State of Frontier AI

Release cadence by lab

Release cadence by country

Release cadence by access

• GLM-5.2 / Kimi K2.7 Code / DeepSeek V4 - top-tier agentic coding (long-horizon tasks, tool-call accuracy). 1T-class MoE; need a server.
• Qwen3-Coder (32B) / Devstral Small - the practical local picks; strong coding on a single GPU.

• DeepSeek V4 / Qwen 3.6 / MiniMax M3 - frontier-class quality, MoE cost advantage when self-hosting at scale.
• Gemma 4 (26B-A4B / 31B) / Mistral Small / Llama - efficient, permissive licensing, easy local deployment.

Dense models up to ~32B and small MoE at Q4: Qwen3-Coder 32B, Gemma 4 31B, DeepSeek 32B distill. A capable local coding copilot.

70B dense and mid-size MoE; 100B+ MoE on a high-memory Mac. Small-team self-hosting via vLLM tensor parallelism.

The 1T-class MoE flagships (DeepSeek V4, Kimi K2.7, GLM-5.2) at INT4/FP8 for production batch throughput. e.g. ~630 GB for a 1T model at INT4.

Anthropic passed OpenAI in run-rate and filed to IPO near ~$47B annualized; OpenAI targets $30B full-year. Google's AI revenue is bundled into Cloud / Workspace and not cleanly isolable; xAI is off a smaller base.

OpenAI's ChatGPT ad pilot launched Feb 2026 (~$100M annualized, ~$60 CPM, $200K minimum). Google now shows ads alongside ~25% of AI Overviews plus shopping ads in AI Mode. Perplexity dropped ads entirely. The AI-ad market is estimated at $15-25B, growing 35-50% a year.

Agentic commerce lets agents discover, compare, and buy inside the chat. OpenAI's Agentic Commerce Protocol powers ChatGPT Instant Checkout (Stripe-settled); Google's Universal Commerce Protocol pairs with shopping in AI Mode. Merchants expose product feeds; the agent handles the buy.

Both leaders filed to go public in mid-2026 (xAI + SpaceX reportedly combined ~$1.25T). Enterprise is where the margins are - Anthropic wins roughly 70% of head-to-head enterprise deals against OpenAI - while consumer subscriptions and ads drive volume.

A model whose trained weights are publicly downloadable, so anyone can run, fine-tune, or self-host it (DeepSeek, Qwen, GLM). It does not always mean the training data is open.

A model offered only through an API or product, with weights not released (GPT, Gemini, Claude).

An architecture where only a subset of parameters ("experts") activate per token, so a model can hold trillions of total parameters but compute like a much smaller one.

A model that spends extra compute "thinking" (chain-of-thought) before answering, trading latency for accuracy on hard problems.

The maximum amount of text, in tokens, a model can consider at once - including the prompt and its own output.

Compressing a model's weights to fewer bits (e.g. 4-bit) to cut memory use with modest quality loss - key to running large models on limited hardware.

An AI system that plans and takes multi-step actions toward a goal using tools - combining autonomy, reasoning, tool use, and memory.

An open standard for connecting AI models to external tools, data, and services - the common interface for agent integrations.

Letting AI agents discover, compare, and buy products on a user's behalf inside the chat (e.g. ChatGPT Instant Checkout).

Answer Engine Optimization / Generative Engine Optimization: making content easy for AI search engines to surface and cite.