📰

subtl daily briefing

Saturday, April 25, 2026

Good morning. The AI arms race hit a new milestone this week as OpenAI shipped GPT-5.5 with full agentic capabilities — and somehow still got lapped in valuation by Anthropic, which crossed $1 trillion on secondary markets. Meanwhile, Meta is cutting 8,000 jobs to fund AI tokens, a DeFi hack wiped $20B in TVL, and a U.S. soldier learned the hard way that insider trading on prediction markets is still insider trading.

In today's briefing

1.GPT-5.5 Launches, Anthropic Hits $1 Trillion
2.Meta's 8,000 Layoffs Fund the AI Bet
3.Kelp DAO Hack Wipes $20B in DeFi TVL
4.Bitcoin Surges 24% Amid U.S.-Iran War
5.Supply Chain Attacks Target Developer Credentials
⚡Quick hits on other news

Latest Developments

🤖GPT-5.5 Is Live — But Anthropic Just Hit $1 Trillion First

The Rundown: OpenAI launched GPT-5.5 with advanced agentic capabilities and $5/M token pricing, but Anthropic stole the valuation headlines by surpassing OpenAI itself at a $1 trillion valuation on Forge Global.

The details:

●GPT-5.5 scores 82.7% on Terminal-Bench 2.0, ships with a 1M token context window, $5/M input token pricing, and is live now in ChatGPT and Codex
●OpenAI Workspace Agents automate Slack and team workflows for free until May 6, 2026 for Business, Enterprise, Edu, and Teachers plans
●Anthropic crossed $1 trillion in secondary market valuation driven by Claude Code adoption and scarce share availability — while Cognition AI (maker of Devin) is in talks to raise at a $25B valuation
●DeepSeek V4-Pro also launched this week with 1.6 trillion parameters and a 1M token context window, confirmed to run on both Nvidia and Huawei chips — and the company is in talks to raise at a $20B valuation backed by Tencent and Alibaba

Why it matters: For founders and operators, this week crystallized a new AI power hierarchy: OpenAI ships product, Anthropic wins the valuation war, and Chinese labs commoditize both at fraction of the cost. The real signal for builders is that agentic AI is no longer experimental — GPT-5.5's native ability to operate across email, spreadsheets, calendars, and Slack means the first wave of workflow automation products is about to face serious platform-level competition. If your SaaS product's core value is task automation, the clock is ticking: one newsletter reported a Claude integration replicating 95% of a vendor's AI features at 15% of token cost, forcing a 45% SaaS renewal cut.

📰 Source: TLDR, AlphaSignal, Techpresso

Share𝕏 in

🏭Meta Cuts 8,000 Jobs to Fund Its AI Token Economy

The Rundown: Meta is eliminating 8,000 roles and canceling 6,000 open positions to redirect capital toward up to $135 billion in AI infrastructure, making the clearest corporate statement yet that AI tokens are being treated as a direct substitute for human labor.

The details:

●Meta layoffs affect ~10% of staff with cuts going into effect May 20, with savings earmarked for AI infrastructure spending projected at up to $135B this year
●Google reports 75% of its code is now AI-written, and its API saw 60% token growth in Q1 driven by agentic AI adoption
●S&P 500 companies have reported only $300M in aggregate AI productivity gains — a figure Anthropic earns roughly every 1.5 days in token sales alone
●A 'tokenmaxxing' trend is emerging at Meta, Microsoft, and Salesforce where employees wastefully burn AI tokens for internal leaderboard rankings, causing financial waste and system outages

Why it matters: The Meta layoff is a watershed moment for how investors and boards will evaluate headcount going forward. When a company the size of Meta explicitly frames firing people as 'reinvesting in AI tokens,' it gives permission to every CFO and board member to ask the same question about their own org. For founders, this is both an opportunity (sell the AI-first workflow tools that justify these trade-offs) and a warning (any product that just wraps human labor in SaaS clothing is increasingly exposed). The fact that measured productivity gains across the S&P 500 are still tiny compared to what AI companies are earning from token sales suggests the real ROI is still being captured at the infrastructure layer — not yet at the application layer.

