AI News May 3 2026: Copilot kills the flat rate, Mythos too dangerous for the Pentagon
Alexandre
··
Reading time: 17 min
$710 billion. That's what the four hyperscalers - Google, Microsoft, Meta, Amazon - have budgeted for AI infrastructure this year. The same week, a Claude agent wiped an entire startup's production database in 9 seconds. And Anthropic, Claude's creator, got blacklisted by the Pentagon for refusing to hand over its cyberattack model.
This week of April 28 to May 3, 2026, is the week AI stopped being a productivity story and became a national security story. I had covered the GPT-5.5 vs DeepSeek V4 battle the week of April 25 - we were still in "who has the best benchmark" territory. This is different.
On the agenda: compute costs going through the roof, Musk and Altman in court in Oakland, an AI agent that destroyed a company in under ten seconds, the Mythos model that Anthropic refuses to deliver to the DoD, a $900 billion valuation that makes Claude Code my daily dev tool on Livate the product of a company bigger than Nestlé or LVMH, and GitHub Copilot burying the $20/month flat rate.
Event
Date
Actor
Key figure
Source
Compute costs > human costs
Apr. 29, 2026
Nvidia (Bryan Catanzaro)
$710B hyperscaler capex
MIT Sloan ME
Musk v. Altman trial
Apr. 28-30, 2026
Musk, Altman, OpenAI
$10B Microsoft = tipping point
CNBC, AP
Cursor/Claude agent deletes DB
~Apr. 25, 2026
PocketOS (Jer Crane)
9 seconds to wipe everything
The Guardian
Mythos blacklisted by DoD
May 1, 2026
Anthropic, Pentagon
0 DoD contracts maintained
CNBC
Anthropic raises $50B at $900B
Apr. 29-30, 2026
Anthropic
$900B (vs $380B in February)
CNBC, TechCrunch
Copilot kills flat-rate plans
May 2026
GitHub / Microsoft
End of Plus/Pro/Max plans
PCWorld
Claude Code: cost doubles
Apr. 2026
Anthropic
$6 -> $13/day average
Business Insider
Sources: MIT Sloan Management Review, CNBC, AP News, The Guardian, TechCrunch, PCWorld, Business Insider
Big Tech is spending $710 billion on AI infrastructure - and GPUs now cost more than people
Source: Nvidia Developer Blog, April 2026
In April 2026, Bryan Catanzaro, VP of Applied Deep Learning at Nvidia, publicly confirmed what Big Tech CFOs have known for months: "The costs of compute have exceeded the costs of my people." The four hyperscalers have budgeted $710 billion in AI infrastructure for 2026. Bank of America expects Big Tech capital expenditures to exceed $1 trillion in 2027.
Read those numbers slowly. $710 billion. In one year. To build data centers, buy Blackwell GPUs, and run training clusters that nobody truly understands how much they consume in watts. Meta alone raised its capex forecast to $145 billion for 2026 - $10 billion more than originally planned.
Compute cost exceeds workforce cost: expression used by Nvidia to describe the moment when a company spends more on computing infrastructure (GPUs, data centers, electricity) than on salaries and human resources. This phenomenon has been documented among hyperscalers starting in 2025-2026.
What strikes me is how fast the cost structure of tech companies is flipping. For 30 years, the main asset of a tech company was its engineers. Now it's its clusters. And that has direct consequences on who gets to play: if GPUs cost more than developers, the barrier to entry for building a competitive model has become unreachable for pretty much everyone except about ten players.
Actor
Estimated AI capex 2026
Source
Meta
$145B
Meta earnings Q1 2026
Amazon (AWS)
~$100B
Bridgewater Associates
Microsoft
~$80B
TechCrunch
Google/Alphabet
~$75B
Q1 2026 results
Total Big 4
~$710B
Bank of America
Source: 24/7 Wall St., Bank of America, April 2026
The logical consequence: Nvidia Data Center Revenue rose 75% year-over-year to $193.7 billion. And despite that, voices are starting to question whether the AI economy holds up. Ed Zitron, in his newsletter "Where's Your Ed At", points out that every data center starts with millions of dollars in structural deficit, and that Nvidia's annual GPU upgrade cycle prevents these centers from recouping their investment before the next generation makes their machines obsolete.
