A Week in AI: May 4-10, 2026 (Week 19)

News for the week: May 4-10, 2026 (Week 19)

The week was mostly about agents and compute. OpenAI swapped the default model in ChatGPT for GPT-5.5 Instant and released GPT-5.5-Cyber for security teams. At its Code with Claude conference, Anthropic showed agents that learn from their own work, and it leased the entire Colossus 1 supercomputer from SpaceX. Nvidia kept pouring billions into suppliers and data centers. The EU reached a tentative deal to delay parts of the AI Act, and Google, Microsoft and xAI will let the US government test their models before release.

New Models

GPT-5.5 Instant Is ChatGPT’s New Default Model

On May 5, OpenAI replaced the previous default model in ChatGPT, GPT-5.3 Instant, with the new GPT-5.5 Instant. It is the fast variant of the model, answering right away without extended “thinking”. It is rolling out to all users, including those on the free tier.

According to internal tests by OpenAI, the model makes up 52.5% fewer false claims (so-called hallucinations) on sensitive questions about medicine, law and finance. The model is better at drawing on earlier conversations, saved files and a connected Gmail account. Subscribers can still choose the older version for three more months (eWeek, Decrypt).

Google Releases Gemini 3.1 Flash-Lite

On May 7, Google released the final version of Gemini 3.1 Flash-Lite, the cheapest and fastest member of the Gemini 3.1 family. The preview version of the model will be retired on May 25. The week also brought new developer features in the Gemini API (Gemini API changelog):

May 4: webhooks, automatic notifications that a long-running task has finished, so apps no longer need to keep polling.
May 5: file search (File Search) now handles images and other media thanks to the gemini-embedding-2 model.

Products and Agents

Anthropic at Code with Claude: Agents That “Dream”

On May 6, Anthropic held its Code with Claude developer conference in San Francisco. The main theme was Claude Managed Agents, agents that Anthropic runs on its own infrastructure on behalf of customers. An agent is an AI that carries out multi-step tasks on its own, such as editing code or processing documents.

Dreaming: at scheduled intervals, the agent reviews its past work and stores the lessons learned in long-term memory. For now it is a research preview with limited access.
Outcomes: the requester describes what success looks like, and the agent checks and improves the result on its own.
Multi-agent orchestration: a lead agent splits the work among specialized subagents that work in parallel on shared files.

Details are summarized by the Claude blog and SiliconANGLE.

Claude for Banks and Insurers

On May 5, Anthropic introduced a package of “agents for financial services“. It consists of ten ready-made templates for Claude Cowork and Claude Code, new connectors to data sources, and add-ins for Excel, PowerPoint and Word. The agents are meant to automate analysis, report preparation and regulatory compliance checks (Anthropic, InvestmentNews).

Meta Is Preparing a Personal Agent Called Hatch

The Information reported on May 5 that Meta is developing a consumer agent code-named Hatch. It is supposed to handle tasks on its own, such as scheduling meetings, writing emails or shopping. Internal testing is expected to wrap up by the end of June.

The agent is eventually meant to run on Meta’s own Muse Spark model, but for now it is reportedly being tested with Claude models. Meta is also preparing a shopping agent for Instagram. The company did not comment on the reports (Engadget, Gizmodo).

Business and Investment

Anthropic Leases All of SpaceX’s Colossus 1

On May 6, Anthropic announced a deal with SpaceX, which has also owned Elon Musk’s xAI since February. Anthropic will use the full capacity of the Colossus 1 data center in Memphis: over 300 megawatts and more than 220,000 Nvidia GPUs. The capacity is expected to come online within a month.

Anthropic will use it mainly for higher usage limits for Claude Pro and Max subscribers. The companies also expressed interest in working together on orbital compute centers. The price was not disclosed (Anthropic, Unite.AI).

Nvidia Invests in Optics and Data Centers

Nvidia announced two big investments this week. It is quickly building a network of the companies that AI infrastructure depends on.

May 6: up to $3.2 billion in glass and optical fiber maker Corning. Corning will build three new factories in Texas and North Carolina and increase its production of optical interconnects for data centers tenfold. The deal will create more than 3,000 jobs (Corning, RCR Wireless).
May 7: the right to invest up to $2.1 billion in data center operator IREN (buying up to 30 million shares at $70 each). The deal includes a plan for up to 5 gigawatts of infrastructure based on Nvidia DSX designs and a five-year cloud contract worth roughly $3.4 billion (Bloomberg, The Star / Reuters).

