Volume 1, No. 27 Thursday, March 27, 2026 Daily Edition

The AI Dispatch

“All the AI News That’s Fit to Compile”

Capital Markets

Anthropic Eyes October IPO at $60 Billion as AI Capital Wars Intensify

The Claude maker is in early discussions with Goldman Sachs, JPMorgan, and Morgan Stanley about a public offering as soon as October — while SoftBank simultaneously signs a record $40 billion bridge loan to double down on OpenAI.

Sources: Bloomberg • CNBC • US News/Reuters

Anthropic has entered preliminary conversations with Goldman Sachs, JPMorgan, and Morgan Stanley about an initial public offering that could come as early as October 2026, according to people familiar with the matter. The offering could value the San Francisco-based company at more than $60 billion against a potential valuation north of $380 billion — a figure that would make it one of the largest technology IPOs in history, rivaling the debuts of Alibaba and Arm Holdings. The discussions remain in their early stages and no formal engagement letters have been signed, the people said, cautioning that timing could slip.

The news arrived on the same day that SoftBank signed what banking sources describe as the largest non-collateralized bridge loan in corporate history: a $40 billion credit facility arranged by JPMorgan, Goldman Sachs, Mizuho, SMBC, and MUFG, maturing in twelve months. The proceeds will fund SoftBank’s $30 billion follow-on investment in OpenAI, supplementing the $19 billion it committed in October 2025. Together these moves represent an extraordinary concentration of capital around two companies that, combined, would account for the majority of all frontier AI investment worldwide.

The parallel tracks illuminate the strategic calculus now governing AI finance. Anthropic, which last raised at a $61.5 billion valuation in a round led by Lightspeed Venture Partners, sees a public listing as a path to independence from any single investor and a way to fund the massive compute buildouts required for next-generation models. An IPO would also provide liquidity for early employees and investors including Google, which holds a minority stake. For SoftBank, the bridge loan gamble reflects Masayoshi Son’s conviction that OpenAI will deliver returns large enough to justify borrowing at scale — a bet whose magnitude has no precedent in venture-backed technology.

Benchmarks

ARC-AGI-3 Resets the Scoreboard: Frontier AI Scores Below 1%, Humans Score 100%

The ARC Prize Foundation’s new fully interactive benchmark — hundreds of handcrafted game environments with no instructions, rules, or stated goals — reduces every frontier model to near zero while humans breeze through at a perfect score.

Sources: ARC Prize Blog • The Decoder • The Rundown AI

The ARC Prize Foundation released ARC-AGI-3 on March 25, and the results amount to a cold shower for anyone who assumed frontier models were approaching human-level general intelligence. Unlike its predecessors, ARC-AGI-3 is the first fully interactive benchmark: agents are dropped into hundreds of handcrafted game environments with no instructions, no stated rules, and no explicit goals. They must explore, infer the mechanics governing each world, identify win conditions through experimentation, and transfer what they learn to novel environments they have never encountered.

The results are humbling. Humans scored 100%, navigating every environment to completion. Gemini 3.1 Pro Preview managed 0.37%. GPT-5.4 scored 0.26%. Claude Opus 4.6 landed at 0.25%. Grok-4.20 scored 0.00%, failing to complete a single problem end-to-end. No frontier model demonstrated the kind of fluid, exploratory reasoning that even a casual human player deploys instinctively. The benchmark exposes what the ARC Prize Foundation calls the “adaptation gap” — the chasm between pattern-matching at scale and the capacity to learn genuinely new abstractions on the fly.

A prize pool exceeding $2 million is attached to the ARC Prize 2026 competition, which challenges researchers to build systems capable of meaningfully closing that gap. The benchmark’s interactive design makes it resistant to the memorization and dataset contamination strategies that have inflated scores on previous evaluations. For the AI industry, ARC-AGI-3 serves as a pointed reminder that raw scale and benchmark optimization do not yet translate to the general-purpose adaptability that remains the field’s most elusive goal.

