Edited from four posts published in April 2025, covering March 31–April 27.
April was the month reasoning became a setting and agents needed a common language. OpenAI’s o3 and o4-mini chose and combined tools while they reasoned, and Gemini 2.5 Flash let developers decide how much thinking a response could use. GPT-4.1 brought million-token context to OpenAI’s API, xAI opened Grok 3 to developers, and OpenAI announced that GPT-4, introduced in March 2023, would leave ChatGPT at the end of the month. At Cloud Next, Google paired its new Ironwood inference chip with Agent2Agent, an open protocol for agents from different providers to coordinate tasks.
Money and governments moved too. SoftBank agreed to lead up to $40 billion of new OpenAI funding, OpenAI countersued Elon Musk, and Isomorphic Labs raised $600 million for AI drug design. Europe published its AI Continent Action Plan, Ireland’s privacy regulator opened an inquiry into Grok’s training data, and a US executive order created an AI education task force.
Agents needed more than models: Databricks, Amazon and Wikimedia each worked on getting business data, browsers and structured knowledge into a form agents could use. The laboratories also began to publish their own accounts of risk. Google DeepMind set out its approach to AGI safety and OpenAI revised its Preparedness Framework, while forecasts grew bolder: the AI 2027 scenario described how fast things might go, and on 60 Minutes Demis Hassabis said AGI could arrive as soon as 2030.
Ironwood, announced here, was readied for general availability in November, with Anthropic planning to use up to a million TPUs. Joelle Pineau’s departure from Meta’s research group opened a year of change in Meta’s AI leadership: the loss of most of Llama’s original authors reported in May, the Scale AI investment in June, Superintelligence Labs at the end of June, a hiring pause in August and reports in November that Yann LeCun planned to leave. Isomorphic’s funding led to preparations for human trials, reported in July. The education order was followed by a task-force meeting in September and a free ChatGPT for US teachers in November.
Isomorphic Labs announced $600 million in its first external funding round, led by Thrive Capital, with participation from GV and Alphabet. The company said the funding would support its AI drug-design engine and therapeutic programs.1
The investment backed an effort to turn computational drug design into treatments; clinical trials were the next milestone.
Amazon introduced Nova Act, a model trained to perform actions in a web browser, and released a research preview of its software development kit. Developers could use the kit to build and test agents for browser-based tasks.2
The browser offered access to workflows that lack a dedicated software interface, provided an agent could complete them reliably.
SoftBank agreed to invest up to $40 billion in OpenAI, with plans to syndicate $10 billion to other investors. Its expected own contribution was up to $30 billion. The agreement provided for staged funding, with the size of the second closing dependent on conditions tied to OpenAI’s restructuring.3
The agreement, paid in stages, signaled the scale of capital now sought for AI development.
Joelle Pineau, the vice president who had led Meta’s Fundamental AI Research group since early 2023, said she would leave the company in May, writing that “it is time to create space for others to pursue the work.” Meta said its AI research plans had not changed.4
Her departure opened a leadership gap in Meta’s research arm as the company increased its AI spending.
Google DeepMind published its approach to technical AGI safety and security. The paper considered misuse, misalignment, accidents and structural risks, with particular attention to misuse and misalignment. It discussed evaluating dangerous capabilities and developing safeguards.5
The paper made the laboratory’s assumptions and proposed protections available for outside scrutiny.
The AI Futures Project published AI 2027, a scenario by Daniel Kokotajlo and colleagues about rapid AI development and its consequences. The authors combined forecasts with a narrative and supplied two alternative endings, describing their work as one possible future.6
Its detail gave readers concrete assumptions to examine and contest.
The MegaMath paper described an open mathematics pre-training dataset containing 371 billion tokens. Its authors combined mathematical web documents, mathematics-related code and synthetic material.7
The release addressed the data used to develop mathematical reasoning, as a resource others could test through training experiments.
Stanford HAI released the 2025 AI Index, bringing together evidence on technical performance, investment, research, governance, education and public opinion. Much of its annual comparison described developments in 2024.8,9
The report supplied a broader frame for weekly announcements, measured mostly on the previous year.
The European Commission published the AI Continent Action Plan. Its program addressed computing infrastructure, access to data, adoption across sectors, skills and implementation of the AI Act.10
The plan treated AI development as an institutional task as well as a technical one, dependent on later investment.
