The Week in AI: April 13-19, 2026 (Week 16)

News of the week: April 13-19, 2026 (Week 16)

The week belonged mainly to Anthropic, which released the Claude Opus 4.7 model, the Claude Design tool and a redesigned Claude Code app, and which, according to Bloomberg, is receiving investor offers valuing it at around 800 billion dollars. OpenAI responded with specialized models for cybersecurity and the life sciences and significantly expanded its Codex agent. Google brought a Gemini app for Mac and a new voice model, while Alibaba released the open Qwen3.6 model. In Washington, the debate centered on how to give federal agencies access to the Claude Mythos model, and in business, TSMC’s record profits and AI-justified layoffs at Snap made waves.

New Models

Anthropic Releases Claude Opus 4.7

On Thursday, April 16, Anthropic launched its new flagship model, Claude Opus 4.7, the successor to version 4.6. The improvements mainly concern programming, long “agentic” tasks (where the model plans and carries out multiple steps on its own) and working with higher-resolution images (The Next Web).

The price stayed the same as its predecessor’s: 5 dollars per million input and 25 dollars per million output tokens (tokens are the chunks of text used for billing). The model now has built-in safeguards designed to detect and block risky use in the area of cyberattacks (Campus Technology).

Anthropic itself admits that Opus 4.7 is less capable than Claude Mythos Preview. Due to safety concerns, the company has made that model available only to a limited circle of partners as part of Project Glasswing.

OpenAI: GPT-5.4-Cyber for Defenders

On April 14, OpenAI introduced GPT-5.4-Cyber, a modified version of GPT-5.4 for defending computer systems. It has fewer restrictions than the standard model and can, for example, perform reverse engineering of binary files, meaning it can analyze a program without access to its source code (SiliconANGLE).

Access is limited to vetted security professionals in the top tier of the Trusted Access for Cyber program, which OpenAI is simultaneously expanding to thousands of individuals and organizations. The media interpret the move as a response to the Mythos model from rival Anthropic (The Next Web).

GPT-Rosalind for Drug Discovery

On April 16, OpenAI came out with GPT-Rosalind, the first in a planned series of models for the life sciences. It helps summarize scientific literature, propose hypotheses and experiments, and analyze data from chemistry, protein engineering and genomics (The Next Web).

For now, the model is in a trial phase only for selected customers in the US; partners include Amgen, Moderna and Novo Nordisk. Access is also restricted because a model trained on biology could be misused to design dangerous pathogens (pharmaphorum).

Alibaba Releases Qwen3.6-35B-A3B

On April 16, the Tongyi lab of Alibaba published the open model Qwen3.6-35B-A3B under the Apache 2.0 license. It uses a mixture-of-experts architecture: the model has 35 billion parameters but activates only about 3 billion of them for each query, making it far more efficient (CnTechPost).

According to Alibaba, it matches larger open models in programming, such as Google’s Gemma 4. The model weights can be downloaded from Hugging Face and ModelScope, so it can also be run on your own hardware (Alibaba Cloud).

Other Models

April 14 – Nvidia released Ising, a family of open models for quantum computers. They help calibrate quantum processors and correct the errors that arise in them (Nvidia).
April 15 – Google launched the voice model Gemini 3.1 Flash TTS for text-to-speech, with finer control over tone and expression. It is available as a preview in the Gemini API, AI Studio and Vertex AI (Google).

Products and Agents

Claude Design: Designs and Presentations from a Chat

On April 17, Anthropic launched the Claude Design tool under its Anthropic Labs division. Users describe what they need, such as a website landing page, a presentation or a flyer, and then refine the result in conversation (Unite.AI).

The tool runs on the Opus 4.7 model and is available to Pro, Max, Team and Enterprise subscribers. Output can be exported to Canva, PDF, PowerPoint or HTML, and a finished design can be handed off to Claude Code for implementation (iPhone in Canada).

Redesigned Claude Code and Routines

Already on April 14, Anthropic released a new version of the Claude Code desktop app for Mac and Windows. It can run several parallel sessions in a single window, switch between them in a sidebar, and introduces a feature called Routines (VentureBeat).

Routines are scheduled automated tasks that run on Anthropic’s servers, so the user does not need to keep their computer on. They can be triggered on a schedule, via the API or by an event on GitHub.

Codex “for (almost) everything”

On April 16, OpenAI expanded its Codex desktop app far beyond programming. The headline feature is computer control: the agent sees the screen, clicks and types in any app on macOS without disturbing the user in other windows (Business Today).

Codex also gained a built-in browser, image generation, memory across conversations and more than 90 new plugins. Computer control is Mac-only for now and will arrive in the EU and the UK later.

Gemini on the Mac

On April 15, Google released a native Gemini app for macOS 15 and later. It opens with the Option + Space shortcut, users can “show” it another app’s window as context, and it is connected to Google Drive, Photos and NotebookLM (TechRadar, Google).