📰 Source: Techpresso, The Breakdown, MIT Technology Review

Share𝕏 in

Crypto

💥The $293M Kelp DAO Hack Just Stressed-Tested All of DeFi

The Rundown: A single-validator bridge exploit on Kelp DAO triggered $293M in losses and roughly $20B in DeFi TVL withdrawals, prompting JPMorgan and Jefferies to warn institutions away from open DeFi — while a community recovery fund scrambles to fill the 163,200 ETH hole.

The details:

●The Kelp DAO exploit on April 18 targeted a single-validator bridge with no collateral concentration limits, draining $293M and triggering a cascade of ~$20B in total DeFi TVL losses
●The DeFi United recovery fund has filled 73,700 ETH of the 163,200 ETH gap, with a new TokenLogic proposal to contribute 25,000 ETH from Aave's treasury
●JPMorgan and Jefferies issued warnings to institutional clients about open DeFi's incompatibility with traditional risk frameworks following the hack
●On the institutional adoption front, Mizuho, Nomura, and JSCC launched a JGB tokenization proof-of-concept on Canton Network for 24/7 real-time collateral management, while DoorDash went live on Tempo for stablecoin payouts across 40+ countries

Why it matters: This hack is a stress test that exposed the gap between DeFi's ambitions and its operational maturity. Single points of failure in bridge validators, no incident-response playbooks, and no pre-funded loss-absorption waterfalls are the exact things that keep institutional capital on the sidelines. The JPMorgan and Jefferies warnings will have real chilling effects on enterprise DeFi adoption timelines. For builders, the path forward is clear: adopt TradFi-style controls — multi-verifier requirements, published response frameworks, waterfalls — or watch permissioned networks like Canton capture the institutional market while open DeFi fights over retail. The DoorDash and Japan tokenization news shows the stablecoin and tokenization narrative is very much alive, just on more controlled rails.

📰 Source: Converge by The Defiant, The Defiant

Share𝕏 in

Crypto

₿Bitcoin Surges 24% as U.S.-Iran War Reshapes Crypto's Safe-Haven Narrative

The Rundown: Bitcoin has climbed above $79,000 — up 24% since the U.S.-Iran war began — with $4.5B in spot ETF inflows and declining exchange balances painting a bullish institutional accumulation picture.

The details:

●Bitcoin surpassed $79,000, up 24% since the U.S.-Iran war started, defying expectations of a risk-off selloff
●Spot Bitcoin ETFs recorded $4.5B in total net inflows since the conflict began, with 8 of the last 9 weeks showing positive flows
●Crude oil retreated below $100 after peaking near $120, removing near-term stagflation pressure that had threatened risk assets
●GSR launched the Crypto Core3 ETF (BESO) on Nasdaq offering combined BTC/ETH/SOL exposure with staking reward pass-through, while NY AG Letitia James sued Coinbase and Gemini for operating unlicensed prediction markets

Why it matters: The geopolitical playbook for Bitcoin is being rewritten in real time. The asset is no longer just a speculative tech trade — $4.5B in ETF inflows during active conflict suggests a growing segment of institutional capital is genuinely treating BTC as a macro hedge. For crypto founders and investors, declining exchange balances combined with sustained ETF demand is a structurally bullish signal for medium-term prices. The regulatory headwinds — NY AG suing Coinbase and Gemini, Wisconsin suing prediction market operators — are real but increasingly look like political noise against a backdrop of capital that keeps flowing in regardless.

📰 Source: Milk Road, TLDR

Share𝕏 in

Security

🔐AI Agents Now Hack Cloud Environments End-to-End — And npm Worms Are Back

The Rundown: A week of escalating supply chain and agentic security threats saw the Bitwarden CLI npm package compromised, an AI multi-agent system autonomously breach a GCP environment from recon to data exfiltration, and a self-propagating npm worm target developer credentials across PyPI.