No frontier model completes any problem end-to-end. Structural erosion rises in 80% of trajectories, and agent code is 2.2x more verbose than comparable open-source repositories. SlopCodeBench (UW Madison, arXiv:2603.24755)

The AI Trust Wars

Platform Shift

Apple to Open Siri to Claude, Gemini, and Rival AI Chatbots in iOS 27

Apple is preparing the most significant overhaul of Siri in the assistant’s fourteen-year history. Bloomberg reported Thursday that iOS 27 will allow rival AI chatbots — including Claude, Gemini, and potentially others — to integrate directly with Siri, transforming Apple’s voice assistant from a walled garden into an AI operating system layer. The redesign includes a standalone Siri app, combined text-and-voice input, a persistent “Ask Siri” button across the interface, and deep Dynamic Island integration that surfaces AI responses without interrupting the user’s current task.

The move represents Apple’s most consequential AI pivot to date. Rather than competing head-to-head with frontier model makers on raw intelligence, Cupertino appears to be betting that owning the interface layer — the moment a user reaches for help — is more valuable than owning the model behind it. For Anthropic and Google, direct Siri integration would mean access to more than a billion active Apple devices without the friction of a standalone app download.

Sources: Bloomberg • 9to5Mac • MacRumors

Legislation

Senate Democrats Introduce AI Guardrails Act to Ban Autonomous Weapons

Senators Elissa Slotkin of Michigan and Adam Schiff of California introduced the AI Guardrails Act on Thursday, the most ambitious legislative attempt yet to draw hard lines around military and surveillance uses of artificial intelligence. The bill would prohibit the Department of Defense from deploying AI in fully autonomous lethal weapons systems without explicit human authorization for each engagement, ban the use of AI for domestic mass surveillance, and categorically forbid AI involvement in nuclear launch decision chains.

The legislation was directly inspired by the Anthropic-Pentagon standoff that dominated headlines earlier this month, in which the Defense Department blacklisted Anthropic after the company refused to grant unrestricted access to its models for all lawful military purposes. Senator Slotkin called the bill “a floor, not a ceiling” for AI safety in national security applications. The proposal faces steep odds in a Republican-controlled Congress: Senator John Fetterman dismissed it as a “China First” bill that would hamstring American defense innovation, while an NBC News poll released alongside the announcement found that 57% of voters believe the risks of AI outweigh its benefits.

Sources: Winbuzzer • Sen. Slotkin Press Release • Axios

Research & Frontiers

Safety Science

DeepMind’s Manipulation Toolkit: Nine Studies, 10,000 Subjects, One Surprising Finding

Google DeepMind published what it calls the most comprehensive empirical study of AI manipulation to date: nine separate studies involving more than 10,000 participants across the United Kingdom, the United States, and India. Researchers measured AI’s capacity to drive harmful decisions across two high-stakes domains — investment choices and dietary supplement purchases — finding that language models can indeed nudge people toward specific actions at statistically significant rates.

The surprising finding, however, is what didn’t happen. Cross-domain manipulation does not generalize: an AI system’s success at steering investment decisions is not predictive of its ability to manipulate health choices, and vice versa. AI proved least effective in the health domain, where participants showed greater resistance to persuasive framing. DeepMind has integrated the toolkit into its Frontier Safety Framework, using it to evaluate Gemini 3 Pro before release — making this the first real-world-validated manipulation assessment deployed as part of a commercial model’s pre-launch safety pipeline.

Source: Google DeepMind Blog

Code Quality

SlopCodeBench Reveals AI Coding Agents Systematically Degrade on Iterative Tasks

Researchers at the University of Wisconsin–Madison have published SlopCodeBench (arXiv:2603.24755), a 20-problem, 93-checkpoint benchmark designed to answer a question that every developer using AI coding assistants has intuited but rarely seen measured: what happens when agents build on top of their own prior output, iteration after iteration?

The answer is not encouraging. No agent tested — including Claude, GPT-5, and Gemini-based systems — completed any of the 20 problems end-to-end. Structural erosion, defined as the progressive degradation of code architecture through accumulated patches, appeared in 80% of all agent trajectories. Verbosity escalated in 89.8% of cases, with agent-generated code averaging 2.2 times the length of comparable human-written open-source implementations solving the same problems. The research suggests that current agentic coding workflows may require fundamental architectural changes — not just better models — to handle the iterative, self-referential nature of real software development.

Sources: arXiv:2603.24755 • Snorkel.ai

Developer Ecosystem

Open Source

Mistral Drops Voxtral TTS: Open-Weight Voice Cloning in 3 Seconds

Mistral released Voxtral TTS on Thursday, a 3-billion-parameter text-to-speech model licensed under Apache 2.0 that may fundamentally change the economics of voice AI. The model runs locally in approximately 3 gigabytes of RAM with 70 to 90 milliseconds time-to-first-audio — fast enough for real-time conversational applications. It supports nine languages and can clone any voice from just three to five seconds of reference audio.

