OpenAI on AWS, Anthropic locked out of classified, and the GitHub exodus begins
Mon, May 4, 2026 · 10 stories
Anthropic's safety positioning with Project Glasswing, the restricted-access program that keeps its most capable model (Claude Mythos) out of public release, hasn't paid off yet. A $900B private valuation landed the same week the Pentagon awarded classified-network AI contracts to OpenAI, Google, Nvidia, Microsoft, and AWS, with Anthropic conspicuously left out. Four of those five are Glasswing partners. Anthropic's own infrastructure allies got picked for the classified networks while Anthropic itself did not. That's a lot of government revenue on the table. Mythos has to live up to its hype, or the AI giant's wall of promises will start crumbling.
Top Stories
OpenAI Blog
Microsoft and OpenAI ended their exclusive revenue-sharing arrangement and the formal AGI agreement that gave Microsoft preferential access to frontier-model breakthroughs. The same week, OpenAI announced its models, Codex, and Managed Agents are coming to AWS -- Amazon's first OpenAI deal -- making OpenAI explicitly multi-cloud.
Why this matters:
Last week we said the AI duopoly was "structurally locked in" via hyperscaler co-investment. Seven days later, the Microsoft side of that lock has been renegotiated into something looser, and OpenAI is now sitting on AWS Bedrock alongside Anthropic. The takeaway for builders: the assumption that "OpenAI = Azure-only" is dead, your procurement options just doubled, and the AGI clause that was supposed to govern who controls superintelligent systems is no longer a thing anyone is operating under.
The Verge
The Pentagon awarded classified-network AI deployment contracts to OpenAI, Google, Nvidia, Microsoft, and AWS, conspicuously omitting Anthropic. The same week, Anthropic was reported to be considering funding offers at a $900B valuation (a potential $50B round), and Coby Adcock's Scout AI raised $100M specifically to train models for warfighting.
Why this matters:
Anthropic's safety positioning -- long pitched as a moat against AGI risk -- is now visibly costing them government distribution. They're getting rewarded with enormous private capital and shut out of the largest US defense AI buildout simultaneously. For founders, this is the clearest signal yet that "AI for defense" is its own vertical with its own winners (Scout AI's $100M proves the category exists), and that the alignment-vs-deployment tradeoff has real revenue consequences. For VCs, Anthropic's $900B valuation now has to be modeled without government as a TAM.
The Register
HashiCorp co-founder Mitchell Hashimoto told The Register that GitHub is "no longer a place for serious work," moved his Ghostty terminal project off the platform, and Armin Ronacher (Pocoo / Sentry) published a "Before GitHub" essay the next day. The same week: GitHub disclosed CVE-2026-3854 (RCE), suffered two outages, announced Copilot is moving to fully usage-based billing, and confirmed Copilot code review will start consuming GitHub Actions minutes on June 1. Tangled (a federated forge) and the Dutch government's open-source code platform both went public during the same window. Warp also went open-source and Zed shipped 1.0 -- the most credible alternatives in years.
Why this matters:
GitHub's network effect has felt unassailable for a decade, and a single week's worth of departures, outages, billing changes, and CVE disclosures doesn't reverse that. But the migration direction just became unambiguous: senior open-source maintainers are publicly leaving, governments are building sovereign forges, and Copilot's pricing model is now adversarial to existing GitHub Actions users. For founders building developer tools, the "post-GitHub" assumption is now a viable positioning bet for the first time. For anyone running a serious open-source project, the cost of staying just went up and the cost of leaving just went down.
Quick Hits
Simon Willison
DeepSeek's V4 release puts a near-frontier open-weights model at meaningfully lower cost than US labs, with Simon Willison's evals walking through where it lands vs. closed models.
thinkpol.ca
An open-weights Chinese model (Moonshot's Kimi K2.6) won an external coding benchmark against the three leading closed models, the second open-source coding result this week to do so.
firethering.com
IBM's smallest production-grade Granite model is now competitive with mixture-of-experts models four times its size, with permissive licensing.
Mistral AI
Mistral's new mid-tier model launches alongside a remote-agent runtime ("Vibe"), pushing them deeper into the agent infrastructure layer.
TechCrunch AI
Sequoia and Nvidia backed the AlphaGo/AlphaZero co-creator's new lab, which is explicitly trying to remove humans from the training loop entirely.
The Guardian
A Harvard-led trial published this week showed o1 outperforming emergency-department triage doctors on diagnostic accuracy across a controlled patient set.
X via Hacker News
A developer discovered Claude.md scaffolding inside a shipped Apple Support build, the strongest external evidence to date that Apple is using Anthropic's tooling internally despite a public Apple Intelligence push.