At Cloud Next, Google announced Ironwood, its seventh-generation tensor processing unit, designed for inference. It also introduced Agent2Agent, an open protocol intended to let agents from different providers exchange information and coordinate tasks.11,12
The announcements addressed two requirements for wider deployment: capacity to run models and a common way for agents to work together.
OpenAI filed counterclaims against Elon Musk, accusing him of a campaign of harassment that included press attacks, legal claims and a “sham bid” for its assets, a reference to the $97.4 billion offer its board had rejected in February. The case was set for trial in spring 2026.13
The dispute over OpenAI’s structure moved from Musk’s original claims to accusations in both directions.
xAI made Grok 3 and Grok 3 Mini available through an API, charging $3 per million input tokens and $15 per million output tokens for Grok 3. The API supported a context window of 131,072 tokens, less than the one million xAI had claimed for the model in February.14
The release put xAI’s flagship model in direct price competition with other developers’ APIs.
OpenAI announced that GPT-4 would leave ChatGPT on April 30 and be replaced there by GPT-4o. The notice stated that GPT-4 would remain available through the API.15
GPT-4, introduced in March 2023, left the consumer product while remaining available to developers.
Ireland’s Data Protection Commission opened an inquiry into whether X lawfully used personal data in public posts by users in the European Union and European Economic Area to train xAI’s Grok models, examining the lawfulness and transparency of that processing under the GDPR.16
European privacy law became a constraint on what public posts could be used for in training.
OpenAI released GPT-4.1, GPT-4.1 mini and GPT-4.1 nano through its API. The models supported context windows of up to one million tokens. OpenAI reported improvements in coding, instruction following and the use of long inputs.17
The range offered developers different choices of capacity, speed and price, with the larger window expanding what a single request could include.
Google began rolling out Veo 2 video generation in Gemini Advanced for Google One AI Premium subscribers. The feature produced eight-second clips from text prompts. Whisk Animate also used Veo 2 to turn images into short videos for subscribers.18
The rollout placed video generation inside consumer creative tools, with subscription access and usage limits.
OpenAI updated its framework for assessing severe risks from advanced models. Its tracked categories covered biological and chemical capabilities, cybersecurity and AI self-improvement. The update defined High and Critical capability thresholds with corresponding safeguard requirements.19
The framework set out how the company intended to connect capability assessments to deployment decisions.
OpenAI released o3 and o4-mini, reasoning models able to select and combine tools in ChatGPT, including web search, Python and image generation. The release also introduced Codex CLI, an open-source experiment for working with code from a terminal.20
Reasoning could now guide a sequence of actions as well as an answer, making the choice of tools part of the task.
Wikimedia Enterprise announced a beta dataset on Kaggle containing structured Wikipedia content in English and French. The April 15 upload included article sections and other fields in a machine-readable format, excluding references and some non-prose material.21
The dataset gave developers an organized route to Wikipedia content, with its scope defined by what was included and left out.
Google released a preview of Gemini 2.5 Flash through the Gemini API in AI Studio and Vertex AI. Developers could switch its thinking mode on or off and set a thinking budget to control the balance between response quality, cost and latency.22
Reasoning became a setting developers could adjust for each application.
In a 60 Minutes interview, Demis Hassabis said artificial general intelligence could arrive as soon as 2030 and demonstrated Project Astra, an assistant that interprets the visual world in real time. He warned that “the race for AI dominance could become a race to the bottom for safety” and called for international cooperation on oversight.23
A Nobel laureate’s timeline for AGI reached a mass television audience.
Databricks described updates to AI/BI Genie, including suggested follow-up questions and file uploads. The upload feature, in gated public preview, let users combine Excel or CSV files with data in Unity Catalog. The roundup also covered conversation APIs in public preview.24
The updates brought external files into a conversational analysis workflow, initially for preview users.
OpenAI made gpt-image-1 available through its API. The model brought the image-generation capability introduced in ChatGPT the previous month into developers’ own applications.25
The API extended image creation beyond the chat interface into other companies’ products.
A US executive order established a White House AI Education Task Force and directed work on student resources, educator training and a Presidential AI Challenge. It called for partnerships involving government, educational institutions and industry.26
The order set a federal program in motion; the task force met to advance the Presidential AI Challenge in September.27
Sources checked October 26, 2026. Dates distinguish announcements, reports and later updates.