The Week in AI: April 20–26, 2026 (Week 17)

News of the week: April 20–26, 2026 (Week 17)

Money and compute dominated the week: Anthropic secured commitments of up to $25 billion from Amazon and up to $40 billion from Google, and SpaceX locked in an option to buy Cursor for $60 billion. OpenAI released GPT-5.5, new image generation, and team agents in ChatGPT. China delivered DeepSeek V4, Kimi K2.6, and Qwen3.6-Max. At its Cloud Next conference, Google unveiled the eighth generation of its TPU chips. The White House pushed back against the copying of American models, and Florida opened a criminal investigation into OpenAI.

New Models

OpenAI Releases GPT-5.5

On April 23, OpenAI introduced GPT-5.5, just six weeks after GPT-5.4. The company describes it as its most capable model for agentic work, meaning tasks where the AI independently carries out a longer series of steps, such as writing and testing code or operating a computer (DataNorth).

The model is available in ChatGPT and in the Codex coding tool for Plus, Pro, Business, and Enterprise subscribers. The more powerful GPT-5.5 Pro variant is reserved for higher tiers. Both models reached the developer API on April 24 with a context window of 1 million tokens, enough to process text the length of several books at once (OpenAI API Changelog).

Performance: according to OpenAI’s figures, GPT-5.5 scores 82.7% on Terminal-Bench 2.0, compared with 75.1% for GPT-5.4 and 69.4% for Claude Opus 4.7. On SWE-Bench Pro, however, Claude Opus 4.7 remains in the lead (AI News).
Price: in the API it costs $5 per million input tokens and $30 per million output tokens, double the price of GPT-5.4. The Pro version costs $30 and $180. OpenAI says the model uses fewer tokens, so actual costs will rise by roughly 20% (AI News).

DeepSeek V4: A Cheap Chinese Rival with a Million-Token Context

On April 24, China’s DeepSeek released two open models in preview: the larger V4-Pro for coding and agents, and the smaller, faster V4-Flash. Both feature a reasoning mode and a 1-million-token context (MIT Technology Review).

The main draw is the price: V4-Pro costs $1.74 per million input tokens and V4-Flash about $0.14. According to the company, V4-Pro matches Claude Opus 4.6, GPT-5.4, and Gemini 3.1 on the main benchmarks (MIT Technology Review).

The link to Chinese hardware matters too. The model is optimized for Huawei Ascend chips, and Huawei announced that its Ascend 950 systems will support V4. For now, though, Chinese chips mainly handle running the finished model, while training has been moved onto them only partially (MIT Technology Review).

Kimi K2.6 and Qwen3.6-Max from China

On April 20, Moonshot AI released the open model Kimi K2.6, focused on coding and agents. By the company’s numbers, it scores 58.6 on SWE-Bench Pro, beating both GPT-5.4 and Claude Opus 4.6. Its Agent Swarm feature can coordinate up to 300 concurrent helper agents (OfficeChai).

On the same day, Alibaba released Qwen3.6-Max-Preview, the most powerful model in its Qwen family so far. Unlike the smaller Qwen3.6 models, it does not have open weights and runs only in Alibaba’s cloud. The company claims it leads six coding and agent benchmarks, including SWE-bench Pro and Terminal-Bench 2.0 (DataNorth).

Products and Agents

ChatGPT Images 2.0

On April 21, OpenAI launched its new image generation, ChatGPT Images 2.0, built on the gpt-image-2 model. The reasoning mode can look up information on the web, plan a layout, and create a set of images with consistent characters and style (eWeek).

The fast Images 2.0 Instant version is available to all ChatGPT and Codex users, including free ones, while the reasoning mode is limited to Plus, Pro, and Business subscribers.
The model handles resolutions up to 2K and aspect ratios from 3:1 to 1:3, and it renders text in images much better, even in Japanese, Korean, and Hindi (eWeek).

Team Agents in ChatGPT

On April 22, OpenAI added a workspace agents feature to ChatGPT. These are agents a team sets up by describing a workflow; they then run independently in the cloud, for example on a schedule or after a specific event, and they also work in Slack (Decrypt).

The feature is in a trial phase for the Business, Enterprise, Edu, and Teachers plans and is built on Codex technology. It is free until May 6, after which it moves to credit-based billing. Existing custom GPTs remain available for now (Decrypt).

Google Cloud Next: An Agent Platform in Place of Vertex AI

At the Google Cloud Next conference in Las Vegas on April 22, Google introduced Gemini Enterprise Agent Platform. It succeeds the Vertex AI platform, and all of its future services will be delivered only through the new platform (Google Cloud Blog).

The platform covers building agents (with the Agent Studio tool), running them with memory and an isolated environment, identity and security management, and evaluation.
It offers more than 200 models, including Gemini 3.1 Pro, the open Gemma 4, and Anthropic’s Claude models (Google Cloud Blog).
The new Workspace Intelligence connects Gmail, Docs, Drive, Meet, and Chat so agents can see context across all of them (The Decoder).