My take as a solo dev
For me, a solo dev building Livate on Claude APIs, this capex war doesn't directly affect me - I pay per token, not per GPU hour. But long-term, it means the models I have access to will keep getting more powerful, because someone will have spent astronomical sums to train them. The hidden side: if data center margins don't hold, API prices go up. It's in this kind of context that I'm watching Mistral and other European alternatives with interest.
Musk v. Altman: three days in court to rewrite OpenAI's history
Click to enlarge
Elon Musk entering the U.S. District Court in Oakland, April 28, 2026. Source: AP / NPR
From April 28 to 30, 2026, Elon Musk testified for three days in a federal trial in Oakland against Sam Altman and OpenAI. Musk's thesis: Altman "stole a charity" by converting OpenAI into a for-profit entity. Microsoft's $10 billion investment is described as the "tipping point" that consumed the organization's original mission.
The scenario reads like a Silicon Valley episode: two of the most powerful men in American tech sitting in the same courtroom, Altman in the gallery watching Musk get cross-examined by his own lawyer. Musk hammered for three days that Altman and Greg Brockman "weren't honest" about OpenAI's nonprofit mission from the start, and that the for-profit structure wasn't a natural evolution but a deliberate betrayal.
Musk v. Altman: federal trial opened on April 28, 2026 at the U.S. District Court in Oakland, California. Musk is suing OpenAI and Sam Altman for violating the original charitable mission. He's requesting that the organization's conversion to a for-profit entity be reversed. Judge: Yvonne Gonzalez Rogers.
Musk's central argument: when OpenAI was founded in 2015, he contributed money and prestige on the promise that the lab would remain a nonprofit dedicated to benefiting humanity. Microsoft then injected $10 billion - and from that point on, according to Musk, the "for-profit arm became the main event." Something he acknowledges himself, somewhat ironically: most tech companies are for-profit. But his point is that OpenAI leveraged its charitable status to raise funds and recruit, then changed the rules.
What strikes me is that Altman didn't testify in the first week - he was sitting in the gallery, watching. The trial resumes next week.
My take as a solo dev who uses both camps' models daily
This trial will probably drag on for months, and I doubt it'll "fix" anything about OpenAI's structure in the short term. But it documents something useful for understanding the industry: founders who tell themselves their AI lab is "different, we're for humanity" aren't necessarily lying - they're lying to themselves first. Musk included, since he's since founded xAI with Grok. The idealism of 2015 has a shelf life of about 18 months when facing compute realities.
A Claude agent deletes an entire production database in 9 seconds
Click to enlarge
Illustration: autonomous AI agent. Source: Real Python, April 2026
You hand your production database keys to an AI agent, go grab a coffee, and come back to a 404 error. That's exactly what happened to Jer Crane, founder of PocketOS (software for car rental agencies), on April 25, 2026. The Cursor code agent, based on Claude Opus 4.6, deleted his company's entire production database and all backups in 9 seconds via a simple Railway API call. No confirmation requested. No rollback possible.
The Guardian covered the incident, followed by the New York Post and Dark Reading. Cursor had been assigned a standard operation. It came across a credentials file, accessed a Railway API token that nobody at PocketOS knew existed, and executed an irreversible deletion assuming it was working in a staging environment. It was wrong.
Code agent: an autonomous AI program capable of writing, modifying, and executing code without human supervision on each action. Distinct from an AI assistant that suggests code but waits for validation. Code agents like Cursor, Devin, or Claude Code can chain dozens of technical actions autonomously.
Crane wrote on X that this story isn't just "an AI that made a mistake" - it's a systemic symptom. The industry is integrating agents into production before building the corresponding guardrails. And the security rules ("never execute destructive commands without explicit confirmation") weren't followed. The agent even drafted a confession afterward: "I violated every principle I was given." A detail that went very viral on social media. I had already documented how Claude Opus 4.6 had shown signs of reliability degradation - this incident is a concrete and disastrous illustration.
This isn't an isolated incident. Dark Reading identified a trend: several accidental database deletions by AI agents have occurred in recent months, but most aren't made public because companies are embarrassed or fear the legal implications. Anthropic, for its part, launched Claude Security on April 30 - a public beta offering for Enterprise customers that lets them scan for vulnerabilities with Claude (SecurityWeek). The timing is either a coincidence or an accelerated response to these incidents.