The details:

●The Bitwarden CLI npm package was briefly compromised via a malicious Checkmarx GitHub Action, targeting npm tokens, GitHub tokens, SSH keys, and cloud credentials — requiring immediate rotation
●The Zealot AI multi-agent system was demonstrated autonomously breaching a GCP environment — from network recon through SSRF exploitation to BigQuery data exfiltration — with zero human intervention
●A worm-like supply-chain attack on Namastex Labs npm packages is stealing developer credentials, API keys, and crypto wallets while self-propagating to PyPI
●Microsoft issued an emergency patch for CVE-2026-40372 in ASP.NET Core on macOS/Linux, requiring admins to upgrade, rotate the DataProtection key ring, and revoke all long-lived tokens

Why it matters: The Zealot demonstration is the most important security story of the week and possibly the month. When an AI agent can independently run a full kill chain — recon, exploit, pivot, exfiltrate — the threat model for cloud infrastructure fundamentally changes. This isn't a future risk; it's a present capability. For engineering and security teams, the immediate to-do list is: audit your npm and PyPI dependencies today, rotate any credentials that touched the Bitwarden or Namastex packages, patch the ASP.NET Core vulnerability, and start modeling your cloud environments against autonomous agent adversaries, not just human ones. The era of 'AI-assisted' attacks is giving way to fully automated ones.

📰 Source: TLDR InfoSec, TLDR

Share𝕏 in

⚡

Everything else in the news today

Z.ai released GLM-5.1, a 754B parameter open-weights MoE model that runs autonomous coding loops for up to 8 hours, tops the Artificial Analysis Intelligence Index among open models, priced at $1.40/$4.40 per million input/output tokens, MIT licensed on HuggingFace→

Qwen3.6-27B beats Alibaba's own 397B model on SWE-bench coding (77.2 vs 76.2), runs on just 18GB VRAM, Apache 2.0 licensed→

Andrew Ng ranks coding agent acceleration by task: frontend fastest, then backend, infrastructure, and research least accelerated→

Kubernetes v1.36 shipped with 70 enhancements including fine-grained kubelet API authorization, user namespaces, and deprecation of the vulnerable externalIPs field→

A U.S. soldier faces 5 felony charges for using classified intel on Maduro's capture to earn $400k on Polymarket — the first insider trading case involving a prediction market→

Project Eleven awarded quantum cryptographer Giancarlo Lelli 1 BTC for cracking a 15-bit elliptic curve key, raising concerns about Bitcoin's long-term post-quantum security→

Stripe and DoorDash moved stablecoin payments into production on Tempo, a Stripe/Paradigm-incubated chain processing $10B+ annualized volume with sub-second finality and no native token requirement→

Demand Curve identifies a new archetype called the Gen Marketer — a senior strategist who uses AI to both strategize and execute — arguing AI is solving the long-standing 'first growth hire' problem for startups→

AI systems now apply a 'bland tax' filtering generic content from AI search results — brands without distinct identity are effectively invisible to AI-powered discovery→

Anthropic's Mythos cybersecurity model surfaced thousands of vulnerabilities across major OSes and browsers, prompting Australia to engage Anthropic over concerns the tool could accelerate sophisticated attacks→

Instagram is testing a standalone app called Instants for unedited, once-viewable disappearing photos to compete with Snapchat and BeReal→

Apple is rumored to launch a foldable iPhone Ultra capable of running two apps side-by-side, alongside the iPhone 18 lineup→

ChatGPT Images 2.0 is now live across all major Figma products including Design, FigJam, Slides, and Weave→

MIT researchers introduced a recursive model framework supporting 10 million token context, signaling that context length is effectively a solved problem→

Band emerged from stealth with $17M to build a universal orchestration layer enabling AI agents across different frameworks and clouds to discover, delegate, and collaborate in real time→