Six months of frontier AI, distilled.

In November 2024, Anthropic released the Model Context Protocol as an open standard for connecting AI systems to external tools. Eighteen months later, it's the de facto infrastructure layer for enterprise AI agent deployment. Sam Altman — who usually doesn't say nice things about competitors' standards — posted simply: "People love MCP and we are excited to add support across our products."

Anthropic donated MCP to the Linux Foundation's Agentic AI Foundation in December 2025, with OpenAI and Block as co-founders and AWS, Google, Microsoft, Cloudflare, and Bloomberg as supporting members. The 2026-07-28 spec — the largest revision since launch — ships a stateless core, MCP Apps, the Tasks extension, and a formal deprecation policy. Twelve months of runway before legacy versions retire.

Why it matters: protocol wars are usually won by the company with the best tooling, not the best paper. MCP won because Anthropic shipped working SDKs in two languages on day one, then let everyone — including competitors — build on it. The cost of switching protocols is now higher than the cost of staying. Network effect locked in.

A year ago, "AI agent" was a slide-deck word. In 2026, it's a product category. 40% of enterprise applications now embed task-specific AI agents (Gartner) — up from less than 5% in 2025. 80% of apps shipped or updated in Q1 2026 embed at least one agent. 31% of enterprises have at least one agent in production. The median enterprise in that 31% is running 4.7 agents.

Microsoft made MCP a first-class standard across Copilot Studio, GitHub, Azure AI Foundry, and Windows at Build 2026. Anthropic shipped Claude Cowork with persistent sessions across devices. OpenAI made computer-use a default capability in GPT-5.4. Google shipped Google Antigravity, an agentic development platform. Every major lab is now in the same fight: who builds the agent runtime everyone builds on?

Why it matters: the agents that get built in 2026 will define the SaaS market for the next decade. The protocol layer (MCP) is settled. The runtime layer is not. Microsoft has the most complete stack. Anthropic has the strongest dev mindshare. OpenAI has the most users. Google has the most data. NVIDIA has the chips all of them run on.

Three different cost curves broke in the same six months. NVIDIA's Vera Rubin platform delivers one-tenth the cost per token of Blackwell. MiniMax M3 launched at $0.30 per million input tokens — 8-12× cheaper than Claude Opus or GPT-5.5. DeepSeek V4-Pro charges $0.435 per million input tokens. The price of intelligence is no longer a market — it's an arms race.

Why it matters: when the cost of intelligence drops 10× in 18 months, the cost of not using AI rises. A customer-service workflow that cost $50/hour to run in 2024 now costs $0.50. The unit economics of every knowledge-work business get rewritten. Watch what happens to SaaS pricing over the next 12 months.

In early 2025, DeepSeek R1 shocked the world with what was possible on a constrained compute budget. Twelve months later, the constraint looks less binding. DeepSeek R2 (Feb 2026) hit 94.2% of GPT-5's MMLU using an estimated one-third the FLOPs. By April, V4-Pro shipped 1M context at $0.435/M input. In July, Moonshot's Kimi K3 launched as the largest open-source model ever released.

The US-China Economic and Security Review Commission put it bluntly in March: "China has opted to go all in on an open-source approach to AI... Permissive licensing, aggressive pricing, and an ecosystem that encourages collaboration are accelerating global uptake of Chinese AI."

China has roughly 400,000 H800-equivalent GPUs. The US has 1.2 million H100/H200. But if the efficiency gap is 3× — and the evidence says it is — the effective compute gap is much smaller. Liang Wenfeng, DeepSeek's founder, told investors in late July: "[We] prioritise AGI over profit, [and] are likely to keep top models open-source."

Why it matters: the export-control narrative assumes compute parity equals capability parity. The data says otherwise. If DeepSeek R3 (expected Q4 2026) closes to within 2% of frontier US models on reasoning, the case for export controls collapses on its own merits.

A year ago, NVIDIA was a GPU vendor. Today, it's an AI infrastructure company. GTC 2026 introduced the Vera Rubin platform — seven new chips, five rack-scale systems, with the Groq 3 LPU integrated. Jensen Huang raised the cumulative Blackwell + Vera Rubin purchase order pipeline to $1 trillion through 2027, double the $500 billion projection from a year ago.

