TL;DR: ElevenLabs raises $500M at $11B; Cerebras lands $1B at $23B; “Agentic Engineering” emerges as 2026’s meta

Major Highlights:

• Double decacorn day: ElevenLabs and Cerebras
  ElevenLabs closed a $500M Series D at an $11B valuation led by Sequoia, a16z, and ICONIQ, underscoring the commercial maturity of SOTA audio models. Hours later, Cerebras announced a $1B Series H at a $23B valuation led by Tiger Global—less than five months after a prior $8B mark—buoyed by its 750MW OpenAI deal reportedly worth $10B over three years. Capital is consolidating behind inference-scale hardware and media-generation platforms.
• Gemini 3 productization and cost curve compression
  Google is baking Gemini 3 into Chrome’s side panel with “Nano Banana” integrations and pushing game-based evaluations (Poker/Werewolf/Chess) via Kaggle’s Game Arena to stress-test soft skills like planning and communication. Sundar Pichai cites Gemini 3 as the fastest-adopted model yet; commentary highlights a 78% unit-serving cost drop across 2025 and a claimed 750M+ MAU for the Gemini app in Q4’25, putting it within range of ChatGPT’s reported MAU.
• Coding agents converge inside the IDE
  VS Code repositioned itself as the “home for coding agents,” introducing Agent Sessions for local/background/cloud agents, parallel subagents, integrated browser, Hooks, Claude.md, and request queuing. GitHub Copilot now lets Pro+/Enterprise users select Claude or OpenAI Codex agents per intent and run them asynchronously to clear backlogs—shifting from chat to “remote async agent” workflows. OpenAI reports 1M+ active Codex users and a shared “Codex harness” via a JSON-RPC App Server.
• Agentic Engineering becomes the new meta
  On the 1-year anniversary of “Vibe Coding,” Andrej elevates “Agentic Engineering” as the 2026 theme: skills, subagents, MCP Apps, and trace-first observability. LangChain’s deepagents formalize .agents/skills, context isolation, and durable execution; OpenAI’s MCP Apps spec aligns ChatGPT integrations. ServiceNow and Monte Carlo case studies emphasize supervisor architectures with hundreds of subagents.

Key Technical Details:

• Funding/valuation: ElevenLabs $500M Series D at $11B (Sequoia, a16z, ICONIQ). Cerebras $1B Series H at $23B (Tiger Global); prior valuation $8B (~5 months ago). Cerebras–OpenAI: 750MW, ~$10B/3 years.
• Google: Chrome side panel “running on Gemini 3”; 78% unit-serving cost reduction across 2025 (reported); Gemini app 750M+ MAU (Q4’25 claim); Kaggle Game Arena for model evaluation via games.
• IDE/agents: VS Code Agent Sessions, parallel subagents, integrated browser, Hooks, Claude.md, queueing. GitHub Copilot agent choice (Claude, Codex) for Pro+/Enterprise. Codex: 1M+ active users; JSON-RPC “Codex App Server” harness.
• MCP/skills: MCP Apps now portable into ChatGPT; “skills” encode domain procedure/knowledge; MCP connects external runtimes.
• Benchmarks: METR reports GPT‑5.2 (high reasoning effort) achieving a ~6.6-hour 50% time horizon on expanded software tasks (CI ~3h20m–17h30m); controversy around runtime reporting.

Community Response/Impact:

• Decacorns are “less rare,” but investors are rewarding hard-infra plus clear enterprise workflows (audio generation; wafer-scale inference).
• Friction: Some Codex users report CPU-only sandboxes; OpenAI DevRel says GPU processes work and requests repros.
• Growing consensus that traditional benchmarks are saturated; shift toward game-based and “economically useful work” evaluations.
• Agent communities (e.g., OpenClaw) are professionalizing with security, distribution, and workflow tooling.

First Principles Analysis:

• The funding signals a flight to scale: capital is backing compute (Cerebras) and media-native models (ElevenLabs) that monetize immediately and compress unit costs.
• Embedding models in browsers/IDEs binds AI to daily workflows, amplifying usage and data flywheels while normalizing agentic, asynchronous patterns.
• “Agentic Engineering” reframes LLM systems as orchestrations of skills, tools, and supervisors—making trace-driven observability the new debugging, and pushing evaluation toward long-horizon task completion rather than static leaderboards.