Chips and Data Centers

Eighth-Generation TPU: Separate Chips for Training and Serving

For the first time, Google has split its TPU chips (its own AI processors) into two variants. TPU 8t is for training models, and TPU 8i is for inference, meaning running finished models when they respond to users (Google Blog).

TPU 8t: a superpod with 9,600 chips, 2 petabytes of shared HBM memory, and 121 exaflops of performance. The Virgo network is designed to allow scaling up to a million chips.
TPU 8i: 288 GB of HBM memory and 384 MB of fast SRAM per chip, three times more than the previous Ironwood generation.
Both variants offer up to twice the performance per watt of Ironwood and will become available later in 2026 (Google Blog).

Intel Benefits from Demand for Server Processors

On April 23, Intel reported first-quarter revenue of $13.6 billion, up 7% year over year, and guided for $13.8 to $14.8 billion in the second quarter. Growth is driven by demand for Xeon server processors for AI data centers. The stock jumped 29% in premarket trading on April 24 (Techzine).

Business and Investment

Amazon and Google Pour Billions into Anthropic

On April 20, Amazon announced it will invest $5 billion in Anthropic right away and up to another $20 billion later, depending on business milestones. Anthropic committed to spending more than $100 billion on the AWS cloud over ten years (Reuters / The Star).

The deal includes up to 5 gigawatts of compute capacity on Amazon’s Trainium chips. Roughly 1 gigawatt is expected to be available by the end of 2026 (Reuters / The Star).

A few days later, on April 24 according to Bloomberg, Google came in with an even larger sum: up to $40 billion. It will provide $10 billion immediately at a $380 billion valuation of Anthropic, with the remaining $30 billion tied to performance targets (The Decoder).

SpaceX Secures an Option to Buy Cursor for $60 Billion

Elon Musk’s SpaceX, which has also owned xAI since earlier this year, announced on April 21 a partnership with Cursor, the popular AI code editor. The agreement gives SpaceX the right to buy Cursor’s parent company Anysphere later this year for $60 billion; otherwise it will pay $10 billion. Through the deal, Cursor gains access to xAI’s Colossus supercomputer (eWeek).

More Business News

Apple announced on April 20 that Tim Cook will become executive chairman on September 1, and that hardware engineering chief John Ternus will become the new CEO. It is Apple’s first CEO change in almost 15 years (Apple Newsroom).
Meta will lay off about 8,000 people, roughly 10% of its workforce, according to an internal memo from chief people officer Janelle Gale reported by Bloomberg on April 23. The layoffs begin on May 20 and are tied to the company’s massive AI spending (Fox Business).
Tesla raised its 2026 capital expenditure plan to $25 billion on April 22, about three times its usual annual level. The money will go to AI training, in-house chips, robotaxis, and production of the Optimus robot (TechCrunch).
OpenAI shut down the Sora video creation app and website on April 26, as announced in March. The Sora API will end on September 24 (The Decoder).

Regulation and Law

Florida Investigates OpenAI over University Shooting

On April 21, Florida Attorney General James Uthmeier opened a criminal investigation into whether OpenAI bears criminal responsibility for last year’s shooting at Florida State University. According to the attorney general, the accused asked ChatGPT a series of questions before the attack (ABC7 / CNN).

The office requested OpenAI’s internal policies and training materials on how the company handles user threats and reports crimes. OpenAI responded that the shooting was a tragedy but that ChatGPT is not responsible for the crime, and that it proactively shared the suspect’s account with police after the attack (ABC7 / CNN).

The White House Takes Aim at “Distillation” of American Models

On April 23, the White House Office of Science and Technology Policy issued memorandum NSTM-4, signed by Michael Kratsios. It warns against “industrial-scale” copying of American models, mainly by entities in China (CNBC).

The issue is so-called distillation: an attacker sends a huge number of queries to someone else’s model and uses the responses to train a cheaper model of their own. According to the memo, this involves tens of thousands of proxy accounts and the circumvention of safeguards. The government wants to share information with AI companies and consider measures against foreign actors (CNBC).

Security and Incidents

Unauthorized Access to Claude Mythos

On April 22, Anthropic confirmed it is investigating a Bloomberg report of unauthorized access to Claude Mythos Preview. Because of the model’s strong cybersecurity capabilities, the company had made it available only to partners in Project Glasswing (Engadget).

According to the report, a group from a private Discord channel reached the model through a third-party vendor’s portal and by guessing the model’s address. A Bloomberg source said the group only wanted to experiment with it, not use it for attacks (Engadget).

Anthropic Explains the Decline in Claude Code

After weeks of complaints about the declining quality of Claude Code, Anthropic published an analysis of the causes on April 23. There were three independent changes; none affected the API, and all had been fixed by April 20 (Anthropic).

The model’s default reasoning effort was lowered from “high” to “medium” on March 4 and restored on April 7.
A caching bug starting March 26 wiped the model’s earlier reasoning at every step; it was fixed on April 10.
A brevity instruction added to the system prompt on April 16 was removed on April 20.
As of April 23, the company reset usage limits for all subscribers (Anthropic).