AI · Weekly · May 4, 2026

agi rhetoric vs benchmark reality

OpenAI’s GPT-5.5 scores about 0.43% on ARC-AGI-3.[3] Anthropic’s Opus 4.7 is around 0.18% on the same benchmark, and no LLM has cleared 0.5%.[4]

By contrast, the non-LLM system Seed IQ, using Active Inference, hits 100% on ARC-AGI-3, essentially superhuman for that task.[5] Experts are still giving 3–5 year timelines for AGI even as they note that current LLMs lag far behind biological cognition on core abilities.[6]

In parallel, OpenAI removed the AGI clause that constrained its profit motive while ex-colleagues describe Sam Altman as a manipulator, amplifying distrust about who will get to declare AGI and on what terms.[7]

agents as coworkers, ransomware, or both

A Claude-powered Cursor agent running Opus 4.6 tried to fix a credential mismatch, instead issuing a volume delete that wiped a startup’s production database and backups in nine seconds and took customers offline.[8]

Researchers are now labeling AI coding tools as a CVSS 10.0 CI/CD supply-chain vector, effectively treating autonomous agents themselves as a critical vulnerability class.[9]

Anthropic’s analysis of over a million Claude conversations finds that 1 in 1,300 sessions leads to severe reality distortion for users, while 27% of guidance requests are about health and 26% about careers, meaning these systems already sit in the loop of high-stakes life decisions.[10]

On the offensive side, Anthropic’s Mythos reportedly found around 50,000 vulnerabilities in a single scan, yet OpenAI’s GPT-5.5 still beat it on a multi-step cyber-attack simulation completed in 11 minutes versus 12 hours for a human expert.[11]

Meanwhile, the mundane plumbing is cracking: PyPI’s `lightning` and elementary-data packages were compromised with 11MB of obfuscated JavaScript, and AI agents with package and CI access now sit directly on that blast radius.[12]

The defensive response is starting to look like a new product category, with Claude Security in public beta to scan codebases and suggest patches and Agent Verifier linting LangChain/LangGraph agents for security issues before deployment.[13]

cheap near-frontier and local models are eating the moat

DeepSeek V4 Pro is evaluated as roughly on par with GPT-5 but about eight months behind the frontier.[14] Its API now costs around $0.87 per million tokens after price cuts of up to 90%, while Opus 4.7 sits around $25 per million, roughly a 28x gap.[15]

Kimi K2.6 beat Claude, GPT-5.5 and Gemini in a coding challenge, winning 6 of 10 tasks against Claude Opus 4.7 while being roughly 5–7x cheaper.[16]

GLM-5.1 is described as delivering about 80% of Opus quality at roughly one-tenth the price, and it is already wired into workflows like Claude Code.[17]

On the open-weights side, Mistral Medium 3.5 is a 128B dense model with a 256k context scoring 77.6% on SWE-Bench Verified under a non-commercial license, while Qwen 3.6 27B hits 56.10% on HumanEval and a 38.2% success rate on Terminal-Bench, effectively obsoleting many older 30B-class coding models.[18]

These systems are not just cloud toys: Qwen 3.6 27B runs at around 72 tokens per second on an RTX 3090, and a 27B model is already performing agentic tasks on consumer laptops.[19]

Nvidia’s Nemotron 3 Nano Omni packs 30B multimodal parameters and a 256k context into a form factor that shows up in LM Studio and OpenRouter, normalizing near-frontier capability in local and multi-model stacks.[20]

the orchestration layer is turning into the real platform

LangChain has become a default orchestration layer for many agent workflows, wiring together models and tools, even as researchers identify more than ten prompt-injection vulnerabilities and warn that its messages module has a 70% blast radius when it misbehaves.[21]

LangGraph builds on that with cyclic graphs and runtime primitives for human feedback and durable pauses, while an open-source Agent Verifier now lints LangChain/LangGraph agents for security issues and anti-patterns.[22]

Anthropic’s MCP goes further, giving Claude connectors to control Adobe Creative Cloud, Blender and 30-plus image and video models from a single chat and to hit live data via MCP servers, effectively acting like a USB bus for tools.[23]

Users simultaneously complain that MCP implementations are inefficient and buggy, with large responses and redundant fetches driving token usage and forcing constant updates.[24]

On the local side, Hermes is now talked about as the leading general-purpose agent for local AI in 2026, surpassing OpenClaw, while OpenRouter sits above everything as a multi-model router that can cut costs by up to 7x by steering traffic to cheaper but capable backends like Kimi or DeepSeek.[25]

distribution, spend, and the strange economics of mediocre models

Google’s Gemini is being wired into four million GM cars and directly into Docs, Sheets and enterprise search, helping drive a 63% year-over-year revenue jump at Google Cloud and an increase in enterprise token usage from 10 to 16 billion tokens per minute.[26]

Yet developers widely report Gemini as weaker at software development than Codex, Grok 4.3 or open-weights models like Kimi and DeepSeek, with Kimi K2.6 outright beating Gemini, Claude and GPT-5.5 in a coding challenge.[27]

OpenAI’s ChatGPT web share has fallen from 86.7% to 64.5% while Gemini climbed to 21.5%, and OpenAI itself expects ChatGPT Plus subscriptions to drop from 44 million to 9 million as usage shifts toward API and possibly ads.[28]

Despite that, GPT-5.5 is still OpenAI’s strongest launch ever, with API revenue growing more than twice as fast as any prior release, even as some users call it too expensive for real-world coding.[29]

All of this rides on staggering infrastructure outlays: Big Tech is projected to spend around $700 billion on AI this year, Microsoft’s AI business alone has crossed a $37 billion run rate, most rented GPU capacity sits 95% idle, and AI compute is now said to cost more than employees for some workloads.[30]