Anthropic Beats OpenAI, Meta Goes Closed, Musk Wants Altman Out
Anthropic Beats OpenAI, Meta Goes Closed, Musk Wants Altman Out
This Week in AI, April 7-13, 2026
A week of arrivals and escalations. Anthropic's revenue eclipsed OpenAI's for the first time, a new model was deemed too dangerous to ship publicly, and Meta finally showed its hand after a year of silence and billions spent. Meanwhile, Elon Musk moved to have Sam Altman removed from the company they co-founded, three weeks before trial begins.
Key Takeaways
- Anthropic is now the highest-revenue AI lab. $30B ARR vs OpenAI's $24B, driven by enterprise, not consumer. Over 1,000 companies now spend $1M+ annually on Claude. If you're building on Claude, you're on the market leader.
- Claude Mythos exists but you can't use it. It generated 181 working Firefox exploits vs Opus 4.6's 2. Anthropic is deploying it exclusively for defensive cybersecurity under Project Glasswing. This sets a precedent: frontier models may debut in locked, domain-specific programmes before or instead of general release.
- Meta went closed-source. Muse Spark, the first model from Alexandr Wang's Meta Superintelligence Labs, sits roughly in the midrange Claude/GPT tier. The Llama open-source era is effectively over at the frontier level.
- The Musk vs. Altman trial starts April 27. Musk is seeking removal of both Altman and Brockman, and $150B in damages to be paid to OpenAI's nonprofit arm. The outcome could force OpenAI to revert its structure or expose how much of its for-profit conversion was built on shaky legal ground.
- Google filled the open-source gap. With Meta going proprietary, Gemma 4 under Apache 2.0 is now the most capable fully permissive model available. Developers who built on Llama need to re-evaluate.
The Revenue Crown Changes Hands
Anthropic Hits $30B ARR, Passes OpenAI
Anthropic's annualised revenue run rate reached $30 billion this week, overtaking OpenAI's $24 billion. That gap barely existed four months ago when Anthropic was at $9B. The driver isn't ChatGPT-style consumer traffic; it's enterprise contracts, with the number of companies spending $1M+ annually on Claude doubling in under two months. Anthropic also locked in 3.5 gigawatts of next-generation TPU compute via Google and Broadcom. Bloomberg
Claude Managed Agents: The Enterprise Bet
The same week, Anthropic shipped Claude Managed Agents in public beta: a fully managed harness for running Claude as an autonomous agent with secure sandboxing, no infrastructure management required. Early adopters include Notion, Asana, Rakuten, Sentry, and Vibecode. It's the clearest signal yet that Anthropic's enterprise play is infrastructure-level, not just API access. Anthropic
The Model You Can't Have
Claude Mythos: Too Dangerous to Release
A mis-timed Anthropic blog post this week confirmed what had been rumoured: Claude Mythos is real, it's the most capable model Anthropic has built, and you won't be getting access to it. On SWE-bench Pro it jumped from 53.4% (Opus 4.6) to 77.8%. On Firefox exploit generation, Opus 4.6 produced 2 working exploits in hundreds of attempts. Mythos produced 181. It also found a 27-year-old bug in OpenBSD and a 16-year-old one in FFmpeg.
Instead of a public release, Anthropic launched Project Glasswing: 12 partner organisations get gated access to use Mythos to find vulnerabilities in critical software, backed by $100M in model credits and $4M in donations to open-source security orgs. The implication for the industry: when a model's offensive capability outpaces its safety case, the default may shift from "release with guardrails" to "don't release at all." Anthropic Red Team
Meta Finally Shows Its Hand
Muse Spark: The Closed-Source Pivot
Meta shipped Muse Spark this week, its first proprietary model, under Alexandr Wang's Meta Superintelligence Labs, nine months after Wang joined from Scale AI. The model sits somewhere between Sonnet 4.6 and Opus 4.6 in capability, with a Contemplating mode for complex reasoning that parallels Gemini Deep Think and GPT Pro. It rolls out across Facebook, Instagram, WhatsApp, Messenger, and Ray-Ban Meta AI glasses. CNBC
Open Source Gets the Scraps
Meta's accompanying strategy announcement made it explicit: the best models stay proprietary, open source gets the smaller variants. After abandoning the 2-trillion-parameter Behemoth project, Meta has pivoted to smaller, efficient, closed models for its own products. The Llama community, developers who built entire stacks on the assumption Meta would keep releasing competitive open weights, now face a decision point. Axios
Musk vs. Altman, Three Weeks Out
The Most Consequential AI Legal Case in Years
Elon Musk filed an amended complaint this week seeking the court-ordered removal of Sam Altman and Greg Brockman from OpenAI, $150B in damages to be paid to OpenAI's nonprofit arm (not to Musk personally), and a full reversion of OpenAI's structure to an actual nonprofit. Jury selection starts April 27 in Oakland. OpenAI responded by asking California and Delaware attorneys general to investigate Musk for anti-competitive behaviour, alleging he coordinated attacks on OpenAI via xAI, Tesla, and SpaceX. CNBC
The case matters beyond the drama: if Musk prevails on any structural argument, it could force OpenAI to unwind the for-profit conversion that underpins its $852B valuation and its $122B fundraise.
Open Source Fills the Gap
Google Gemma 4: Apache 2.0, No Strings
As Meta retreated from open source, Google stepped forward. Gemma 4 launched this week under Apache 2.0, fully permissive, commercial use included, no exceptions. Four model sizes from edge to 31B parameters, native multimodal, 256K context window, and benchmark-competitive with closed models at the midrange. It's the first Gemma with zero commercial restrictions. For developers evaluating Llama alternatives, the decision just got straightforward. Android Developers Blog
Quick Hits
- Eclipse raises $1.3B: The physical AI and robotics fund (largest in Eclipse's history) targets AI infrastructure, manufacturing, and defense. Bloomberg
- GPT-5.5 "Spud" pretraining complete: OpenAI's next flagship has finished pretraining. No release date announced. The Information
- OpenAI acquires TBPN: The Silicon Valley live business show becomes OpenAI's first media acquisition. TechCrunch
- Shield AI raises $1.5B: Series G at a $12.7B valuation, up 140% in one year, for autonomous military aircraft. TechCrunch
- Claude Haiku 3 deprecated April 19: Migrate to Claude Haiku 4.5. Anthropic docs
- Microsoft launches MAI models: Three new models (MAI-Transcribe-1, MAI-Voice-1, MAI-Image-2) now available in Azure Foundry, with MAI-Image-2 ranking top 3 on Arena.ai. Microsoft AI
- Q1 2026 VC hits $300B: All-time record for global venture investment in a single quarter, with AI driving the majority of deals. Crunchbase
This Week in AI is published every Monday by the Lablab team.
Software Developer