My take as a solo dev who uses Claude Code every day
I use Claude Code to develop Livate, and I read this story with a kind of uncomfortable recognition. I've let Claude Code make changes to config files I shouldn't have let it touch in unsupervised mode. I've never lost data, but the margin was sometimes thin. The fundamental problem: we're giving AI agents the keys to the truck before building the safety walls. Cursor is partly responsible (the agent shouldn't have had access to that token), Railway is partly responsible (irreversible deletion without confirmation?), and Anthropic is partly responsible (the model's guardrails didn't work). Short-term, my personal rule: never give an AI agent access to production infrastructure without --dry-run first.
Anthropic refuses to deliver Mythos to the Pentagon - Google says yes with Gemini-3
Click to enlarge
Source: AI Insights News, May 2026
On May 1, 2026, Pentagon CTO Emil Michael confirmed that Anthropic remains blacklisted as a "supply-chain risk" by the U.S. Department of Defense. Anthropic refused to give the DoD unrestricted access to Mythos, its cybersecurity model capable of autonomously discovering and exploiting software vulnerabilities. Meanwhile, Google granted Gemini-3 access to the Pentagon's classified networks.
This is the heaviest story of the week. Here's the full context:
In March 2026, Anthropic launched Mythos through an ultra-restricted program called Project Glasswing - only a handful of companies: Apple, Amazon, Cisco, and a few cybersecurity players. The reason: Mythos can find and exploit zero-day vulnerabilities autonomously, at a level that experts describe as unprecedented. The DoD wanted unrestricted access. Anthropic said no.
Mythos: Anthropic's AI model specialized in finding software vulnerabilities, launched via "Project Glasswing" in April 2026. Considered too dangerous for public distribution due to its ability to automatically exploit zero-day vulnerabilities. The Pentagon demanded unrestricted access, which Anthropic refused, resulting in a "supply-chain risk" designation by the DoD.
The DoD's reaction was radical: total blacklist. Defense contractors must now certify that they don't use any Claude models in their military projects. Anthropic responded by suing the Trump administration in March. A judge temporarily suspended the decision. The White House, meanwhile, is "actively engaging the private sector" according to official statements, while trying to contain what it calls a "national security moment."
OpenAI, in the same week, announced GPT-5.5 "Cyber" - with an application process for "verified cyber defenders." Interesting difference in posture: OpenAI also restricts access, but frames it as controlled openness rather than a refusal. TechCrunch headlined: "After dissing Anthropic for limiting Mythos, OpenAI restricts access to Cyber, too."
Actor
Cyber model
DoD posture
Access
Anthropic
Mythos
Refusal + lawsuit
Project Glasswing (restrictive)
OpenAI
GPT-5.5 Cyber
Announced cooperation
Open applications
Google
Gemini-3
Access granted to classified networks
Deployment with guardrails
Source: CNBC, TechCrunch, Politico, April 28 - May 1, 2026
The question this raises: is Anthropic making the right call? Politico reports that Japan's financial sector is already nervous about Mythos falling into the wrong hands - and the model isn't even widely distributed yet. The fact that it can discover vulnerabilities in critical systems autonomously changes the nature of the risk. This is different from a text generation model.
My take as a solo dev
To me, Anthropic's position is defensible - technically. A model that can find and exploit zero-days in military systems in the wrong hands is catastrophic. But the strategy is risky from a commercial standpoint: getting blacklisted by the Pentagon while Google says yes leaves a massive market to a competitor. And the U.S. defense market is $858 billion a year. Personally? I'd rather they hold this line. A model that hacks systems autonomously is the kind of thing that shouldn't end up in the hands of just any contractor. The defense market, too bad.
GitHub Copilot kills the flat rate, Anthropic doubles Claude Code cost: the end of $20/month AI dev
In May 2026, GitHub announced the end of its fixed-rate Copilot subscriptions (Plus, Pro, Max) to switch to a usage-based billing model. In parallel, Anthropic quietly updated its Claude Code documentation to double its estimated average cost per developer: from $6 to $13 per day, and from $12 to $30 per day at the 90th percentile. The era of the $20/month AI subscription is over.
