Andrej Karpathy joins Anthropic for frontier LLM R&D

<p>Andrej Karpathy, one of OpenAI’s original co-founders and former Tesla AI director, has joined Anthropic. He’s not advising or taking a figurehead role — he’s on the pre-training team, starting immediately, reporting to Nicholas Joseph (himself an ex-OpenAI researcher).</p>
<p>His mandate is specific: use Claude to accelerate the research that builds the next Claude. Pre-training is the most compute-intensive phase of frontier model development, and the bet here is that AI-assisted research loops can compress the timeline meaningfully.</p>
<p>This isn’t just a theory Anthropic is testing. Karpathy already ran a version of it himself. In March he released an open-source project called <code>autoresearch</code> — give an AI agent a training script, a frozen eval metric, and a five-minute compute budget per experiment, then let it iterate overnight. It worked. The autonomous research loops Anthropic is now building around are loops Karpathy already validated in public.</p>
<p>He’ll work alongside John Schulman, another former OpenAI co-founder who joined Anthropic last year.</p>
<p>The hire comes during a busy month for Anthropic on the talent front. Ross Nordeen, a founding xAI member and ex-Tesla engineer, joined earlier in July. Chris Rohlf, a security engineer with over 20 years of experience including time at Meta, joined the frontier red team the same day as Karpathy.</p>
<p>Karpathy is probably best known outside research circles for coining the term “vibe coding” in early 2024. In his own announcement, he said he thinks the next few years at the frontier will be “especially formative” and that he plans to return to his education work in time.</p>
<p>For Anthropic, the timing matters. The past two months have been rough — a stretch of user complaints about Claude gave OpenAI room to push forward with GPT-4.5. Whether Karpathy’s arrival changes the trajectory of the next model is an open question, but it’s hard to read this as anything other than a serious signal about where Anthropic thinks the leverage is: not just raw compute, but using existing models to accelerate the research itself.</p>


Andrej Karpathy, OpenAI co-founder and former Tesla AI director, has joined Anthropic to lead Claude pre-training research. He reports to Nicholas Joseph and will work alongside John Schulman, another ex-OpenAI co-founder. His mandate is to use Claude itself to accelerate the research that builds the next Claude — a self-improving research loop he already validated publicly with his open-source 'autoresearch' project in March. The hire is part of a broader talent push at Anthropic in July, which also brought in Ross Nordeen from xAI and security engineer Chris Rohlf. Karpathy coined the term 'vibe coding' and plans to eventually return to education work. The move signals Anthropic's belief that AI-assisted research loops, not just raw compute, are the key lever for frontier model development.

Standard Chartered to cut 7,800 back-office jobs to AI by 2030

Standard Chartered CEO Bill Winters announced at an investor day in Hong Kong that the bank will cut approximately 7,800 back-office jobs — over 15% of roles in HR, risk, and compliance — by 2030 as part of an AI-driven efficiency push. The goal is to raise income per employee by 20% by 2028. Winters described the shift as replacing 'lower-value human capital' with machines, language that is drawing scrutiny from unions and regulators. Standard Chartered is the first major bank to publicly attach a specific percentage and functional area to such a commitment. The announcement fits a broader cross-sector pattern, with JPMorgan, Citi, HSBC, Meta, and Klarna all making similar AI-driven headcount moves.

Meta reassigns 7,000 employees to AI roles the same week it cuts 10% of workforce

Meta is reassigning 7,000 employees to AI-focused roles — covering agents, apps, and infrastructure — as part of a broad corporate restructuring, while simultaneously preparing to cut 10% of its workforce later the same week. The move, described in an internal memo from Chief People Officer Janelle Gale, reflects Meta's strategy of redeploying institutional knowledge toward AI while shedding other headcount. The restructuring coincides with a $145bn Q1 2026 capex run-rate and a first-ever quarterly decline in daily active users. The company has not disclosed which workers are being moved versus let go, or whether this is a one-time event or a rolling reorganization.

Linux Foundation finds AI security readiness tripled as top enterprise adoption barrier since 2024

The Linux Foundation's 2026 State of Tech Talent Report identifies AI security readiness as the top barrier to enterprise AI adoption, with security and privacy concerns nearly tripling from 17% in 2024 to 48% in 2026. Despite 97% of organizations committing to AI implementation, 57% report significant gaps in security and risk management. The report also challenges the narrative that AI kills jobs, finding net hiring is projected at 31% for 2026. Organizations are responding by prioritizing upskilling existing staff over new hires, with data showing upskilling outperforms hiring across dimensions like retention, cost, and team cohesion.

