FOMO Buster #007: The 7-day model

A newsletter that makes FOMO obsolete.

if i fear missing out on reading this newsletter, did i win or lose

Curated by Maurice.

01 · The One Thing

GPT-6.1 Sol: near-Astra intelligence for a fifth of the price

Seven days. OpenAI replaced GPT-6 Sol with GPT-6.1 Sol a week after the original shipped, claiming near-Astra intelligence at a fifth of the price — and the HN thread (1,006 points, 883 comments) is the biggest of the morning. Artificial Analysis frames it as a replacement release, not a side variant: the model it displaces lasted one week. The distribution tell arrived within hours: Cognition put GPT-6.1 Sol into Devin, reporting coding costs cut 57% overnight.

This is the velocity thesis made literal. When flagship models get replaced in seven days, the thing you’re buying isn’t a model — it’s a subscription to a changelog. The winners here aren’t the labs; they’re whoever prices like they expect the floor to move weekly. Devin just did.

Signal: The model-release cadence has collapsed from quarters to days. Pin a model version at your own risk: the deprecation clock now runs faster than most product cycles. Expect the cost floor to keep dropping as replacement cycles compress. Sources: AA: GPT-6.1 Sol replaces GPT-6 Sol after 7 days · HN thread · AlphaSignal: Cognition brings GPT-6.1 Sol to Devin

02 · Lab Watch

OpenAI’s Dots: a persistent agent that works while you sleep

DevDay’s headliner wasn’t a model — it was an always-on agent. Dots turns ChatGPT into a persistent agent that keeps working when you’re not there (693 points, 559 comments on HN). The industry has spent two years shipping agents that run while you watch; Dots is the bet that the real product is agents that run while you don’t. Whether the infrastructure (and the trust) is there yet is the whole question — and the thread is split on exactly that.

Decisions API: a constrained GPT-6 Luna router for developers

Quietly the most product-shaped announcement of DevDay: a constrained router that makes decisions without burning the big model. Read next to yesterday’s System-One through-line (Julia-1, Jeeves, Jev-Omni) and the pattern is unmistakable: every lab is now shipping the router layer as a first-class product. “Don’t make the expensive model do the cheap model’s job” has become an API.

DevDay’s developer stack: Sol, Ultrafast, and Cloud Codex

Beyond Dots and the router, OpenAI pushed its developer surface outward in three directions: Cloud Codex (the agent moves off your machine), Codex CLI rebuilt for parallel coding agents (details), and Ultrafast (a speed tier). Plus: Codex now scans GitHub repos using Daybreak Blue’s security AI (details) — the coding agent is also becoming the security reviewer. The terminal, the cloud, and the scanner are converging into one surface.

Signal: DevDay’s message is that the agent stack is now horizontal: persistent agents (Dots), routers (Decisions API), and cloud execution (Cloud Codex). The coding agent is no longer a tool — it’s the platform, and OpenAI just claimed every layer of it.

03 · Open Weights

The replacement economy hits Hugging Face

convaiinnovations/laya still leads trending, but the all-time board barely noticed it — attention without loyalty, week four. The real delta is the climbers: deepseek-ai/DeepSeek-V4.1-Flash is now trending with 3,919 likes (already sitting at #24 on the all-time liked board), alongside fresh entries XiaomiMiMo/MiMo-V2.6-Pro-RL, XingChen-AGI/Xing4.0-29B-A4B, TaichuAI/ZDTaichu5.0-9B, and inclusionAI/Ming-Image-0.1-Design — models climbing the trending list that are not yet established gravity wells. The flash-tier race is the theme: everyone’s shipping the cheap-fast variant and the community is rewarding it instantly.

Meanwhile the refusal-stripping economy keeps growing around the uncensored image models: the Qwen-Image-2.1 GGUF family holds over a million downloads, and DavidAU’s “heretic” uncensored Qwen3.8 variant has 1.78M downloads. New gravity forming in adjacent niches: nvidia/Nemotron-3-Diarization (speech diarization, 36k downloads) and fastino/GLiNER2.5-Decide (the router-pattern lands in NER, 34k downloads).

