Weekly Digest
April 6 – April 12, 2026 · 8 stories
The frontier model race is converging on a paradox: capability is surging while the economics get worse. Meta and Z.ai shipped competitive models this week from opposite directions — proprietary and open-source — while leaked financials showed OpenAI and Anthropic burning through cash faster than they can grow revenue. Meanwhile, the physical AI thesis is gaining real capital, with $2.6B flowing into robotics and hardware in a single week.
Meta debuted Muse Spark, a natively multimodal reasoning model and the first major release from Meta Superintelligence Labs. It's small, fast, and proprietary — a sharp departure from the open-source Llama strategy. Scores 52 on the Artificial Analysis Intelligence Index, behind only Gemini 3.1 Pro, GPT-5.4, and Claude Opus 4.6.
Zhipu AI (now Z.ai) released GLM-5.1, a 744B-parameter mixture-of-experts model under an MIT license that topped SWE-Bench Pro at 58.4, edging out GPT-5.4 and Claude Opus 4.6. It can run autonomous coding loops for up to eight hours — the first open-weight model to break into the top 3 on Code Arena.
Confidential documents show inference costs now exceed half of revenue at both companies. Anthropic has passed OpenAI in ARR ($30B vs $24B). OpenAI projects $121B in compute spending by 2028 and doesn't break even until the 2030s. The economics of frontier AI are getting harder, not easier.
OpenAI published a 13-page policy paper proposing public wealth funds, robot taxes, four-day workweeks, and massive grid investment to prepare for superintelligence. Critics called it regulatory capture dressed as policy vision. Either way, it signals OpenAI is staking out political positioning ahead of its IPO.
Just five months after GEN-0, Generalist AI shipped GEN-1 — a robotics foundation model trained on raw movement data that handles high-dexterity physical tasks. The $440M-valued startup is betting that general-purpose robotic intelligence follows the same scaling laws as language.
The Cerebras-backing VC firm closed a $1.3B fund targeting physical AI — robotics, manufacturing, and defense. Eclipse plans to incubate companies internally, not just write checks. The fund signals a capital shift from software-only AI toward real-world deployment at the hardware layer.
Poke, which lets you run AI agents through iMessage, SMS, and WhatsApp, raised $10M on top of a $15M seed at a $300M post-money valuation. The bet: most people won't adopt AI through chat apps or IDEs — they'll adopt it through the messaging interface they already live in.
Anthropic pushed back the release of Claude Mythos, its next-generation reasoning model rumored to be a major leap in autonomous vulnerability discovery and coding. The delay comes as the company navigates IPO preparations and growing scrutiny of frontier model safety claims.