Meta cuts 8,000 jobs while raising AI infrastructure spend to $145B

Meta is cutting 8,000 jobs starting May 20 — its largest single layoff round since 2023 — while simultaneously reporting record Q1 2026 revenue of $56.31 billion and raising its AI infrastructure spending guidance to $125–145 billion for 2026. The cuts are not driven by financial distress but by a strategic decision to convert payroll into AI capital expenditure, funding data centers, Nvidia GPUs, and custom silicon for the Llama ecosystem. Employee morale has collapsed amid shrinking median compensation (down to $388,200), a controversial keystroke/screenshot surveillance program called the Model Capability Initiative, and the expectation of further layoffs in August and autumn 2026. Zuckerberg is simultaneously recruiting AI researchers with packages up to $100 million for Meta Superintelligence Labs. The restructuring is part of a broader Big Tech trend: over 110,000 tech jobs have been cut in 2026 so far, with Microsoft, Oracle, and Amazon making similar moves.

Study finds 31% of developer time now consumed by AI validation work invisible to DORA metrics

A study of 700 engineering practitioners across five countries reveals that generative AI has created a significant 'invisible' workload for developers. While AI accelerates code production, 31% of a developer's day is now consumed by AI-related validation work — reviewing code accuracy, fixing subtle bugs, and explaining AI-generated code to teammates. Despite 81% of engineering leaders reporting increased code review time since deploying AI, only 38% of organizations track this overhead. The report highlights a disconnect between leadership and practitioners, with 94% agreeing that tech debt, validation time, and burnout are missing from current metrics like DORA. Five recommendations are offered: measure validation work, prioritize ship rate over code volume, audit measurement frameworks, plan for governance complexity, and build developer trust through clear data policies.

xAI promised employees $420 for tax data two months ago and still hasn't paid

Elon Musk's xAI offered employees $420 each to submit their personal US tax returns as training data for Grok ahead of the April 15 deadline. Two months later, Bloomberg reports the payments have not been made. The incident is framed against xAI's turbulent restructuring following its acquisition by SpaceX in February 2026, the departure of all eleven original co-founders, and ongoing layoffs. The missing payments raise questions about data-handling practices, employee trust, and the operational state of a company mid-rebuild. Regulators in Europe already scrutinizing xAI's data practices may view the unpaid compensation as an additional red flag.

Anthropic splits Claude subscription billing starting June 15 with separate API credit pool

