WEEK OF MAR 29 – APR 4, 2025

OpenAI promised open weights and closed $40B the same day Anthropic raised its own bar

The largest private fundraise in history came with a governance condition attached. Anthropic quietly tightened its scaling policy on the same day, and every major coding tool shipped an agent mode.

OpenAI closed $40 billion at a $300 billion post-money valuation, and $30 billion of it is contingent on the company completing a for-profit conversion by the end of the year. Wealthy as it looks, that makes a corporate restructuring the load-bearing document underneath OpenAI's entire 2025 roadmap — pricing, model access and Stargate capacity all sit downstream of a conversion two state attorneys general can block.

Anthropic published version 2.1 of its Responsible Scaling Policy the same day, adding a CBRN capability threshold and splitting its AI R&D thresholds into two levels. And the agent harness stopped being a product and became a default: GitHub put agent mode and MCP support in front of every VS Code user, Cognition cut Devin's entry price from $500 a month to $20, and Amazon shipped its first browser agent.

FRI · Apr 4, 2025agent-harnessmcptoolinggithubmicrosoft

Agent mode with MCP support rolls out to every VS Code user as Microsoft turns 50

On Microsoft's 50th anniversary, GitHub began rolling agent mode — plus Model Context Protocol support — out to all VS Code users in Stable, and published an open-source local GitHub MCP server so any MCP-capable LLM client can call GitHub tooling. Agent mode runs multi-step tasks: it reads the workspace, proposes edits, runs terminal commands it asks you to approve, watches compile and lint output, and self-corrects until the task completes. Tool surface is now three-way — built-in VS Code tools, MCP servers, and VS Code extensions that contribute language-model tools — configured through .vscode/mcp.json with ${env:...} input variables instead of hard-coded secrets.

GitHub also made Anthropic, Google and OpenAI models generally available inside Copilot and introduced a new Copilot Pro+ tier metered by premium requests, with allowances beginning May 5, 2025. Satya Nadella's keynote added a Code Review Agent and a multi-agent framework in Azure AI Foundry, alongside red-teaming agents and code-vulnerability tooling.

Agent mode with MCP support rolls out to every VS Code user as Microsoft turns 50
GitHub

WHY IT MATTERS

MCP went from an Anthropic-specific integration to a first-class transport inside the world's most-used editor in about four months. If you maintain a dev tool, a design system, or internal infrastructure, the cheapest distribution channel for it is now an MCP server — and the attack surface moved with it: tool descriptions and tool responses arrive in the model context from third parties, so an MCP server you did not write can carry instructions the agent treats as trusted.

THU · Apr 3, 2025agent-harnesspricingcoding-agentscognition

Devin 2.0 cuts the autonomous coding agent from $500/month to a $20 entry point

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.

Devin 2.0 cuts the autonomous coding agent from $500/month to a $20 entry point
cdn.sanity.io

WHY IT MATTERS

Agent pricing is converging on metered compute plus a near-zero entry fee, which changes how you budget: a seat is no longer a fixed line item, and a runaway agent loop is a variable cost on your card. The $2.25 per ACU figure is the number to anchor on — 15 minutes of active agent time for the price of a coffee is roughly an order of magnitude below human contractor rates, so the build-versus-agent call now turns on review time, not generation cost.

MON · Mar 31, 2025open-weightssafetypolicyopenaianthropic

OpenAI promises an open-weight reasoning model — the same day Anthropic adds a CBRN threshold to its scaling policy

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.

WHY IT MATTERS

An open reasoning model from a frontier US lab is the single most decision-relevant thing that could happen to inference economics this year: it moves reasoning capability out of per-token API billing and onto hardware you control, which matters for regulated workloads, long agent loops and anything with a token bill that scales with usage. The contrast with Anthropic's same-day tightening is worth reading carefully — the framing of safety as a scaling gate that only closed labs can satisfy loses force when the weights ship, and the licensing terms of whatever OpenAI actually publishes will decide whether it is usable rather than merely open.

MON · Mar 31, 2025agentsbrowser-agentssdkamazon

Amazon enters browser agents with Nova Act, a research-preview SDK built around atomic commands

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.

WHY IT MATTERS

The interesting engineering claim is not the model, it is the reliability model: breaking agent work into small, testable, atomic commands with human intervention points is the opposite of the give-it-a-goal-and-hope pattern most harnesses shipped in 2025, and it maps onto how you would actually write tests for an agent. If your product has no API — internal admin panels, legacy vendor portals, a checkout you do not control — a browser harness priced on a commodity model is now a viable integration path rather than a research demo.

MON · Mar 31, 2025fundingcomputestargateopenai

OpenAI closes $40B at a $300B post-money valuation, with a for-profit conversion clause attached

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.

WHY IT MATTERS

The deal is a compute financing story with a governance condition baked in: $30B of the $40B is contingent on a legal restructuring, so OpenAI's 2025 roadmap, pricing and model-access decisions now sit downstream of a corporate conversion that Microsoft's board and the California attorney general have to bless. For anyone building on OpenAI, that is the risk to price in — the capital, the Stargate capacity and the API roadmap are all tied to a conversion deadline that third parties can block.