AI · Weekly · May 27, 2026

the tokenmaxxing crash

Quarterly token volume is up ~17,000x in four years while prices collapsed, creating a culture of tokenmaxxing as a proxy for progress.[3]

Now the bill has arrived: Microsoft is canceling internal Claude Code licenses, explicitly citing unsustainable token-based costs and even halting AGI projects.[4]

Uber’s COO reports no measurable productivity gains from AI despite burning through the budget early, and says AI tools are pricier than engineers.[5]

Salesforce still plans to spend around $300M on Anthropic tokens this year with AI doing only 30–50% of its workload, while some startups blow $1.3M a month on tokens.[6]

Median coding-agent requests already stuff in 96k tokens—longer than The Great Gatsby—so the default workflow is literally to overcontext everything.[7]

deepseek and the commoditization of tokens

Into that mess walks DeepSeek V4 Pro, permanently cutting prices by 75% to about $0.435/M input and $0.87/M output tokens—roughly 11.5× cheaper than GPT‑5.5.[8]

DeepSeek is simultaneously raising $10.29B specifically to scale open-source-style models rather than a closed SaaS platform, a very different capital story from OpenAI.[9]

China is now restricting overseas travel for DeepSeek and other AI talent, effectively treating its model weights and people as strategic assets.[10]

OpenRouter and similar brokers are already routing huge volumes—25T tokens/week—through cheaper models like DeepSeek after its cuts, making per-token price a first-class competitive lever.[11]

The background hum in forums is that AI vendors aren’t trustworthy and token pricing is opaque, which makes a permanently cheap, open-leaning player look less like a discount and more like a wedge.[12]

meta-optimizers are the real 'new models'

A single PyTorch-based 'universal optimizer' almost tripled Gemini Flash’s ARC‑AGI score from 32.5% to 89.5% while cutting cloud costs by ~40% across six tasks, without a new base model.[13]

On the code side, GPT‑5.5 hits 70% on the new DeepSWE benchmark, which involves editing ~668 lines across seven files per task—real engineering, not toy LeetCode.[14]

SWEBench Pro now looks artificially harsh for GPT‑5.5 because 68.5% of its 'failures' came from broken tests, implying an effective score closer to 86.7%.[15]

Meanwhile an OpenAI general-purpose model just disproved a long-standing Erdős conjecture in discrete geometry and discovered a new, better family of constructions, a qualitative shift from 'autocomplete for math proofs.'[16] The pattern is that orchestration, evaluation, and targeted fine-tuning are driving the biggest step-changes, while the marketing still talks like it’s all about bigger base models.[17]

local-first is no longer cosplay

On the hardware fringe, AMD + Vulkan is quietly becoming a serious LLM platform: users report ~20% speedups over ROCm and RX 7900 cards outpacing Nvidia 3090s for local inference.[18]

A dual-RTX 3060 setup is decoding Qwen 3.6‑27B at 30–50 tokens/s, while a single 3090 can push Qwen 3.6 27B to ~164 tokens/s with the right configuration.[19] llama.cpp keeps squeezing more from this hardware, with Multi-Token Prediction updates yielding up to 7× faster generation on Qwen 27B-class models and BeeLlama hitting ~178 tokens/s on a 3090.[20]

In parallel, WebGPU is turning browsers into runtime environments: PrismML’s 3GB Bonsai Image 4B text-to-image models run fully client-side with 1‑bit/ternary weights, and Local Ghost serves Qwen2.5 offline in-browser.[21]

All this is happening against a backdrop where GPUs are still painfully expensive and many users on lower-end hardware treat lightweight quantized models as the only way open source is actually usable.[22]

agents are now a security problem with a UI

The most interesting agent work right now reads like security engineering, not UX: Runtime’s YC-funded sandboxed coding agents, Gemini Managed Agents’ Linux sandbox, and Edge.js’s Node-in-WASM all treat code execution as the primitive.[23]

Protocol-wise, MCP is standardizing how models talk to tools and data across 10,000+ servers, now moving to a stateless design while already showing that ~15.3% of scanned public servers are vulnerable enough for NSA warnings.[24]

RAG pipelines still fail mainly on retrieval—about 60% of breakdowns—yet those same pipelines are being wired directly into agent toolchains.[25]

The surrounding software supply chain is porous: a poisoned VS Code extension exfiltrated ~3,800 private GitHub repos, the Megalodon campaign hit 5.5k more, and a Hugging Face dataset stayed poisoned for six months.[26]

In other words, 'autonomous agent' increasingly means 'scriptable front-end on your production environment,' with an attack surface expanding faster than most security teams are staffed.[27]