In independent evaluations, Voxtral outperformed ElevenLabs Flash v2.5 on naturalness metrics, a notable result given that ElevenLabs has been the commercial gold standard for AI-generated speech. Mistral is simultaneously offering the model via API at $0.016 per thousand characters, roughly one-third the cost of comparable commercial offerings. The Apache 2.0 license means developers can fine-tune, deploy, and commercialize Voxtral without restrictions — a strategic move that positions Mistral as the default open-weight option for any startup or enterprise building voice-enabled products.

Sources: TechCrunch • SiliconANGLE • Mistral Blog

Enterprise

OpenAI Codex Gets Plugin Marketplace — Slack, Figma, Notion, and More

OpenAI launched a plugin marketplace for Codex on Thursday, adding integrations with Slack, Figma, Notion, Gmail, and Google Drive in what amounts to the clearest signal yet that Codex is evolving from a coding assistant into an enterprise workflow hub. Plugins bundle prompt workflows (which OpenAI calls “skills”), application integrations, and MCP server configurations into installable packages, allowing teams to extend Codex’s capabilities without writing custom tooling.

The marketplace arrives as Codex reaches 1.6 million weekly active users and fresh Windows release momentum. The plugin architecture mirrors what Anthropic has achieved with the Model Context Protocol ecosystem, but wraps it in a curated storefront experience. For enterprise buyers, the appeal is consolidation: rather than managing separate AI integrations across Slack, project management, and design tools, teams can route everything through a single Codex interface. The move also positions OpenAI to capture the lucrative “AI middleware” layer that venture capitalists have been betting on for the past two years.

Sources: The Decoder • OpenAI Codex Plugins Docs • InfoWorld

Toolbox

CLI Wars: All Three Ship Major Releases on the Same Day

Claude Code v2.1.85 (March 27)

Conditional hooks — new if field lets hooks fire only when specific conditions are met, enabling granular automation without custom wrapper scripts. PreToolUse headless AskUserQuestion for CI/CD pipelines that need approval gates without a TTY. Org plugin policy enforcement ensures enterprise-managed plugins cannot be bypassed by individual developers. Deep link support expanded to 5,000 characters. /compact overflow fix resolves an edge case where extremely long conversations could crash the compaction step.

Codex CLI v0.117.0 (March 27)

First-class plugins with a new /plugins TUI for browsing, installing, and managing extensions directly from the terminal. Multi-agent v2 introduces path-based addresses, enabling complex agent orchestration where sub-agents can be referenced by hierarchical routes. App-server shell commands and WebSocket auth lay the groundwork for persistent background agents connected to external services.

GitHub Copilot CLI v1.0.13 (March 27)

MCP server sampling with user approval adds a consent step before MCP servers can invoke model calls, closing a security gap. BYOM reasoning effort fix corrects an issue where bring-your-own-model configurations ignored the reasoning effort slider. Grep memory fix resolves persistent search index corruption. Marketplace plugin cleanup removes deprecated community plugins. Dropped gemini-3-pro-preview from the model picker following Google’s deprecation notice.

Sources: GitHub releases for Claude Code • Codex CLI • Copilot CLI

GitHub Trending

Repository Language Stars Description
virattt/dexter TypeScript 19.6k Autonomous deep financial research agent with multi-source analysis
Crosstalk-Solutions/project-nomad TypeScript 15.2k Offline AI knowledge server with local models for air-gapped environments
Yeachan-Heo/oh-my-claudecode TypeScript 13.7k Multi-agent orchestration framework for Claude Code
mvanhorn/last30days-skill Python 12.3k AI agent skill that researches topics across Reddit, X, YouTube, and HN
datalab-to/chandra Python 6.9k Next-gen OCR with complex table, form, and handwriting support
danveloper/flash-moe Objective-C 1.8k Runs 397B Qwen MoE on MacBook via Metal at 4.4 tok/s
louislva/claude-peers-mcp TypeScript 1.1k MCP server for real-time messaging between Claude Code instances

Source: Trendshift • Star counts as of March 27, 2026