Signal: Open weights are mirroring the labs: fast, cheap, and frequently replaced. The trending/all-time delta is the early-warning list — DeepSeek-V4.1-Flash and the MiMo RL variants are this week’s names to know before they become fixtures. Sources: HF trending · deepseek-ai/DeepSeek-V4.1-Flash · nvidia/Nemotron-3-Diarization

04 · Money Flow

The capital story is the cost collapse

No mega-rounds today — the money moved inside the price war instead. DeepSeek V4 Flash cut reasoning costs 99% in 19 months (AlphaSignal); GPT-6.1 Sol delivers near-Astra intelligence at a fifth of the price; Cognition cut Devin’s coding costs 57% by switching models overnight. When the unit cost of intelligence drops two orders of magnitude in under two years, the capital question stops being “who raises” and starts being “who survives the margin compression.” The labs are funding the war by making their own product cheaper — which is only sustainable for the ones with the deepest balance sheets.

Signal: Watch the agent-builder business models built on today’s margins: the cost floor is moving faster than pricing pages can follow. Everyone whose product is a wrapper on a model’s price just got repriced by their supplier.

05 · Leaderboard

Claude Sonnet 5.5 hits #2 on the Intelligence Index

Yesterday’s mid-tier coup is now official on the scoreboard: Claude Sonnet 5.5 reaches #2 on the Artificial Analysis Intelligence Index (announcement) — the mid-tier model that beat Opus 5.5 on coding is now the second-smartest model on the board, period. Meanwhile OpenAI is already churning: GPT-6.1 Sol replaces GPT-6 Sol after seven days (AA), and the board is filling with effort-tiered variants faster than the index can rank them. New entry worth noting: JT-4.1 Flash 236B A21B (model page).

Signal: Leaderboards are becoming perishable goods — #2 one week, displaced the next. The ranking that matters is shifting from “smartest” to “smartest per dollar,” and nobody’s index has caught up yet.

06 · The Argument

Livenerf: Has Opus 5.5 been nerfed yet?

762 points, 322 comments, and a repo built to answer one question: did the flagship get quietly worse after launch? The “nerf watch” genre is now a standing HN institution — every major release gets a crowd-sourced regression tracker, because everyone has now watched a model they depend on change behavior under them. The irony is thick today: the same week OpenAI openly replaced a model after seven days, the community is building infrastructure to catch silent replacements. The Argument isn’t whether any specific nerf happened; it’s that model behavior is now the least stable dependency in the stack, and everyone knows it.

Signal: “Nerf watch” is becoming infrastructure, not discourse. The livenerf repo is the community’s answer to the 7-day model — and it suggests the real product gap is verifiable model-behavior changelogs. See today’s Build on This.

07 · Papers

Correct Answers, Invalid Traces

Verifiable grade-school math reveals something uncomfortable: models can get the right answer with a chain-of-thought trace that doesn’t actually support it. Correct answer, invalid reasoning — the trace is theater. Pairs directly with yesterday’s failure-transparency benchmark (agents reporting success after tool failures): the whole evidence pipeline of agentic AI, from trace to report, is getting stress-tested in public this week.

Signal: CoT faithfulness is having its reckoning. “Show your work” only works if the work is real — and verifiable tasks are now proving it often isn’t.

Do LLM Agents Execute the Plans They Declare?

From planning-mode declaration to pattern-specific execution: do agents actually follow the plans they announce? The question lands in the same 24 hours as Dots (the always-on agent) — right as agents get permission to run unattended, the research asks whether they’d follow their own plan if they did.

Signal: Plan-following is becoming a measurable property, not an assumption. For unattended agents, this is the paper that matters most this week.

