Volume 1, No. 43 Tuesday, April 15, 2026 Daily Edition

The AI Dispatch

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


Controversy

Anthropic Faces Backlash Over “Nerfed” Claude Performance

Power users and developers report that Claude increasingly fails to follow complex instructions and takes shortcuts — traced to a quiet change reducing the model’s default effort level to conserve tokens.

Anthropic is facing a growing backlash from its most dedicated users after reports surfaced that the company quietly reduced Claude’s default “reasoning effort” level to medium — a change that conserves tokens and lowers compute costs but leaves many complex tasks half-finished. Developers who rely on Claude for multi-step coding workflows, document analysis, and agentic pipelines say the model now frequently takes shortcuts it never used to, skipping edge cases, truncating outputs, and failing to follow detailed instructions that it previously handled with ease. The complaints, which first gained traction on Reddit and Hacker News before being picked up by Fortune and VentureBeat, paint a picture of a model that feels noticeably “dumber” to the people who use it most intensively — even as Anthropic’s marketing continues to tout Claude’s benchmark-leading performance.

What has turned frustration into genuine anger is the lack of transparency. Anthropic never publicly announced the effort-level change, and users had to discover it through trial and error or by noticing that explicitly setting the effort parameter to “high” restored the behavior they had come to expect. The pattern — degrading default performance while leaving a hidden toggle for those who know where to look — strikes many as antithetical to the trust-first brand that Anthropic has carefully cultivated. Critics on social media have drawn comparisons to the “shrinkflation” tactic common in consumer goods: same packaging, less product inside. Several prominent developers have publicly questioned whether the company is prioritizing margin optimization ahead of a rumored IPO over the user experience that built its reputation.

The timing is particularly awkward. Anthropic is widely reported to be preparing for a public offering, and the controversy raises uncomfortable questions about whether the compute constraints behind the effort-level change reflect genuine infrastructure limitations or a deliberate choice to improve unit economics before going to market. In a company blog post that addressed the issue only obliquely, Anthropic said it “continuously tunes default parameters to balance quality and efficiency” and pointed users to the API documentation for adjusting effort levels. For a company that has built its identity on being the responsible, transparent alternative to OpenAI, the evasive response has done little to quell the growing sense that Anthropic’s actions are diverging from its stated values.

Developer Tools

OpenAI Updates Agents SDK with Sandboxing and Long-Horizon Harness

OpenAI shipped a major update to its Agents SDK on April 15, introducing three capabilities that enterprise builders have been requesting for months: sandboxing, a long-horizon harness for multi-step agentic workflows, and subagent spawning for parallel task decomposition. The sandboxing feature isolates each agent in its own workspace with controlled file-system access and network permissions, addressing the security concerns that have kept many regulated industries from deploying agentic systems in production. The long-horizon harness manages state persistence and checkpoint recovery across workflows that can run for hours or days, handling the messy reality of network failures, rate limits, and model timeouts that plague long-running agent tasks.

Perhaps most significantly, the SDK now supports more than 100 non-OpenAI LLMs via any Chat Completions-compatible API — a tacit acknowledgment that enterprise customers want framework portability rather than vendor lock-in. The subagent spawning system allows a parent agent to decompose complex tasks into parallel subtasks, each running in its own sandbox, with results aggregated back to the parent. Enterprise AI now accounts for roughly 40% of OpenAI’s revenue, and the SDK update reflects the company’s aggressive push to own the agentic infrastructure layer regardless of which models customers ultimately choose.

Public Sentiment

Public Turns Against AI and Data Centers Just as IPOs Loom

A CNBC analysis published April 15 documents a sharp turn in public opinion on artificial intelligence that threatens to collide with the industry’s most important financial milestone in years. More than $156 billion in data center projects have been cancelled or delayed across the United States as local opposition intensifies, with 142 activist groups across 24 states now organizing against AI infrastructure expansion. The concerns span water usage, power grid strain, noise pollution, and property value impacts — bread-and-butter local issues that have proven far more effective at blocking projects than abstract debates about existential risk ever were.

The timing could not be worse for OpenAI and Anthropic, both of which are preparing IPO filings. Public companies face a level of scrutiny that private ones can largely avoid, and a hostile public environment raises the risk of regulatory intervention at the state and municipal level that could constrain expansion plans investors are counting on. OpenAI policy chief Chris Lehane, speaking at a recent industry event, warned that both “boomers” — those who promise AI will solve everything — and “doomers” — those who predict civilization-ending catastrophe — are failing to offer practical solutions to the concrete, local impacts that are actually driving opposition. It is the mundane politics of zoning boards and water rights, not the grand narratives of either camp, that may ultimately determine how quickly the AI buildout proceeds.

