Anthropic just made the word "frontier" awkward for everyone. Claude Sonnet 5 dropped last week with scores that beat every competitor - including Anthropic's own Opus model - at mid-tier pricing. At the same time, AI companies are spending hundreds of millions trying to shape who writes the rules for this technology. The stakes don't get much higher.
🧠 Sonnet 5 Is the Best Coding Model Alive, and It Costs Sonnet Money
Anthropic released Claude Sonnet 5 on April 1st, and the benchmark numbers are hard to believe. It scored 92.4% on SWE-bench Verified - the industry's standard test for real-world coding tasks. That beats Anthropic's own Claude Opus 4.6 (80.8%) and absolutely buries GPT-5.4 (57.7%).
The pricing hasn't moved. Same $3/$15 per million tokens as Sonnet 4.6. The 2M token context window is now out of beta. Sonnet 5 also leads on PhD-level science questions (96.2% on GPQA Diamond) and abstract reasoning (84.7% on ARC-AGI-2).
The practical implication is significant: the gap between "flagship" and "mid-tier" models is collapsing. Developers who were waiting to use Opus-class reasoning can now get it at Sonnet prices. That changes cost calculations for anyone building AI-powered products.
🗳️ AI Companies Are Buying the Midterms - and They're Not All on the Same Side
The AI industry is flooding the 2026 midterm elections with money, and the split is more interesting than you'd expect. "Leading the Future," backed by OpenAI co-founder Greg Brockman, Andreessen Horowitz, and other tech heavyweights, has raised over $125 million to elect candidates who oppose state-level AI regulation.
Anthropic is funding its own PAC - "Public First Action" - on the opposite side, raising $75 million to support candidates who want to preserve states' rights to regulate AI. Meta has a third PAC spending $65 million. Innovation Council Action, tied to Trump advisors, says it'll spend at least $100 million.
This is the AI industry's version of the crypto lobbying playbook, and it's moving fast. What makes it unusual is the internal split: the biggest AI labs aren't voting as a bloc. The fight over who controls AI regulation - federal or state governments - is real, and the answer will shape everything that comes after.
🌍 Yann LeCun Left Meta to Build AI That Actually Understands the World
In March, Turing Award winner Yann LeCun officially left Meta and launched AMI Labs with a $1.03 billion seed round - the largest seed ever raised by a European company, backed by Nvidia, Bezos Expeditions, and a long list of tech luminaries including Eric Schmidt and Jim Breyer.
AMI's bet is that large language models are hitting a ceiling because they learn from text, not reality. World models - AI that builds internal representations of how the physical world works - are LeCun's proposed fix. It's the same argument he's been making publicly for years, now with a billion dollars behind it.
Whether world models pan out is genuinely uncertain. But LeCun's departure from Meta signals that the people closest to the frontier think there's a different path worth chasing. AMI will operate across Paris, New York, Montreal, and Singapore, drawing key talent from Meta and Google DeepMind.
🚀 OpenAI Is Minting $2B a Month and Moving Toward an IPO
OpenAI is now generating $2 billion in monthly revenue, and an IPO is looking increasingly real. The company is valued at $852 billion following its latest funding round - a number that would have seemed fictional two years ago.
It also acquired TBPN, the daily tech news show hosted by John Coogan and Jordi Hays that's interviewed Zuckerberg, Satya Nadella, and Sam Altman himself. That's an unusual move for an AI lab, and signals OpenAI is thinking about distribution and media presence, not just models.
The next model - GPT-5.5, internally called "Spud" - has finished pretraining and is expected to drop sometime in the next few weeks. OpenAI is running hard.
🔌 97 Million MCP Installs Later, Agentic AI Has Its Plumbing
Anthropic's Model Context Protocol crossed 97 million installs in March 2026. That number matters because MCP started as an experimental spec about 18 months ago - now it's the default way AI agents connect to tools, data sources, and services.
The Agentic AI Foundation, formed under the Linux Foundation, has pulled together MCP, OpenAI's AGENTS.md, and Block's Goose framework into a shared infrastructure layer. When competing labs start contributing to the same plumbing, that's a sign the ecosystem is maturing.
We're past the "will agents become real" question. The question now is which apps and workflows get rebuilt around them first.
Number of the week
92.4% - Claude Sonnet 5's score on SWE-bench Verified, the go-to benchmark for real coding ability. For context: the previous record was Anthropic's own Opus 4.6 at 80.8%. GPT-5.4 sits at 57.7%. This is the biggest single jump on a major benchmark in recent memory.
Keep an eye on
GPT-5.5 "Spud" - OpenAI has finished pretraining and a mid-to-late April launch looks likely. It could ship as GPT-5.5 or jump straight to GPT-6.
State AI bills - Georgia's legislature wraps up April 6th with three AI bills heading to the governor, including chatbot disclosure rules and a child safety bill. Idaho just passed four. The patchwork of state laws is arriving faster than any federal framework.
AMI Labs' first research - LeCun's team is hiring fast from Meta and DeepMind. Watch for the first technical papers to signal whether the world model thesis has legs.
You're reading AgenticBrief, the newsletter that brings you relevant AI signals without the noise. New issues every Monday and Thursday.