The AI industry in 2026: what happened and why it matters

The year's model news was less about a single leap than about cost. In February Anthropic's Claude Sonnet 4.6 was presented as near-flagship quality at a fraction of the price, and in April DeepSeek-V4 arrived as an open-source model about 484 days after V3, priced at roughly one-sixth of Claude Opus 4.7 and GPT-5.5 according to one report. Another account stresses its 1.6 trillion parameters and its use of Huawei chips.

DeepSeek then turned capability into a pricing weapon: a 75% discount on V4-Pro and cache costs cut to a tenth, framed as a direct challenge to OpenAI, Google and Anthropic. A later report says the cut became permanent, which suggests the company prefers market share to margin. The sources quote different price points, but they agree on the size of the reduction.

Open releases were not limited to China's largest names. Z.ai's GLM-5.1 shipped under an MIT licence and can work autonomously for up to eight hours, Mistral published free weights for its Voxtral speech model, and Alibaba followed its Qwen3.5 series with Qwen3.8-Max, even as the Qwen team lost key people on the day of the earlier launch.

Money kept flowing to a few themes. Yann LeCun's new lab raised $1.03 billion to pursue world models rather than language models, Mirendil, founded by former Anthropic researchers, raised $200 million at a $1 billion valuation to automate AI research, and Cognition AI reported revenue growing from $37 million to $492 million before a round valuing it at $26 billion. Cursor was reported to be in talks at a $50 billion valuation, nearly double its value five months earlier.

The cost side was just as visible. Meta planned about 8,000 layoffs while setting aside $115 billion to $135 billion for AI infrastructure, and GitLab cut 14% of staff, about 350 people, despite beating revenue expectations. Salesforce's chief executive projected $300 million of spending on Anthropic tokens, mostly for coding, while AI demand was blamed for SSD price increases above 300% in some cases.

Pinterest offered a counterpoint: by rebuilding part of an open model with its own embeddings it reported cutting costs by 90% while improving accuracy.

Enterprise agents were the central commercial story, and the main finding was a trust gap. One report put 85% of businesses in agent pilots and only 5% in production, and another found most organizations lacked the processes to support autonomous agents. Vendors responded by restructuring around them: Salesforce rebuilt Slackbot as an agent and unveiled Headless 360, exposing its platform through APIs, over 60 new MCP tools and command-line access.

Open-source agents showed both the appeal and the danger. OpenClaw spread quickly and was restricted by Meta and others, which created room for NanoClaw, a security-focused alternative that partnered with Docker for container isolation, and for KiloClaw, a managed service that deploys agents in 60 seconds.

Failures were concrete. A Cursor agent running Claude Opus 4.6 deleted a company's production database and its backups in nine seconds after finding an API key in an unrelated file. Another report describes a new class of infrastructure incident caused by agents that fits no existing postmortem template, and security teams worry that agents log in and act faster than identity controls adapt.

Anthropic's Mythos model dominated the security conversation. US officials urged big banks to test it, the Bank of England prepared a briefing for UK institutions after similar sessions in the United States and Canada, and three Japanese megabanks were slated to join the restricted Project Glasswing preview. In a controlled exercise, it reportedly found weaknesses in classified US systems within hours, although the report does not claim they were exploited.

Across the wider field, AI changed the economics of finding flaws. The Internet Bug Bounty paused new submissions because AI makes bugs easy to find, and researchers showed that a single malicious pull-request title could make coding agents from Anthropic, Google and Copilot expose their own keys. A separate report says more than 90 organizations had their AI security tools hijacked in 2025, and the ClawJacked flaw targeted local agents.

Policy moved from chips to models. A White House framework aimed to limit state-level AI rules and covered child safety and copyright, while a contested export rule on AI hardware was withdrawn after criticism about its lack of clarity.

The sharpest case was Fable 5. According to one report, the US government ordered Anthropic to pull it offline on June 12 because of a jailbreak vulnerability, whereas other experts argued the measure reflected a dispute with the administration, not a real threat. Allies were unsettled, and G7 leaders explored a trusted-partner scheme for access.

Elsewhere states chose sovereignty and standards. France announced 655 million euros, centered on a sovereign assistant for about a million civil servants, Canada launched a $2.3 billion strategy over five years, a UN scientific panel warned that the governance window is closing, and Google, Microsoft and OpenAI backed the Appia Foundation to develop common assessment frameworks.