Frameworks

Google ADK Reaches 1.0 Across Java, Go, and TypeScript

Google’s Agent Development Kit has hit version 1.0 for Java and Go, with TypeScript joining the stable release lineup. The Java SDK adds native Agent2Agent (A2A) protocol support for cross-framework agent interoperability, Human-in-the-Loop approval workflows, and event compaction for reducing context window bloat in long-running sessions. The Go SDK ships with YAML-based agent configuration, OpenTelemetry tracing out of the box, and full A2A interop — making it the first Go framework to support the emerging agent communication standard. Three separate posts on the Google Developers Blog detail the releases, signaling that Google views multi-language agent infrastructure as a strategic priority rather than a Python-only affair.

Open Source

HuggingFace Ships Transformers v5

HuggingFace has released Transformers v5, a significant architectural update to the library that underpins much of the open-source AI ecosystem. The release simplifies model definitions with a cleaner internal API, promotes Mamba and hybrid cache architectures to first-class citizens, and adds support for new multimodal and OCR models including PP-OCRv5, SLANeXt, and VidEoMT. Separately, HuggingFace Buckets — persistent storage for Spaces — has reached general availability, solving a long-standing pain point for developers who needed durable file storage in their deployed applications without wiring up external cloud services.

Society

Pause AI and Stop AI Face Hard Questions After Firebombing

A Fortune investigation has examined the organized anti-AI movements in the wake of recent violent incidents, exposing the uncomfortable gap between groups advocating lawful civil disobedience and the lone-actor attackers who claim similar motivations. Pause AI and Stop AI, the two most prominent advocacy organizations, have publicly condemned violence while struggling to distance themselves from individuals who cite their rhetoric. The investigation reveals a sharp generational divide: younger social media users are significantly more sympathetic to the fears that motivated the attacker, even when they reject the methods, raising difficult questions about radicalization pipelines that movement leaders say they are only beginning to grapple with.

Research

HarmonyGNN Sets New Accuracy Records on Graph Neural Networks

Researchers at NC State University have introduced HarmonyGNN, a self-supervised training framework for graph neural networks that handles graphs with mixed node types — a longstanding challenge in the field. The architecture uses a teacher-student design with partial graph masking: the teacher network sees the full graph structure while the student learns from deliberately incomplete views, forcing the model to develop robust representations that generalize across both homophilic graphs (where similar nodes connect) and heterophilic ones (where dissimilar nodes connect). The distinction matters because most real-world graphs — social networks, molecular structures, weather systems — contain a mixture of both patterns, and existing methods tend to excel at one type while struggling with the other.

HarmonyGNN matched state-of-the-art performance on seven standard homophilic benchmarks while setting new records on four heterophilic benchmarks, with accuracy gains ranging from 1.27% to 9.6% over the previous best methods. The practical applications span drug discovery (predicting molecular interactions across different compound types), weather forecasting (modeling atmospheric systems where neighboring regions may have very different conditions), and materials science. The paper was accepted at ICLR 2026, and the code has been released under an open-source license.

GitHub Trending

Repo Language Stars Description
NousResearch/hermes-agent Python 91.6k (+53k/wk) Self-evolving AI agent framework
forrestchang/andrej-karpathy-skills Markdown 47.9k (+30.9k/wk) CLAUDE.md for better Claude Code
thedotmack/claude-mem TypeScript 59k (+10.7k/wk) Session recorder with AI compression
vercel-labs/open-agents TypeScript 3k (+735/day) Cloud-native agent system foundation
milla-jovovich/mempalace Python 23.9k (new) Memory palace architecture for AI agents
addyosmani/agent-skills TypeScript 6.7k Production-grade skills library for coding agents
microsoft/markitdown Python +15.7k/wk Convert Office docs to Markdown
Toolbox

This Week in AI Coding Tools

Three major CLI coding agents shipped updates on the same day — a sign of just how competitive the agentic coding space has become.

Claude Code v2.1.110 (Apr 15)

OpenAI Codex CLI v0.121.0 (Apr 15)

GitHub Copilot CLI v1.0.27 (Apr 15)