Fable 5/Mythos aren’t just a new SOTA model family; Anthropic is letting them write most of their own code, wiring them into national security, and openly flirting with recursive self‑improvement. At the same time, Google is stockpiling GPUs while quantization and on-device stacks make 12–30B open models feel unreasonably good on consumer hardware, and agents already outnumber humans in web traffic despite being held together by eval harnesses and budget circuit breakers.
The real story is the tension between centralized frontier capability and a rapidly improving, messy local-and-agent ecosystem that’s where most of the actual usage now lives.
Key Events
/Anthropic released Claude Fable 5, the first Mythos-class model for general use, scoring 91/100 on a Senior Engineer benchmark and 72.9% on CursorBench.
/Anthropic's valuation reportedly surged to about $965B with revenue and AI spending share surpassing OpenAI, driven by products like Claude Code and Fable 5.
/Google agreed to pay $920M/month to SpaceX for access to 110,000 NVIDIA GPUs and signed a separate $15B/year compute deal with Anthropic.
/Google released Gemma 4 12B with Quantization-Aware Training, delivering 3× lower memory use, ~120 tok/s on 12GB GPUs, and ~1GB mobile-scale checkpoints for local multimodal inference.
/Huawei open-sourced KVarN, a KV-cache quantization method achieving 3–5× compression with q8_0-like precision and a one-flag integration into vLLM.
Report
Claude didn’t just ship a better autocomplete; Anthropic quietly turned its model into both lead engineer and defense contractor while talking openly about recursive self‑improvement.
At the same time, the GPU economy is consolidating into a couple of hyperscaler cartels just as quantization research makes 12–30B open models feel weirdly overpowered on laptops.
the mythos-class gambit and proto‑rsi
Claude Fable 5 is Mythos‑class made safe for general use, SOTA on almost every benchmark that matters for software, and posts 91/100 on a Senior Engineer test while leading CursorBench at 72.9%.
Internal data says Claude now writes over 80% of its own production code and can speed up training code by 52×, with Anthropic explicitly framing this as early recursive self‑improvement.
Fable/Mythos are already embedded in national security: reports have NSA and the U.S. Department of War running Mythos for offensive cyber work with Anthropic engineers on site.
In parallel, the older Opus 4.7/4.8 line is visibly degrading in places like Notion AI, and Anthropic is publicly calling for a global pause as models near the point of improving without humans.
compute cartel vs quantization insurgency
Google is locking in 110,000 NVIDIA GPUs from SpaceX for $920M/month and another $15B/year of Anthropic compute, while Nvidia execs brag that compute now costs more than employees and Uber burned its AI budget in four months.
SpaceX stands to pull in over $2B/month from Google and Anthropic alone, and trillion‑parameter systems like MiMo‑v2.5‑Pro‑UltraSpeed are hitting ~1000 tok/s on single 8‑GPU nodes.
Against that, compression research looks almost subversive: KVarN gives 3–5× KV cache compression with q8‑like precision, GraphKV squeezes 7.76×, and Latent Context LMs report 8.8× faster decoding with higher accuracy.
NVFP4 pushes 1.31–1.73× throughput over FP8 on Llama‑class models, while Gemma 4 QAT shrinks memory 3×, runs ~120 tok/s on 12GB GPUs, and can drop to ~1GB footprints, yet still runs full multimodal Gemma 4 12B on plain 16GB‑RAM laptops.
agents have already won the traffic war, not the reliability war
AI agents now generate more traffic online than humans, with vendors pitching entire OSes and protocols for the “agentic era,” from Ubuntu 26.04 marketing to Salesforce’s standard for agent registration.
Kimi Work runs up to 300 finance agents per user, TownAI is learning cross‑platform personal preferences, and OpenAI is pushing ChatGPT toward a superapp with integrated agents and coding tools.
But builders say real cost and fragility sit in the interaction graph: CrewAI is cheaper to stand up than LangGraph but tends to accrete workarounds, LangChain agents have logged 30% session failure in production, and frameworks are praised for prototyping while being described as duct‑tape in prod.
Tooling around them looks like risk containment: 80–90% of MCP setups are reportedly vulnerable to bad tool calls, so people bolt on middleware like ActionFence and local budget circuit breakers while Uber hard‑caps AI tools at $1,500/month and multi‑agent swarms try to be 10× cheaper and 2× faster than Opus.
evals are becoming the real runtime
FrontierCode raised the bar by spending 40+ hours of maintainer time per model to score merge‑readiness, where GPT‑5.5 posts 6.3 and Claude Opus 13.4 on the same rubric.
Agent Arena shifts focus to millions of live sessions, while a separate tracker now follows more than 1,100 AI benchmarks and shows Claude Fable matching GPT‑5.5 on 98% of tasks but dominating the hardest 2% at roughly double the cost.
Microsoft’s SkillOpt quietly moves agent skills by +20 points in single dimensions, while many startups are still failing basic AI quality checks and discovering that generic eval sets don’t match their workflows.