The 200,000-GPU Colossus supercomputer that xAI built in Memphis in 122 days has become the world's largest single training cluster. It's also become a landlord. Anthropic signed a deal in May to rent all of Colossus 1's compute. In June, Google agreed to rent 110,000 GPUs at Colossus 2 — at $920 million per month, for three years. $920M/month from one customer is bigger than most model API businesses.

Why it matters: the AI industry's economic center of gravity has shifted from models to compute. The companies that own the chips and the data centers make money on every API call, regardless of which model is being called. Jensen Huang called the Vera Rubin + Groq LPX combination a $300B annual revenue opportunity. That number doesn't include the rental economy — that's on top.

At WWDC 2026, Apple did something almost no one expected: it gave away the on-device model for free. Every Apple Intelligence-capable iPhone ships with AFM 3 Core — a 3-billion-parameter model accessible via native Swift API. No per-token cost. No telemetry. No setup. The Foundation Models framework lets any developer call it with a few lines of Swift.

Microsoft made the same bet at Build 2026. The Windows AI Platform SDK — a system-level API exposing on-device AI inference to any Windows app — ships with Windows 11 24H2 in Q3 2026. Aion 1.0 (Instruct + Plan 14B) ships in-box. Phi-4 Mini and Phi-4 Multimodal are MIT-licensed on Hugging Face.

Why it matters: the cloud-only assumption is dead for inference at the edge. For privacy-sensitive work, low-latency interactions, and cost-sensitive deployments, on-device is now genuinely viable. Apple showed that giving away the model is the right move when your moat is the silicon and the OS.

A year ago, "multimodal" meant a model that could read an image. Today, it means a model that takes text, images, and video as input, generates text, and (in some cases) generates video with synchronized audio. The single-modality model is now the exception. Every flagship release in H1 2026 is multimodal-first.

• OpenAI — GPT-5, GPT-5.4, GPT-5.5 all accept images natively. GPT-Live-1 brings natural voice with simultaneous listen/speak.
• Anthropic — Claude Sonnet 5 and Opus 4.8 process images and documents. Computer use is the new modality.
• Google DeepMind — Gemini 3, 3.5, Omni Flash are all multimodal-first. Veo 3.1 generates video with audio. Lyria 3 generates music.
• Meta — Llama 4 Scout and Maverick process text + image from day one (early fusion, not bolted on).
• Microsoft — Phi-4 Multimodal handles text + image + audio. GitHub Copilot reads your screenshots.
• MiniMax — M3 accepts text + image + video input. Hailuo 2.3 generates 1080p video with native audio.
• Apple — AFM 3 Core accepts images in prompts. Vision framework tools are callable from the model.

Why it matters: text-only models are now a feature flag, not a product line. The agent layer assumes the model can see what the user sees. The cost of building vision-capable apps just dropped to zero for any developer on Apple silicon or Windows 11 24H2.

In April 2025, Meta released Llama 4 Scout and Maverick with open weights. A year later, the open-weight flagship at Meta — Llama 4 Behemoth — is effectively shelved. The successor, Muse Spark, is Meta's first closed-weight, API-only frontier model. No weights. No architecture paper. The era of "frontier open" at Meta is over.

Anthropic, the originator of MCP, runs its frontier models on rented xAI hardware. OpenAI, despite being the most valuable AI company in the world, ships its flagship on a closed API with retirements on a 30-day sunset. Apple doesn't have a frontier model at all — it partnered with Google to use Gemini for Siri's cloud reasoning, with cryptographic attestation so Apple and Google can't see your data.

Meanwhile, the Chinese AI labs doubled down on open weights. DeepSeek (MIT). Moonshot (Kimi K3 — largest open-source model ever, July 2026). MiniMax (M3, June 2026, weights within 10 days of launch). Z.AI (GLM-5.2). Alibaba (Qwen 3, Apache 2.0).

Why it matters: the frontier is now genuinely bifurcated. Closed + integrated (OpenAI, Anthropic, Google, Meta, Apple) vs open + cheap (DeepSeek, MiniMax, Moonshot, Alibaba, Mistral). For enterprises, this is a real choice: pay for the best, or self-host something that's within 5-10% of the best. Both paths are now defensible.