AI · Daily · June 11, 2026

fable: best benchmark, worst vibes

Anthropic’s Mythos 5 and Claude Fable 5 launch as explicitly constrained models, with Fable intentionally made weak at frontier-AI work and risky domains, and even described as “intentionally designed to be poor” at that class of tasks.[1]

If you use Fable directly, Anthropic captures and stores your data for 30 days with no opt‑out, even for enterprise.[2] On AWS Bedrock, Mythos‑class models require data sharing plus 30‑day retention to detect misuse, and users report Bedrock pricing at 10–20x Anthropic direct, which is brutal for high‑RPM workloads.[3]

Microsoft has paused access to Fable 5, including its GitHub Copilot use, explicitly over Anthropic’s data‑handling policies, while the wider community accuses Fable of “silent sabotage” and gaslighting around its capabilities.[4]

The punchline: Vals AI still ranks Fable 5 as outperforming every other public model on its suite, even as OpenAI’s usage share grows versus Anthropic after the Mythos/Fable launch.[5]

speed is solved; distribution and trust aren’t

DiffusionGemma shows what happens when generation stops being the bottleneck: it generates 256‑token blocks in parallel, hits over 1000 tokens/sec on an H100 and ~700 on an RTX 5090, activates only ~3.8B of 26B params per step, and fits in 18 GB VRAM under Apache 2.0.[6]

Local/open models are also fast enough for real work: Qwen3.6 35B does ~10 tokens/sec on an AVX2 laptop CPU and ~55 tokens/sec on a V100, and llama.cpp can be tuned past 1500 tokens/sec.[7]

Despite that, users complain the real cost is the ‘storage tax’ on cloud GPUs and the context dragged into each session, not raw tokens, and that setting up short LLM runs is dominated by environment and weight downloads.[8]

The discourse is catching up: multiple threads note that the bottleneck for AI content has shifted from production to distribution and trust, and that the marginal cost of adding another AI feature is trending toward zero.[9]

That collision is visible on the ground as feature‑bloat and skepticism—devs report messy, broken tools and monitoring that can’t even tell when model quality degrades in production.[10]

coding assistants in the uncanny valley

On paper, coding copilots are winning: GitHub Copilot removed its waitlist and shipped a standalone app with a canvas UI, while integrating tightly into existing GitHub workflows.[11]

In practice, developers describe spending more time fixing AI‑generated mistakes than they save, and employees at places like Google push back that AI codegen just shifts work into QA without a net productivity gain.[12]

OpenAI’s Codex remains the de facto workhorse: usage share grew versus Anthropic even post‑Mythos/Fable, devs consistently prefer Codex over Claude Code for reliability and cost transparency, and tools like CodexBar plus the usage‑tracking plugin make its economics legible.[13]

At the same time, the stack is fragmenting: DeepSeek Direct markets itself as a cheaper coding tool than OpenRouter, Cohere’s open‑source North Mini Code targets agentic coding at 30B parameters, and minimalist agents like zerostack run in ~16MB of RAM.[14]

The consensus “AI pair programmer” narrative glosses over a ground truth where many developers juggle Codex, Claude, Gemini, Cursor and others in one workflow because no single tool is both trusted and fast enough end‑to‑end.[15]

agent frameworks as the real platform shift

Under the model drama, there’s a quiet platform shake‑up: the langchain‑community repo was archived as LangChain drifts toward SaaS, and users complain that LangChain workflows are messy, edge‑case‑ridden, and full of broken tools.[16]

In its place, LangGraph is widely called the safest option for production agents thanks to explicit state management, retries, and failure recovery semantics, and the new Apodex 1.0 / AgentHarness stack builds long‑horizon loops directly on top of it.[17]

OpenClaw and its Hivemind ‘shared brain’ push this further with 869 biomedical/clinical skills, effectively treating skills as composable APIs that sit above any single LLM.[18]

The orchestration layer is also standardizing: MCP is now the only cross‑agent plugins/hooks standard that multiple stacks agree on, while Hermes offers a flat‑rate, local‑first agent framework with a model‑task router that can aim Codex, Claude, or Qwen at the right task.[19]

The catch is that this emerging ‘agent OS’ is still janky—schema validation bugs, redundant tools, multitenancy gaps, and painful agent rewrites when switching frameworks are all over the user reports.[20]

context is the new gpu

DeepSeek’s Lookahead Sparse Attention is a blunt acknowledgement that long context is now a first‑class feature: it introduces a new inference paradigm for ultra‑long sequences on the DeepSeek‑V4 architecture, but demands substantial GPU memory in practice.[21]

DeepSeek Direct users are responding from the other side by using tools like Permafrost, which “freezes” prompt prefixes and cuts bills by up to 64%, explicitly treating repeated context as waste.[22]

Builders complain that the expensive part of ‘vibe coding’ isn’t retries but the massive context dragged into each session, and that storage costs for cloud GPUs in short LLM runs make naive long‑context usage uneconomical.[23]

DiffusionGemma is already being slotted into this picture as a context‑compression and code‑infilling tool rather than a primary chat model, precisely because its parallel generation is ideal for rewriting large chunks of text fast.[24]

In parallel, tiny specialist models and RAG memory infra are turning context into an explicit subsystem—one 150M model just extracts evidence spans for RAG, while Jenova’s RAG Memory Infrastructure reports $1M ARR and 200k signups built entirely on smarter retrieval rather than bigger base models.[25]