AMD Benefits from Demand for AI Chips

On May 5, AMD reported its first-quarter results. Revenue rose 38% year over year to $10.25 billion, and the data center division hit a record $5.78 billion (+57%). For the second quarter, the company expects about $11.2 billion, more than analysts had forecast. In the second half of the year, it plans to start shipping chips to Meta and OpenAI (Zacks, TweakTown).

More Business News

May 8: Chinese model developer StepFun is reportedly closing a funding round of nearly $2.5 billion and preparing for an IPO in Hong Kong (CnTechPost, Yicai Global).
May 5: Digi Power X announced a 40-megawatt AI campus in Alabama whose tenant will be chipmaker Cerebras (SEC).

Regulation and Law

The EU Delays Part of the AI Act

On May 7, representatives of the member states and the European Parliament reached a tentative agreement on the Digital Omnibus on AI package, which simplifies the AI Act. Obligations for high-risk systems (such as AI used in hiring or lending) are pushed back by a year, to December 2027. For AI built into products, such as elevators or toys, the new deadline is August 2, 2028.

The deal also adds a ban on so-called nudify apps, which turn photos into fake nude images, and on AI that generates sexual content without consent. The package still needs formal approval from both Parliament and the member states (Euronews).

The US Government Will Test Models Before Release

On May 5, Google, Microsoft and xAI signed agreements with the government’s Center for AI Standards and Innovation (CAISI) at the US Department of Commerce. The center will examine their most powerful models before they are released, mainly for risks in cybersecurity and in biological and chemical weapons. OpenAI and Anthropic have had similar agreements since 2024 and have now renegotiated them (Euronews, SiliconANGLE).

Musk v. OpenAI: Brockman and Zilis Take the Stand

The trial in Oakland continued, in which Elon Musk is suing OpenAI over its shift from a nonprofit to a for-profit company. On May 4, OpenAI president Greg Brockman acknowledged that his stake in the company is worth close to $30 billion, even though he did not put any of his own money into it. He also testified that in 2017 Musk wanted full control of the company and talked about $80 billion for a city on Mars (Quartz, The Star / Reuters).

On May 6, former OpenAI board member Shivon Zilis testified. According to her testimony, Musk had considered having Tesla take over OpenAI (Courthouse News).

Apple to Pay $250 Million over Siri Promises

Early in the week, Apple agreed to a $250 million settlement in a class action lawsuit. Customers claimed they paid more for iPhones because of advertising for a smarter Siri and Apple Intelligence features that were not available at the time. The settlement covers buyers of the iPhone 16 and 15 Pro in the US and still needs court approval. Apple did not admit wrongdoing (Yahoo Finance / Reuters, MediaPost).

Security and Incidents

GPT-5.5-Cyber for Vetted Defenders

On May 7, OpenAI launched limited access to GPT-5.5-Cyber. It is not a smarter model, but a version that refuses security tasks less often, such as penetration tests (controlled attacks on your own systems). It is available only to verified teams in the Trusted Access for Cyber program. Starting June 1, individuals with access must use enhanced account security (IT Brief, Neowin).

TrustFall: One Keypress Is Enough

On May 7, security firm Adversa AI described a weakness called TrustFall in the coding agents Claude Code, Gemini CLI, Cursor CLI and Copilot CLI. When a developer downloads a malicious project and confirms the prompt asking whether they trust the folder, the tool may automatically run harmful code. And the default answer is “yes”.

According to the researchers, Anthropic considers this behavior intentional. The key takeaway is not to trust unknown repositories (Help Net Security).

Research

May 7: Google DeepMind summed up the first year of AlphaEvolve, an agent built on Gemini that discovers and improves algorithms on its own. According to DeepMind, it has helped with DNA analysis, power grid management, quantum computing and logistics (Ground News).
May 7: Microsoft published a study on how quickly AI is spreading around the world, “The state of global AI diffusion in 2026”. The author is the company’s chief data scientist Juan Lavista Ferres (Microsoft On the Issues).