Agents in Enterprise Software

April 13 – According to TechCrunch, Microsoft is preparing yet another autonomous agent for Microsoft 365 Copilot, modeled on the popular open source project OpenClaw (TechCrunch).
April 15 – Adobe introduced Firefly AI Assistant: the user describes the desired result and the “creative agent” carries out the steps across Adobe apps on its own. The company also wants to offer it on third-party platforms, including Claude (Adobe).
April 15 – At its TDX conference, Salesforce announced Headless 360. The entire platform is now available via APIs, MCP servers (a standard for connecting AI agents to tools) and the command line, so agents can operate it without a browser (VentureBeat).

Business and Investment

Anthropic Attracts Investors at an 800 Billion Dollar Valuation

According to Bloomberg on April 14, Anthropic is receiving investor offers that would value the company at roughly 800 billion dollars. That is more than double the February round, in which the company raised 30 billion dollars at a 350 billion valuation (Bloomberg).

The interest is linked to rapid revenue growth: its annualized run rate reportedly stands at about 30 billion dollars. The company has not yet accepted any offer (The Next Web).

Microsoft Takes Over Norway’s Stargate

On April 14, Bloomberg reported that Microsoft will lease capacity at a data center in Narvik, Norway, originally intended for OpenAI’s Stargate project. The 230 MW site is being built by Nscale, and Microsoft plans to install 30,000 Nvidia Vera Rubin chips there (Bloomberg).

It marks another retreat by OpenAI from its own European data centers, following the pause of the Stargate UK project. At the same time, OpenAI opened a new office in London on April 13 (TechRepublic, CNBC).

TSMC Reports Record Profit

On April 16, Taiwanese chipmaker TSMC announced that its first-quarter net profit rose 58% to a record 572.5 billion Taiwan dollars, or about 18 billion US dollars. Revenue grew by roughly 35%, and the company raised its full-year outlook thanks to demand for artificial intelligence chips (DigiTimes).

More Business News

April 13 – OpenAI bought the startup Hiro, which offered a “personal AI CFO”. The deal was mainly about acquiring the team; the price was not disclosed and the Hiro service is shutting down (TechCrunch).
April 13 – Vercel CEO Guillermo Rauch signaled readiness for an IPO, as revenue is being driven by AI coding agents (TechCrunch).
April 15 – Snap, the owner of Snapchat, will lay off about 1,000 people, or 16% of its workforce. CEO Evan Spiegel attributed the move to advances in AI, which reportedly writes more than 65% of new code, and the company expects savings of over 500 million dollars a year (CNBC).
April 15 – Shoemaker Allbirds sold its footwear business, will rename itself NewBird AI and will rent out GPUs. Its shares soared by several hundred percent during the day and then fell sharply the next day (The Next Web).

Regulation, Politics and Law

The White House Wants to Give Agencies Access to Mythos

According to Bloomberg on April 16, the Office of Management and Budget (OMB) is preparing rules that would allow federal agencies to use the Claude Mythos model. Federal CIO Gregory Barbaccia wrote to departments that safeguards are being put in place because the model can find thousands of software bugs (Bloomberg Law).

The move is notable because the Pentagon designated Anthropic a supply chain risk in March, and the company is in a legal dispute with the government. On Friday, April 17, Anthropic CEO Dario Amodei met with White House Chief of Staff Susie Wiles; the White House described the meeting as productive (Fox News).

Research

Stanford AI Index 2026

On April 13, Stanford’s HAI institute published its annual AI Index report mapping the state of the field. One of the key findings is the growing gap between how AI is perceived by people in the industry and by the general public (Stanford HAI, TechCrunch).

AI Agents Outperform Anthropic Researchers

On April 14, Anthropic published a study on automated alignment researchers. Nine agents running in parallel on the Claude Opus 4.6 model tackled the “weak-to-strong supervision” problem, that is, how a weaker model can effectively teach a stronger one (Anthropic).

Two human researchers closed about 23% of the performance gap in seven days, while the agents closed around 97% in a further five days. The cost came to roughly 22 dollars per hour of work per agent; commentators note, however, that the results may not transfer well to larger models (eWeek).

April 13 – Quanta magazine described how AI is beginning to contribute to new results in mathematics (Quanta Magazine).

Security and Incidents

April 14 – Fortune described growing user dissatisfaction with the quality of Claude‘s responses and criticism of Anthropic for a lack of transparency during a compute shortage (Fortune).
April 14-16 – Both major players are responding to the risks of models capable of finding vulnerabilities: OpenAI is releasing GPT-5.4-Cyber only to vetted defenders, and Opus 4.7 received filters against misuse for cyberattacks (SiliconANGLE, Campus Technology).