Everything that mattered in AI, one page a week.

Cognition shipped Devin 2.0, an agent-native IDE with interactive planning, answers about a codebase that cite files, and generated code wikis — and dropped the entry price from the $500-per-month team plan Devin launched on to $20, then pay-as-you-go. The $20 buys 9 ACUs (Agent Compute Units) at $2.25 each, and Cognition prices roughly 15 minutes of active Devin work at 1 ACU, so the entry tier is about 2.25 hours of agent work before metering; ACUs are more expensive per unit on the cheap plan than they were on the $500 subscription, where they cost $2.

The price cut came weeks after Cognition reportedly raised hundreds of millions, and into a field where GitHub Copilot agent mode, AWS Kiro and Q Developer, Windsurf and Magic's models all compete on the same autonomous-coding pitch. The bet is that a low floor converts trials, with the real margin sitting in compute credits rather than seat fees.

Sam Altman said on X on March 31 that OpenAI is excited to release a powerful new open-weight language model with reasoning in the coming months — a reversal of the closed-by-default posture, reported by WIRED as a response to DeepSeek's R1 and the reach of Meta's Llama weights, and following Altman's own earlier admission that OpenAI sat on the wrong side of history over open models. OpenAI technical staff publicly added that the model must run on your own hardware, and a safety researcher said the company would test it against its Preparedness Framework and not release anything judged catastrophically risky. No parameter count, license or context window was disclosed; this is a commitment, not a release.

That same day, Anthropic published version 2.1 of its Responsible Scaling Policy, effective March 31. The revision added a new capability threshold for CBRN development — the point at which a model could substantially uplift a moderately resourced state program — split its AI R&D thresholds into two levels (fully automating entry-level AI research work, versus dramatically accelerating the rate of effective scaling), and committed to re-evaluating thresholds whenever it upgrades safeguards. It did not raise or lower Anthropic's existing ASL-3 protections.

Amazon's AGI Lab in San Francisco released a research preview of the Amazon Nova Act SDK, a model trained to act inside a web browser. Rather than one long autonomous rollout, the SDK decomposes workflows into atomic commands — search, checkout, answer a question about what is on screen — that developers can annotate with constraints, interleave with Python for asserts, breakpoints and parallel thread pools, and mix with direct Playwright manipulation for steps like password entry. Amazon claims 0.939 on ScreenSpot Web Text against 0.900 for Claude 3.7 Sonnet and 0.883 for OpenAI's computer-use agent.

It ships as pip install nova-act with API keys from nova.amazon.com, alongside a new portal for the whole Nova family (text, Canvas images, Reel video), and The Verge notes the Nova lineup is positioned on price — Amazon claims at least 75 percent cheaper than comparable rivals. Nova Act already powers parts of the delayed Alexa+ upgrade, and Amazon's framing is deliberately narrow: short, reliable tasks with explicit human-in-the-loop handoffs rather than full autonomy.

OpenAI announced the largest private fundraising on record: $40 billion at a $300 billion post-money valuation, led by SoftBank, which committed $10 billion in mid-April and $30 billion in December, the second tranche contingent on OpenAI completing its conversion to a for-profit entity by the end of the year. SoftBank disclosed that its total investment could fall to as little as $20 billion if that restructuring does not happen; the remaining $10 billion comes from a syndicate including Microsoft, Coatue, Altimeter and Thrive. Roughly $18 billion is expected to fund OpenAI's commitment to Stargate, the SoftBank–OpenAI–Oracle data-center joint venture.

OpenAI also said ChatGPT now has 500 million weekly users, up from 400 million a month earlier, and reports expectations of tripling revenue to $12.7 billion by year-end. SoftBank is financing the first tranche with borrowings from Mizuho and other lenders rather than cash.

OpenAI added native image generation to GPT-4o, replacing DALL-E 3 as ChatGPT's default image generator and reclassifying DALL-E as legacy. Rather than handing requests to a separate diffusion service, image output is produced by the same transformer that handles text and vision, which is why the model can follow detailed instructions, render legible text inside images, edit an existing picture, and hold visual consistency across a conversation in a way earlier tools could not.

The reaction was the story. Altman said roughly a million new users arrived in about an hour, against about five days for ChatGPT's first million in 2022 — 'biblical demand' of a kind he said he had never seen — and then asked users to ease off because the team needed sleep and the GPUs were melting. OpenAI temporarily delayed the free-tier rollout, throttled generation, and moved free users toward three generations per day. Usage roughly doubled in the weeks around the launch; Altman cited around 500 million weekly active users by April, against 400 million in February.

What went viral was a style: users redrew family photos and memes in the visual language of Studio Ghibli, which ignited a copyright and attribution argument within days. OpenAI added a refusal that triggers when a user attempts to generate an image in the style of a living artist, and critics resurfaced a 2016 clip of Hayao Miyazaki calling an AI-driven animation demo an insult to life itself. The feature also had a documented bug — the model struggled to render the word 'sexy women' at all, which Altman called a bug to be fixed.

Google put Gemini 2.5 Pro out as an experimental preview under the identifier gemini-2.5-pro-exp-03-25 and made it available to all Gemini users. It is a thinking model that reasons before responding, covers text, audio, images, video and code, and carries a 1M-token context window with a 64K-token output limit.

It arrived at the top of the leaderboards and stayed there for months. On the June 2025 model card, Gemini 2.5 Pro GA posts 88.0% on AIME 2025, 86.4% on GPQA Diamond, 82.2% on Aider Polyglot, 54.0% on SimpleQA and 82.0% on MMMU. The experimental 03-25 build alone scored 86.7% on AIME and 84.0% on GPQA Diamond at launch. Google said all new Gemini models would include reasoning by default.

A file called DeepSeek-V3-0324 appeared on Hugging Face under DeepSeek's account with an MIT license and no accompanying announcement, technical report or changelog. Community members and benchmarking services found it within hours. It is a 671-billion-parameter mixture-of-experts model activating about 37 billion parameters per forward pass, keeping V3's multi-head latent attention, auxiliary-loss-free load balancing and multi-token prediction objective, with a 128K context window and a July 2024 knowledge cutoff.

The gains came from a better post-training pipeline rather than a new architecture, and they are large. MMLU-Pro rose from 75.9 to 81.2, GPQA from 59.1 to 68.4, AIME from 39.6 to 59, and LiveCodeBench from 39.2 to 49.2, with WebDev Arena improving by about 240 points and noticeably better front-end code and function calling. The license changed too: the original V3 shipped under DeepSeek's own license, while this checkpoint moved to MIT, the same terms as R1. Artificial Analysis flagged it as the first open-weight model to exceed all commercially available non-reasoning proprietary models on their composite quality score.

The release format is itself the statement. By uploading weights with no fanfare, DeepSeek bypassed the press cycle that Western labs build major launches around — and it did so weeks into a US debate in which OpenAI had accused DeepSeek to Congress of free-riding on the capabilities developed by American frontier labs.