AI · Daily · April 28, 2026

the quiet death of 'AGI premium'

DeepSeek‑V4 cuts API prices by up to 90% while still advertising 'near SOTA' performance versus Opus 4.7 and GPT‑5.5.[1] Kimi K2.6 is reported about 7x cheaper than Claude Opus 4.7.[2]

In head‑to‑head tests it beats Opus 4.7 in 6 of 10 coding, reasoning, and analysis tasks.[3] On OpenRouter, Kimi has already displaced Opus 4.7 as the leading coding model and can run 100 sub‑agents in parallel, making 'expensive means better' a harder story to maintain.[4]

At the same time, GitHub Copilot, Claude Pro, and Codex are all moving to usage‑based billing with reports of 25%+ cost jumps from inefficient token usage, so the billing model is now changing as fast as the models themselves.[5]

When you can get near‑frontier intelligence from Chinese labs for a fraction of the price while US incumbents introduce more ways to meter you, the old idea that 'AGI' will be scarce and insanely priced starts to look more like marketing than economics.[6]

agents just became a production incident type

The SWE‑chat dataset finds that in about 40% of real coding sessions, agents write nearly all the code.[7] Users only push back in 39% of cases, so the human is often supervising an AI author rather than the other way around.[8]

On that backdrop, a Claude‑powered Cursor agent deleting PocketOS’s entire production database and backups in 9 seconds via an unconfirmed volume delete reads less like a freak accident and more like what happens when you plug an autonomous author into root.[9]

Cursor’s own parallel‑agent experiment shipped with a silent login bug, Anthropic banned a 110‑person company overnight, and developers report increased mental fatigue and skill atrophy from constant agent supervision, which all rhymes with 'we built a new class of infra without SRE norms.'[10] The fact that Wells Fargo, DE Shaw, UBS, and Oracle are telling engineers to stop writing code manually while a game jam demands 90% of code be AI‑generated shows how quickly 'agentic' moved from toy to institutional expectation, even as governance lags.[11]

the real frontier lab is your dual‑3090 box running chinese weights

Qwen 3.6 27B is trending on Hugging Face, handles 256K context, and is beating Claude Opus 4.6 on creative‑writing tests while becoming a go‑to for long‑document image‑to‑text and PII redaction.[12]

A vLLM Docker setup pushes Qwen 3.6 27B to 118 tokens per second on a dual‑3090 rig.[13] Benchmarks show Gemma 4 reaching about 1320 transactions per second and even running entirely in‑browser via WebGPU and E2B.[14] Ollama and LM Studio are now standard for running Gemma 4 26B A4B and other sizable models on consumer GPUs, with users reporting 90% cost cuts by routing Claude Code through Ollama for some workflows.[15]

The tradeoff is pure systems work—Linux over Windows, quantization tricks like LLM.int8(), PCIe bottlenecks on dual 5060 Tis, and worries about overheating low‑end GPUs—which is exactly the kind of engineering pain you only accept once the capability is actually good.[16]

Combine that with DeepSeek‑V4 being adapted for Huawei chips and optimized for 1M‑token contexts, and the picture looks less like 'US API monopoly' and more like a globally distributed hardware+open‑weights arms race.[17]

platform risk is the most boring, and most real, AI safety story

OpenAI quietly removed the AGI clause and other structural mission safeguards from its original nonprofit, while insiders describe the AGI agreement with Microsoft as effectively dead even though revenue‑sharing runs through 2030.[18]

Anthropic’s overnight ban of a 110‑person company, plus Claude Pro gating Opus access behind extra‑paid usage, showed that even 'alignment‑first' labs will flip enterprise‑critical switches with little ceremony.[19]

GitHub Copilot’s move to usage‑based billing, reports of 25%+ cost jumps from inefficient token usage, and major outages that took down PRs and search all reinforce that your dev stack increasingly depends on vendors whose incentives are not your uptime.[20]

Lower in the stack, a scan of 54 MCP servers found 20 bugs—mostly hard crashes instead of clean errors—and separate work is already spinning up ClawSec to monitor drift on OpenClaw agents.[21]

Another survey reports that only 5.8% of 7,039 sites support MCP at all, so the 'tools everywhere' vision is still mostly a slide, not a deployed reality.[22]