of the documents tested leaned on them to make their point. Tested across Claude variants and GPT-5.6, the pattern held regardless of model. The authors' line is the one worth remembering: reliability gets won or lost at ingestion, not at reasoning. arXiv

• SWE-bench grades a fiction. Real coding sessions aren't a single tidy GitHub issue typed once; they're multi-turn, with requirements changing mid-conversation. SWE-DIVE replaces the clean prompt with a simulated user who adds tasks, shifts requirements, and keeps judging the agent's output turn by turn, orchestrated through a framework the authors call Harbor. If your agent's scores only look good against one-shot issues, this is the benchmark built to catch that. arXiv

• Long-context RL without the long-context data. Training a model to reason over huge contexts has needed huge, verified training examples, exactly the thing that's scarce and expensive to hand-label. LongZero has the model write its own compact question, answer, and supporting evidence first, then expands that evidence into a sprawling multi-document context afterward, no external corpus or teacher model involved. Trained with GRPO across two LLMs and five long-context benchmarks, it beat baselines built on real external data. Self-generated curricula, it turns out, don't need to start big to end up useful.

🏛️ POLICY & REGULATION

• OpenAI published a blueprint asking Washington to legislate, not just read white papers: clear legal responsibility for securing frontier models and agents, public money for training the people who patch after an attack, and new government bodies built to move at AI speed instead of hearing-cycle speed. Nothing here is law. It's OpenAI lobbying for rules it would also have to live under, which is worth remembering before you take the framing at face value. Blueprint

🛠️ TRY THIS

Audit who your agents are, before a legal AI firm with your data handles the audit for you

Legora just raised $550M at an $8.8B valuation building AI that reviews contracts and does legal diligence. That's real stakes attached to an agent's output now, not a demo. Most people still run their agents on one shared API key with no idea what it can touch or how fast they could kill it.

1. List every agent, script, or automation with standing access to a credential, an inbox, or a repo.

2. For each one, write down what it's actually authorized to do, not what you assume.

3. Swap any shared, long-lived key for a scoped token tied to one agent, one task.

4. Pick your riskiest agent and time how long it takes to revoke its access mid-task. If it's over a minute, fix that first.

Prompt: List every API key, service account, or credential my agents currently use. For each: what can it do, is that the minimum needed, and how would I revoke it in under 60 seconds if it misbehaved.

Worth a look

• AWS published a reference architecture for cryptographic agent identity: signed request envelopes, user-signed delegation, and real-time OCSP revocation via Keyfactor EJBCA. Heavier than step 3 above, but the blueprint to graduate to once an agent is doing something that matters. link
• NxcoreAI's EverRoom is a workspace that remembers an agent's projects, decisions, and sources across sessions. Useful for the diligence-style work Legora is selling, where losing context between sessions is how mistakes slip through. link

🔗 QUICK LINKS

• DeepSeek Bloomberg reports a roughly $15B raise ahead of a possible IPO, a number nobody else has confirmed yet. Bloomberg
• OpenAI is buying Orchestra Bio to run AI-powered operations inside biopharma R&D, betting drug development is now an agent problem. OpenAI
• DUOS Technologies landed a $500M+ Axe Compute contract for 55MW of AI-datacenter colocation space, per its 8-K. SEC filing
• OpenAI pushed ChatGPT Ads into 26 more countries; over 100,000 advertisers have now run on the platform. OpenAI
• OpenAI, PerturbAI and Scripps turned thousands of AI agents loose on data from millions of brain cells to surface research leads a human team would take years to find. OpenAI
• Cloudflare is using OpenAI's Daybreak to investigate its own vulnerabilities and draft fixes, with Codex handling the day-to-day code. OpenAI
• NVIDIA argues power, not chips, is the real AI bottleneck now, and its DSX platform treats the grid and the data center as one system, substation to rack. NVIDIA

See you tomorrow.

Pradeep Perugu