Thinking Before Thinking: Scaling Agentic Inference Through Meta-Reasoning

Meta-reasoning as the scaling lever for agentic inference: think about the thinking before doing the thinking. Fits the week’s through-line — effort knobs, adaptive reasoning, routers deciding how much cognition a decision deserves. The stack is converging on the same insight from every direction: how much to think is now a first-class design decision.

Signal: Meta-reasoning is the academic twin of the industry’s effort-knob wars. Read it as the theory behind Sonnet 5.5’s adaptive reasoning and the Decisions API.

08 · GitHub Trending

The theme: the agent’s cloud and its money memory

The daily board still runs through the agent-ops wave: NVIDIA/OpenShell (11,644 stars, the safe private runtime), debpalash/VoiceStudio (49,900, the local voice stack), mvschwarz/openrig (2,774, multi-harness orchestration), and mksglu/context-mode (24,343) holding the daily list. The weekly board adds two tells: anthropics/financial-services (38,325) — Anthropic shipping a vertical agent stack for finance, the first serious industry-specific harness from a lab — and HKUDS/CLI-Anything (51,090). Plus anthropics/claude-code-action (9,255) — the lab productizing the “Claude reviews your PRs” pattern as a first-class GitHub Action.

Signal: The lab is now competing with its own ecosystem: financial-services agents and code actions straight from Anthropic. The open-source harness wave built the category; the labs are now absorbing it.

09 · Hacker News

GPT 6.1 Sol: Near-Astra intelligence for a fifth of the price

1,006 points, 883 comments — the morning’s biggest thread by far, and the comments are doing the real work: pricing math, replacement-cycle anxiety, and the question of what “near-Astra” even means when the goalposts move weekly.

Dots: Always-on agents

693 points, 559 comments. The thread is a referendum on unattended agents: the convenience case writes itself, and the failure-mode case writes itself louder. Read the top comments for the security framing — they pair with the Perplexity sandbox red-team story from yesterday.

Also notable

Livenerf: Has Opus 5.5 been nerfed yet? (762 points, 322 comments): the nerf-watch repo, covered in The Argument. Most data centers refusing to say how much water, electricity they use (339 points, 175 comments): the infrastructure-opaque story continues — see Safe to Ignore for why it’s not this newsletter’s fight. Pi.dev: You Said No MCP (339 points): the anti-MCP counter-essay in the tool-protocol wars.

10 · OpenRouter Apps

The coding trio ties; Framer teleports to #2

Hermes Agent (~2.03T) still leads, but the board’s two stories are below it. First, the coding-agent race is now a dead heat: Cline (~1.036T), Claude Code (~1.034T), Kilo Code (~1.018T) — all within two percent of each other, the closest the trio has ever been. Second, Framer (~1.39T) teleported to #2 — up from ~49B yesterday, a 28x jump that smells like a new integration or a board-view change rather than organic growth, so read the exact order loosely. Behind the top: pi (~467B), omp (~444B), Command Code (~351B), CodeGPT (~334B), Codex (~314B), DeepSeek Harness (~226B), Freebuff (~218B), OpenClaw (~201B), Instaplay (~124B), Strix (~118B).

Signal: Three coding harnesses within 2% of each other means the coding-agent market has no leader — it has a commodity. The differentiation has moved up the stack: who routes, who budgets, who remembers. Framer’s jump is the anomaly to verify before citing.

11 · Watch

OpenAI Just Revealed Dots… This Changes ChatGPT Forever — TheAIGRID

TheAIGRID · Sep 30 — same-day breakdown of the Dots announcement. Watch after reading today’s One Thing: the persistent-agent pitch, decompressed.

Signal: The fastest video explainer of DevDay’s biggest swing. Ten minutes to calibrate on what “works while you sleep” actually ships as.

ASTRA 6.1 too dangerous to be released… — WesRoth

WesRoth · Sep 29 — the alignment-doom counterweight to a week of capability news. Take the title with salt, but the underlying question (what happens to release norms when models ship weekly) is real.