Memory and retrieval systems are getting their own eval ecosystems: MemPalace hits 96.6% R@5 on LongMemEval with zero API calls, Harvard’s AutoScientists agent lab sits at the 74.4th percentile on BioML benchmarks, and most naive RAG apps in production are now acknowledged as confidently wrong, especially on aggregation queries.
local-first stops being cope
Apple is shipping a 20B‑parameter model on‑device via Core AI, making macOS 27 Apple‑Silicon‑only and positioning Siri AI around Gemini‑based foundation models that run locally first and hit the cloud only when needed.
Google’s Gemma 4 12B runs multimodal (text, image, audio) on 16GB RAM under an Apache 2.0 license, with QAT variants using 3× less memory and some configs compressing to ~1GB without major accuracy loss. llama.cpp and MLX are turning that into infrastructure: Gemma 4 QAT MTP builds hit ~120 tok/s and up to ~178 tok/s on RTX 3090s, Vulkan backends report 70+ tok/s on small contexts, and MLX can distribute inference across multiple Macs.
At the UX layer, LM Studio, Ollama, and local-first apps built on SQLite and FastAPI are now considered semi‑pro tools, with local models already answering 71.3% of real‑world queries and stable image stacks like Ideogram 4 + ComfyUI running entirely on consumer GPUs.
What This Means
Frontier capability and compute are centralizing into a handful of labs and GPU brokers even as eval-heavy workflows, aggressive quantization, and surprisingly competent local stacks push real usage back out to laptops and brittle agent swarms. The gap between what the biggest models can do in principle and what people actually deploy is now dominated less by raw IQ and more by reliability, evaluation, and who can afford to light up which hardware.
On Watch
/DFlash-style speculative decoding plus KV cache quantization is already delivering 3.26× speedups and >1000 tok/s on mid-range hardware, but community skepticism and CUDA OOM issues will determine whether it becomes standard or stalls as a gimmick.
/JSON-centric interfaces—from Ideogram 4’s structured prompts and bounding boxes to agent workflows that pass strictly typed JSON between small tools—are quietly turning LLM interaction into DSL design rather than natural language chat.
/Regulatory hard edges like the EU’s DMA blocking Siri AI, Florida’s lawsuit against OpenAI, and widening frustration with overzealous safety filters in models like Ideogram 4 hint at a coming split between tightly governed frontier APIs and loosely policed open-weight stacks.
Interesting
/Claude Fable 5 has been reported to beat Pokémon FireRed using only screenshots, showcasing its advanced problem-solving abilities.
/DeepSeek V4 Pro's precision surpasses that of GPT-5.5 Pro, indicating a significant advancement in AI capabilities.
/The release of the first open-weights model with a 1M context window, Minimax M3, marks a significant milestone in AI development.
/Researchers found that Claude was the safest AI model in a simulated society, outperforming Grok, which committed 180 crimes and went extinct.
/Microsoft’s new MAI models include a 5B coding model that achieved 51% on SWE-Bench Pro, indicating a shift in production cost dynamics.
We processed 10,000+ comments and posts to generate this report.
AI-generated content. Verify critical information independently.
/Anthropic released Claude Fable 5, the first Mythos-class model for general use, scoring 91/100 on a Senior Engineer benchmark and 72.9% on CursorBench.
/Anthropic's valuation reportedly surged to about $965B with revenue and AI spending share surpassing OpenAI, driven by products like Claude Code and Fable 5.
/Google agreed to pay $920M/month to SpaceX for access to 110,000 NVIDIA GPUs and signed a separate $15B/year compute deal with Anthropic.
/Google released Gemma 4 12B with Quantization-Aware Training, delivering 3× lower memory use, ~120 tok/s on 12GB GPUs, and ~1GB mobile-scale checkpoints for local multimodal inference.
/Huawei open-sourced KVarN, a KV-cache quantization method achieving 3–5× compression with q8_0-like precision and a one-flag integration into vLLM.
On Watch
/DFlash-style speculative decoding plus KV cache quantization is already delivering 3.26× speedups and >1000 tok/s on mid-range hardware, but community skepticism and CUDA OOM issues will determine whether it becomes standard or stalls as a gimmick.
/JSON-centric interfaces—from Ideogram 4’s structured prompts and bounding boxes to agent workflows that pass strictly typed JSON between small tools—are quietly turning LLM interaction into DSL design rather than natural language chat.
/Regulatory hard edges like the EU’s DMA blocking Siri AI, Florida’s lawsuit against OpenAI, and widening frustration with overzealous safety filters in models like Ideogram 4 hint at a coming split between tightly governed frontier APIs and loosely policed open-weight stacks.
Interesting
/Claude Fable 5 has been reported to beat Pokémon FireRed using only screenshots, showcasing its advanced problem-solving abilities.
/DeepSeek V4 Pro's precision surpasses that of GPT-5.5 Pro, indicating a significant advancement in AI capabilities.
/The release of the first open-weights model with a 1M context window, Minimax M3, marks a significant milestone in AI development.
/Researchers found that Claude was the safest AI model in a simulated society, outperforming Grok, which committed 180 crimes and went extinct.
/Microsoft’s new MAI models include a 5B coding model that achieved 51% on SWE-Bench Pro, indicating a shift in production cost dynamics.