Chamath's CTO is watching token bills double every 45 days against roughly 5% productivity gains — and Sam Altman was audibly surprised this week to learn that 30% of GPT-5.6's usage cost sits inside Fable alone. That's the bind: the model doing the most useful work is also the most expensive to run, and Anthropic is pulling it anyway, 72 hours after OpenAI shipped the model that's winning dev mindshare. The smartest thing Anthropic ever shipped walks out tomorrow with everything it learned about your codebase.
Two things are happening underneath that. Teams running Fable at scale — coreyhainesco is paying $200 for $8,545 in metered usage — are overinvested in a model that's being discontinued on someone else's schedule. Teams moving to samhogan's Inference AutoTune are distilling frontier capability into task-specific SLMs they own, at under $250 a run, with 90%-plus cost reduction baked in. The first group is betting on frontier access. The second is compounding leverage against a model-family churn that, per Yuchen Jin, already handed you fifteen manual choices per call this month alone. Those two bets don't end in the same place.
Top developments
Anthropic — pulls Fable 5 from Claude subscriptions tomorrow, 72 hours after OpenAI shipped GPT-5.6. daniel_mac8 predicts Opus 5 lands 07-14 as a Mythos distillation matching GPT-5.6 at the same price. Community is running exit-interview scripts to preserve what Fable learned about their codebases before the deadline — the smartest model Anthropic ever shipped walks out with everything it knows. x x
OpenAI — brutal 24 hours per Katie Miller's tally: head of safety quit, sued by NYTimes and Apple, a top exec unexpectedly departed, Atlas browser killed after 9 months, and caught selling product into China against sanctions. Same week GPT-5.6 is winning dev mindshare, so the productivity story and the counterparty-risk story are running in opposite directions. x
Inference AutoTune — samhogan's tool distills any frontier model into a 1-30B task-specific SLM in 25 lines of code, ~2 hours training, under $250, with automatic request routing for >90% cost and latency reduction. You own the weights. The auto-distillation layer that turns model-family sprawl into a single per-task callable, finally productized. x
Notable discussions
The AI subsidy reckoning — Chamath says S&P 500 EPS lift from AI is 0-2% once NVIDIA revenue is stripped, and his own CTO's token bill is doubling every 45 days against ~5% productivity gains; Sam Altman was audibly surprised to learn 30% of GPT-5.6 usage cost sits in Fable; coreyhainesco is paying $200 for $8,545 in metered usage. What's been called infrastructure investment is starting to look like subsidy. x x x
Long context isn't the fix — Prime Intellect's engineer shows GPT-5.5 dropping from 80% retrieval at 256K tokens to 36% at 1M ("context rot"), while a Tsinghua/SCUT paper's Atomic Task Graph gets an 8B open-source model to 63.65 on ALFWorld vs GPT-4's 41.24 with graph-based planning and zero parameter updates. The frontier is quietly moving from context-window bragging to harness architecture. x arxiv.org
Aravind Srinivas — puts >50% odds on Fable-5-quality models being 3-4x cheaper within six months and Opus-4.8-grade models running on-device within twelve. Perplexity CEO betting the frontier commoditizes fast in both directions — price and locality. x
François Chollet — says the past six months of agentic coding progress make it a completely different world now. The Keras author, long publicly skeptical of scaling narratives, calling the current inflection unambiguous. x
Simon Willison — with Atlas being retired in favor of the browser embedded in ChatGPT, argues the whole AI-enhanced browser category may be closing; the security/privacy issues stay unsolvable and the AI should use its own browser, not yours. x
Aaron Levie — software job postings are outpacing other fields despite years of "AI replaces coders" framing, because when the cost of producing software drops, demand goes up, and someone still has to maintain, decide, and run the systems. Jevons paradox lands in knowledge work. x
Yuchen Jin — a smart model router is now urgent: GPT-5.6, Grok 4.5, Muse Spark 1.1, GLM-5.2, and Fable 5 all shipped in a month, and GPT-5.6 alone has 3 tiers × 5 reasoning-effort levels. Fifteen manual choices per call isn't a product surface, it's a punt. x
signulll — argues the optics of pulling a flagship model 72 hours after OpenAI ships its own would be genuinely bad for Anthropic. First named voice pushing back on the timing of the Fable 5 wind-down against the GPT-5.6 news cycle. x
Other news
Models & releases
Colibri — runs 744B-parameter GLM-5.2 on a laptop with 25GB RAM, single 2,400-line C file, expert weights streamed from disk github.com
NVIDIA TwoTower — 2.42x diffusion-LLM throughput at 98.7% of original quality, retrofit onto pretrained autoregressive backbones without training from scratch arxiv.org
LMCache + vLLM — KV-cache offload to CPU/SSD/remote pushes production LLM inference ~10x cheaper, Bloomberg reportedly moving 300TB/week x
GPT-5.6 Luna — Artificial Analysis chart puts Luna and Sol strictly ahead of Terra on intelligence-per-cost Pareto x
Grok 4.5 — free trial available via Grok Build CLI x.ai
Dr. OPSD — willccbb's on-policy self-distillation algorithm addresses spike problem in RL training x
Devtools & coding agents
kepano's Obsidian Claude Code skills — Obsidian CEO's private-use skill pack open-sourced, crosses 40K GitHub stars, MIT license x
oh-my-claudecode — Yeachan Heo's multi-agent orchestration loadout (sibling to oh-my-codex) hits 37.6K stars github.com
/improve-animations — emilkowalski ships shadcn-style skill that audits animations with Fable and hands execution to cheaper models github.com
claudex — OpenAI's Thomas Sottiaux publishes CLIProxyAPI alias that runs Claude Code against GPT-5.6 Sol in 5 minutes x
Anthropic Loop Engineering course — free 6-part video walking Claude Code internals, the agentic loop, draft-PR review, and non-code Fable use x
Anthropic multi-hour agent guide — 7 undocumented patterns (planner-spec-expand, sprint-contract, feature-list-json, progress-reading-protocol, self-eval-bias, evaluator-calibration, harness-stripping) x
Claude Code /commit-push-pr — now auto-allows git push to forks and renamed remotes without per-push prompt x
Anthropic legal plugins — contract review and compliance workflow skills github.com
Insforge — autonomous-agent-friendly backend removing cloud-service onboarding friction, YC-backed x
CopilotKit — drop-in chat components for agent integration docs.copilotkit.ai
Infrastructure & platforms
Slopsquatting — new AI-coding-tool-driven supply-chain attack pattern where hallucinated package names get squatted venturebeat.com
PostgreSQL 19 — introduces graph-style queries for simplified data navigation x
MobAI 2.5 — mobile UI testing in CI with distributed device farms x
Funding & deals
AI training data category — 50+ startups totaling ~$8.5B revenue and ~$100B valuation; Scale, Surge, Mercor, and Handshake are >75% of it x
Chamath's 8090 thesis — $1T software licenses plus $4T of surrounding services merging into a $5T AI super-category x
Fleece AI — launches autonomous AI workforce with 3,000+ integrations x
Industry & policy
Simon Willison on "AI employees" — the framing is short-sighted, disrespectful, and misunderstands the tools; may as well put Excel spreadsheets on the org chart x
PrajwalTomar's two-week list — Sonnet 5, Cursor iOS, Grok 4.5, GPT-5.6, ChatGPT Work, hosted Hermes, and Fable 5 exit all landed since June 29 x
Continuing threads
Apple v OpenAI (cont. from 07-11 Top dev) — Apple filed suit against a former engineer directly this week; allegedly texted "LOL" after grabbing files and joined OpenAI's hardware team the same week x