<h2>What’s changing</h2>
<p>Starting June 15, Anthropic is separating programmatic Claude usage from standard subscription credits. Each paid plan gets a dedicated monthly credit pool for Agent SDK usage, <code>claude -p</code>, Claude Code GitHub Actions, and third-party apps built on the Agent SDK:</p>
<ul>
<li><strong>Pro</strong>: $20/month</li>
<li><strong>Max 5x</strong>: $100/month</li>
<li><strong>Max 20x</strong>: $200/month</li>
</ul>
<p>Credits don’t roll over, aren’t transferable, and are billed at API rates. If you hit the cap, you can either enable pay-as-you-go API billing or wait for the next cycle.</p>
<h2>Why developers are angry</h2>
<p>The credit amounts sound reasonable until you compare them to what was available before. Tools like T3 Code, Zed, and OpenCode previously accessed Claude through the Agent SDK — the only officially supported path for building on Claude — drawing from the same subscription pool that was effectively worth an estimated $5,000–$7,500/month in inference. Under the new structure, that same usage is capped at $200/month. That’s a 25–40x reduction.</p>
<p>Developer Theo, who built T3 Code as an open-source Claude wrapper, called it a betrayal after months of explicit encouragement from Anthropic to build on the SDK. He’s canceling his subscription and switching to Codex. He’s not alone — many developers who built automation workflows around the old flat-rate plans are now looking at 5–10x cost increases they didn’t budget for.</p>
<p>This also reverses a policy from just a month ago that had removed third-party tool coverage from subscriptions entirely, which itself followed a series of changes over the past year: weekly limits, silent cache TTL reductions, third-party harness blocks, and at least one billing bug that charged a user $200.98 in API fees due to a false-positive detection of third-party tool usage.</p>
<h2>The business logic behind it</h2>
<p>The official framing is billing simplification. The less charitable reading: Anthropic is routing valuable training-signal interactions through its own surfaces (<a href="http://Claude.ai" target="_blank" rel="noopener nofollow">Claude.ai</a>, Claude Code) while reclassifying third-party usage as “programmatic” and billing it at full API rates. This improves unit economics ahead of what’s reported to be a $950B funding round and eventual IPO, where public markets will demand actual margin.</p>
<p>Salesforce’s numbers illustrate the scale of the problem Anthropic is solving. Marc Benioff projects Salesforce will spend $300 million on Anthropic tokens in 2026, almost entirely on AI coding agents. That kind of consumption-based spend is what Anthropic wants to capture directly, not subsidize through flat-rate plans.</p>
<p>The broader industry is moving the same direction. GitHub Copilot is already shifting to usage-based billing in June 2026 because flat-fee models collapsed under agentic workloads. OpenAI is heading the same way. The current pricing was always a loss-leader — a knowledge worker on a $20/month Claude Pro plan can consume $200–$400 worth of compute monthly. That math doesn’t work at scale, and it especially doesn’t work when agents run autonomously overnight.</p>
<h2>What this means for teams using Claude in production</h2>
<p>The shift from flat-rate to metered billing introduces real budgeting complexity. Credits don’t pool across team seats, runaway agents can exhaust a month’s budget in hours, and cost forecasting requires tracking token consumption you probably weren’t measuring before.</p>
<p>The practical advice is straightforward: audit your actual token consumption now, model what your costs look like at API rates, and set hard budget alerts before June 15. If you’re running CI pipelines or long-running automations, treat them like cloud infrastructure — with the same cost controls you’d apply to EC2 instances.</p>
<p>It’s also worth building multi-vendor optionality. Local models like Qwen 3.6-27B and Gemma 4 have matured enough to handle many everyday coding tasks on consumer hardware. Claude is still the best agentic coding tool for hard reasoning tasks, but the terms of any cloud model relationship can change at any time. A locally hosted model’s access conditions are fixed once downloaded. That’s worth something.</p>
<h2>The bigger picture</h2>
<p>Anthropics’s Code with Claude event in early 2026 showed impressive technical progress — SWE-bench Verified scores improved from 62% to 87% over the past year, Q1 2026 revenue grew 80x versus a planned 10x, and the platform is adding managed agents, sandboxed execution, and cron-triggered routines. The product is genuinely getting better.</p>
<p>But the billing relationship with developers has gotten less predictable with each passing month. The June 15 change is probably not the last one. Enterprises and independent developers alike should plan accordingly.</p>


Anthropic is splitting Claude subscription billing starting June 15, creating separate monthly credit pools for programmatic usage (Agent SDK, claude -p, Claude Code GitHub Actions, third-party apps). Pro gets $20/month, Max 5x gets $100/month, and Max 20x gets $200/month in credits at API rates — a 25–40x reduction from what was previously available through the shared subscription pool. Developers who built tools on the Agent SDK, like T3 Code and OpenCode, are facing 5–10x cost increases. The change is framed as billing simplification but effectively routes third-party usage to full API pricing, improving unit economics ahead of a reported $950B funding round. Practical advice includes auditing token consumption now, setting budget alerts before June 15, and building multi-vendor optionality with local models for everyday tasks.

Cerebras IPO raises $5.55B at $95B valuation in largest US tech debut since Snowflake

Cerebras Systems debuted on Nasdaq at $311.07, up 68% from its $185 IPO price, reaching a ~$95 billion market cap and raising $5.55 billion — the largest US tech IPO since Snowflake in 2020. The AI chip company's success signals strong public market appetite for AI hardware exposure. However, attention is already shifting to the potential $3 trillion in combined IPOs from SpaceX (targeting $1.75T valuation, possibly listing June 12), OpenAI (targeting ~$852B in Q4 2026), and Anthropic (~$800B, possibly October). The combined fundraising demand of $150B+ raises questions about market liquidity and whether non-AI companies can compete for institutional capital in this environment. Cerebras itself benefited from a $20B multi-year inference contract with OpenAI, which resolved earlier customer concentration concerns around Abu Dhabi's G42.

BLS data shows AI-exposed US occupations shed jobs as overall employment grows

Several US occupations flagged by the Bureau of Labor Statistics as AI-exposed saw a 0.2% employment decline between May 2024 and May 2025, contrasting with 0.8% overall employment growth. The 18 affected occupations — covering roughly 10 million jobs and including customer service representatives, secretaries, and salespeople — recorded heavy job losses for a second consecutive year in 2025.