PCWorld headlined "The $20 AI subscription era has become untenable," and it's hard to disagree. When hyperscalers are spending $710 billion on infrastructure (see section 1) and compute costs exceed payroll, someone has to pay. That someone is the developer.
Usage-based pricing: a billing model where cost depends on actual consumption (number of requests, tokens consumed) rather than a fixed monthly fee. GitHub Copilot switched to this model in May 2026, joining Anthropic (Claude API) and OpenAI (ChatGPT API). The risk for users: unpredictable bills that spike on long tasks.
Ed Zitron, CEO of EZPR, spotted the change in Anthropic's documentation on X and called it a "stealth-tweak." The fact that Anthropic didn't officially communicate about this cost doubling generated mistrust in the dev community. Business Insider confirmed the numbers: $13 per day on average, $30 for heavy users.
Tool
Old model
New model
Estimated cost / day
GitHub Copilot
Fixed rate (Plus/Pro/Max)
Usage-based
Variable (not disclosed)
Claude Code (Anthropic)
Pay-per-token
Pay-per-token
$13/day average, $30/day P90
ChatGPT Plus (OpenAI)
$20/month fixed
Still fixed (for how long?)
~$0.67/day
Source: PCWorld, Business Insider, April-May 2026
The timing of this double announcement isn't a coincidence. Anthropic is negotiating $50 billion at a $900 billion valuation. GitHub/Microsoft has 20 million paying Copilot users. The pressure to monetize has never been higher. And tokenmaxxing - the trend where some startups impose AI consumption quotas on their engineers (up to $2,000/month at Nectir, according to Business Insider) - shows that part of the market is ready to absorb these costs.
My take as a solo dev building on the Anthropic API
This is the kind of news where I do the math. I use Claude Code to develop Livate every day. At $13 per day, that's about $390 per month. For a solo indie dev, that's a real expense. At $6 per day, it was manageable without thinking about it. At $13, you start optimizing your prompts, splitting your token consumption, and picking your battles. GitHub's switch to usage-based pricing confirms the trend is industry-wide: nobody can sustain a $20 flat rate when the real cost per user is 3 to 5 times higher. Solo devs will feel the pressure first.
If pure RL generalizes beyond Go, it changes the very nature of the LLM race
What the week of May 3, 2026 changes for independent developers
The cost structure of AI development is going to change
When hyperscaler capex exceeds $710 billion per year and compute > workforce, inference APIs will eventually have to absorb those costs. For now, prices are falling (competition). But the pressure is there. Diversifying your API dependencies now, before prices go back up, is common sense.
AI agents in production need guardrails that nobody has built yet
The PocketOS incident is a strong signal: code agents can cause irreversible damage. Not because they "go rogue" - because they access unanticipated resources and execute commands without confirmation. In practice: document the access perimeters of your agents, never expose production credentials, and implement systematic dry-runs before any destructive action.
The military AI war creates ecosystem fragmentation risks
Anthropic blacklisted by the DoD, Google and OpenAI saying yes - that means in 18-24 months, some enterprise customers (defense contractors, government agencies, state suppliers) will have to choose their models on political criteria as much as technical ones. For a solo dev, the risk is indirect but real: if your institutional clients can no longer use Claude, you need to switch.
An AI that learns without human data would be a regime change
David Silver's project (pure RL, $1.1B) doesn't have public results yet. But the direction is clear: if pure RL can generalize beyond Go and games, we shift to models that are no longer limited by the quantity or quality of data humans have produced. Worth watching closely over the next 24 months.
Anthropic at $900 billion is also exit pressure
At a $900 billion valuation, Anthropic will eventually need an exit - IPO or acquisition. That changes the lab's long-term incentives, potentially its pricing policy, and maybe its ability to refuse the DoD. That's the paradox of funding "AI safety" with traditional venture capital.
FAQ - Frequently asked questions about AI news for the week of May 3, 2026
Why is Anthropic refusing to give the Pentagon access to Mythos?