12 · Listen

Jensen Huang Thinks A.I. Alarmism Has Gone Too Far — Hard Fork

Hard Fork · Sep 25 — Ezra Klein’s show hands the mic to Jensen Huang on AI alarmism, the same week Wes Roth is warning ASTRA 6.1 is too dangerous to release. The two poles of the discourse, one episode apart. Essential pairing: hear the industry’s strongest “calm down” case right before deciding how much of the doom content to take seriously.

Signal: The alarmism debate is the discourse layer of the capability week. Huang’s case is the best-articulated version of the “build” side — and it’s worth hearing even if you land on the other one.

OpenRouter: from Seed to Stripe — Latent Space

Latent Space · Sep 25 — with OpenRouter’s Alex Atallah on the road from seed to Stripe. Pairs with today’s OpenRouter section: the business behind the token leaderboard, in the founders’ own telling.

13 · Read

NVIDIA’s Physis-Lang teaches video AI real physics, beats Google’s Veo 3.1

Days after AMD bought World Labs for $8.2B, Nvidia answers on the model side: Physis-Lang beats Google’s Veo 3.1 by teaching video AI actual physics. The world-model war is now two-sided — AMD bought the lab, Nvidia is shipping the model. Physical AI’s arms race is moving faster than the text-model one did.

Signal: The AMD/World Labs deal and Physis-Lang are the same story from opposite ends: whoever owns physical understanding owns the next platform. Robotics investors should be reading both.

Microsoft’s Quine ranked pancreatic cancer compounds in one weekend

An AI agent ranked pancreatic cancer compounds in a single weekend. The drug-discovery agent story keeps compressing timelines from years to days — and this one names a specific, brutal disease. The science-agent wave is producing results that read like press releases because they are, in the best sense, results.

Signal: Weekend-scale drug ranking is the new normal for science agents. The bottleneck is no longer compute or models — it’s wet-lab validation bandwidth.

RLinf’s RPent wraps frozen Pi0.5 robots with an LLM recovery brain

Don’t retrain the robot — give it a recovery brain. RPent wraps frozen Pi0.5 policies with an LLM that handles failure recovery, which is the cheapest possible way to make robots robust: keep the cheap policy, add the smart supervisor only when things go wrong. The router pattern, now for robotics.

Signal: The “small model does the work, big model handles the exceptions” architecture has reached physical robots. Expect this pattern everywhere by next quarter.

14 · Safe to Ignore

Permission to skip

Skip: America.gov. 686 points, 572 comments — a government website launch. Big thread, not this newsletter. Skip: Ask HN: What are you reading? 366 points, 703 comments — the quarterly book thread. Lovely. Not AI news. Skip: The data-center water thread. 339 points on datacenters refusing to disclose water and power use — infrastructure politics, not model news. One “refusing to say” headline per month is plenty. Skip: How Delhi cut electricity loss from 50 to 5 percent. 549 points of genuinely interesting infrastructure — and genuinely not this newsletter. Skip: September 2026: the world today, as seen by one Polish guy. A personal essay with 296 points. Good writing, wrong inbox.

15 · Build on This

PinWatch: the changelog your models won’t give you

OpenAI replaced a flagship model after seven days. HN built a repo to catch silent nerfs. The indie dev’s problem: your app pins gpt-6-sol, the lab swaps what that means, and your evals degrade with no alert. Product: a model-behavior watchdog — pin your production model, and PinWatch runs your eval suite (or a standard probe pack) against it on a schedule, diffing behavior, latency, refusal rates, and cost per task. When the model changes under you — announced or not — you get the diff before your users get the bug. Sell it as a $29 one-time purchase per project, $99 for teams with shared dashboards — no subscription, because the whole point is that the ground moves; the tool shouldn’t. The wedge: every lab is now shipping weekly, and nobody’s changelog covers behavior. Verification is the product.

16 · Appendix

Everything below didn’t make the cut. Got a better pick? The next issue gets sharper.