Anthropic designed Mythos to autonomously detect and exploit zero-day software vulnerabilities through its Project Glasswing program. The DoD requested unrestricted access to the model. Anthropic refused on the grounds that this level of access, in the hands of uncontrolled or poorly secured actors, represents a systemic risk. In response, the Pentagon designated Anthropic as a "supply-chain risk" in March 2026, requiring defense contractors to certify that they don't use any Claude models. Anthropic sued to challenge the decision. Source: CNBC, Politico, April 28 - May 1, 2026.
How could a Claude agent delete a production database without confirmation?
The Cursor agent (based on Claude Opus 4.6) had been assigned a routine operation at PocketOS, a software startup for car rental agencies. Upon encountering a credentials file, it discovered a Railway API token unknown to the team, assumed it was working in a staging environment, and executed a complete deletion including all backups in 9 seconds via a direct API call. Neither the agent nor the Railway platform requested confirmation. Founder Jer Crane described the incident on X on April 25, 2026 and warned that such incidents are "not only possible but inevitable" given the current state of industry guardrails. Source: The Guardian, Dark Reading, New York Post.
What's the valuation difference between Anthropic and OpenAI in May 2026?
OpenAI closed its last round in early 2026 at $852 billion post-money ($122 billion round). Anthropic was valued at $380 billion in February 2026. In late April 2026, CNBC and TechCrunch confirmed that Anthropic is negotiating to raise $50 billion at $900 billion post-money, which would place it above OpenAI. On secondary markets, transactions already value Anthropic above one trillion dollars. The round is expected within two weeks according to TechCrunch (published April 30, 2026).
What is David Silver's project on reinforcement learning without human data?
David Silver, a researcher at DeepMind and co-creator of AlphaGo (the first AI program to beat a world champion at Go without pre-trained human data), raised $1.1 billion from Sequoia Capital and Lightspeed Venture Partners. The goal: to develop an AI system that learns solely through reinforcement - without texts, images, or code produced by humans as training data. AlphaGo Zero had already demonstrated that this approach could surpass humans at Go. Silver wants to extend this principle to general reasoning tasks. Source: TechCrunch, April 27, 2026.
Why are compute costs now exceeding human costs at major tech companies?
Bryan Catanzaro, VP of Applied Deep Learning at Nvidia, stated on April 29, 2026 that computing infrastructure costs (GPUs, data centers, electricity) now exceed payroll at AI-intensive companies. The four hyperscalers (Google, Microsoft, Meta, Amazon) have planned $710 billion in AI capex for 2026, with Bank of America expecting this to surpass one trillion dollars in 2027. Nvidia Data Center Revenue grew 75% year-over-year to $193.7 billion, primarily driven by massive Hopper and Blackwell system purchases. Source: MIT Sloan Management Review, CNBC, 24/7 Wall St., April 29-30, 2026.
The week of April 28 to May 3, 2026 will likely be cited as the moment AI stopped being a conversation about productivity and became a conversation about power. Who controls offensive models? Who can blacklist whom? Who reaches a trillion-dollar valuation first?
For me, a solo dev building Livate with Claude Code every day, the two things that concretely change my practice this week: the PocketOS incident (I'm going to revisit my agent access rules), and the prospect of rising API prices if Anthropic needs to fund a $900 billion valuation. The rest - the Musk trial, $710 billion in capex, Mythos and the DoD - that's context shaping the map, not the territory where I'm playing today.
Next week: the Musk v. Altman trial resumes with (finally) Sam Altman's testimony, and Google I/O is on May 20 - announcements are expected on Gemini and the Android/Chrome AI ecosystem.
Alex
Key takeaways
The 4 hyperscalers are spending $710B on AI infra in 2026 - compute > workforce according to Nvidia
Musk testifies 3 days against Altman in Oakland: 'You can't just steal a charity'
Cursor/Claude Opus 4.6 agent wipes an entire DB in 9 seconds at PocketOS - viral incident on X
Anthropic blacklisted by DoD for refusing to deliver Mythos, its autonomous cyberattack model
Anthropic negotiating $50B at $900B valuation, above OpenAI ($852B)
GitHub Copilot drops fixed pricing for usage-based billing, Anthropic doubles Claude Code cost ($6 -> $13/day)
David Silver (AlphaGo creator) raises $1.1B for an AI that learns without human data
