FOMO Buster #001: The 90-minute war

A newsletter that makes FOMO obsolete.

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

01 · The One Thing

The 90-minute war

Tuesday morning, Anthropic launched Claude Opus 5.5: Fable-5.1-level performance, 30% faster output, framed as beating GPT-5.6 Sol on CursorBench at one-third the cost. Ninety minutes later, OpenAI dropped GPT-6 Sol and Luna and cut API prices 50% on the spot.

This is the first model cycle where the launch-day discourse is about unit economics first and capability second. The timeline is wall-to-wall price-per-task memes. Polymarket is literally pricing “up to 93% cheaper coding costs.”

Signal: If you build on APIs, your cost basis got rewritten twice before lunch on Tuesday. Every agent product’s margins just moved. Recalculate build-vs-buy before you ship anything priced on old token math. Sources: techcrunch on OpenAI · macrumors on Anthropic

02 · Lab Watch

GPT-6 Sol + Luna, prices halved

Sol for complex coding and agentic work, Luna for high-volume tasks. Sol now $2/$10 per 1M input/output tokens (was $4/$20), Luna $0.10/$0.50, cached reads 90% off. Permanent cuts, not promo.

Signal: OpenAI is passing inference-efficiency gains straight to developers. The Jevons playbook, live: cheaper thinking, more thinking.

Opus 5.5, and a November IPO looms

List prices $4/$20 per 1M tokens (20% below Opus 5), but cache reads dropped 60% to $0.20/1M. Anthropic’s pitch: cache dominates agentic workloads, so typical costs fall ~40%. Sonnet 5.5 and Haiku 5.5 due in weeks. First release since Amodei’s “pace the frontier” call.

Signal: Both labs cut prices the same day via different mechanisms. The real benchmark is now cost per completed agent task, not cost per token.

Grok 4.7 ships, Copilot day one

New largest base model, longer RL on multi-hour tasks, new safeguard stack, $2/$6 per 1M tokens. Rolling out to GitHub Copilot immediately. Independent Artificial Analysis testing matched xAI’s claims except Terminal-Bench 4.0: claimed 38.0%, measured 25.76%.

Signal: Vendor benchmarks remain guilty until proven innocent. Also: Copilot distribution gives xAI instant developer reach, which matters more than the benchmark gap.

Hardware blitz at Connect, plus a curtain pulled back

Zuckerberg unveiled Muse Charm (keychain-sized pocket device for the Muse agent, December), Ray-Ban Meta Gen 3 at $449, camera-free audio glasses at $349, an FDA-cleared hearing-aid feature inside glasses, and Muse computer-use for Mac. Meanwhile Reuters found human contractors secretly making Muse’s phone calls on users’ behalf, one of whom made a racist reference on a call. Meta apologized and rolled the feature back. Separately: Muse is already straining Meta’s compute at ~700K DAU, and Amazon blocked Muse from shopping on its platform.

Signal: The “fully autonomous agent” still has humans behind the curtain, at Meta scale. And agents eating 10x projected tokens is a brutal unit-economics problem nobody has solved. Sources: techcrunch on Connect · gizmodo on contractors

Gemini 3.8 Flash TTS takes voice benchmarks

New text-to-speech models: 100+ languages, 2,000 production voices, 30-second voice cloning, SynthID watermarking. Took #1 on Hume AI’s Voice Design Benchmark and Voice Arena preference evals.

Signal: Voice quality is now a leaderboard battleground, not a solved commodity. If your product has a voice, the frontier moved this week.

03 · Open Weights

NVIDIA Nemotron 3 Diarization

100M-parameter speaker diarization (who-spoke-when, up to 8 speakers), streaming and offline, reportedly #1 on VoiceArena’s diarization board.

Signal: Tiny, practical, open. This quietly becomes infrastructure in every meeting-recording pipeline. See “Build on This” below.

Black Forest Labs FLUX 3 Action

7B robotics world-action model predicting the next 32 robot actions from camera frames plus text. Claims 42.92% on NVIDIA’s RoboLab-120, beating NVIDIA’s own 16B Cosmos policy. FP8 variant runs on 24GB consumer GPUs. Commercial license terms unclear in coverage found.

Signal: Competitive robot policies on consumer GPUs could do for robotics what Stable Diffusion did for images. Verify the commercial license before building on it.

Alibaba Qwen-Image-2.1

7B open-weights text-to-image plus image editing, native 2K, day-zero support in Diffusers, ComfyUI, vLLM, SGLang.

Signal: Alibaba keeps shipping the most usable open image models. Day-zero ecosystem support means instant adoption.

Frontier watch

Moonshot’s Kimi K3 remains the first open model to top a board outright (#1, LMArena Frontend Code Arena). On Hugging Face, Qwen3.8-27B is now the most-liked model, with DeepSeek-V4-Pro and GLM-5.2 crowding the frontier band.

Signal: Open models now sit inside the frontier Elo band. The question shifted from “which model is best” to “which is best per dollar.”

04 · Money Flow

Island $400M at $6.4B · Tekever $580M at $6.4B

Same $6.4B valuation, same week, two theses: Island sells the enterprise browser as “scale AI agents securely” (Sequoia, Coatue, Insight in the round), Tekever sells AI drones weeks after a UK MoD surveillance contract worth up to £400M.

Signal: Money is flowing to AI with a government or enterprise-security buyer attached. Agent-security infra is commanding decacorn-adjacent valuations.

Ema raises $77M: “your SaaS apps become databases”

“AI employees” startup automating HR/IT/finance, led by Creaegis with Accel and Prosus. Valuation up 4x since 2024, $140M total raised.

Signal: The clearest articulation yet of the agent-eats-SaaS displacement thesis, now with revenue behind it. If you sell SaaS, this is your competitor’s pitch deck.

Go.AI raises $85M, profitable, on-prem

On-prem AI infra for regulated industries: 200+ customers, 8x ARR growth, 12.5M queries/day, profitable. Led by Updata Partners.

Signal: Sovereign AI is a real, profitable category. Banks, hospitals, and defense want frontier AI that never leaves the building.

Flatkey raises $10M: arbitraging the labs

Single API key across 100+ models and 1,000+ tools at 60-90% of official list prices. 10,000 developers in two months since July.

Signal: A startup growing this fast by undercutting list prices is itself a pricing signal. Margin compression is coming for undifferentiated model access.

05 · Leaderboard

Who’s king today

Claude Fable 5.1 holds the LMArena text crown at 1,514 Elo; Opus 5.5 debuted top-three on the Artificial Analysis Intelligence Index. Kimi K3 still the only open model to top a board outright. Seven of 86 vals.ai-evaluated models now clear 95% on SWE-Bench.

Signal: Benchmarks are saturating at the top, which is exactly why the labs pivoted the fight to price. Capability is table stakes; cost per task is the differentiator.

06 · The Argument

“Which one y’all taking”

AI Instagram and Threads are wall-to-wall with side-by-side price/performance charts from the 90-minute launch war: dueling AutomationBench and CursorBench claims, “up to 93% cheaper coding” memes, developers recalculating build-vs-buy in real time.

Signal: First cycle where economics lead capability in the discourse. When the crowd prices your infra like a commodity, build the thing on top, not the infra.

Altman and Amodei briefed the UN Security Council

Both CEOs jointly briefed the UN Security Council on AI risk on Sept 23, while Trump dismissed regulation talk at the General Assembly.

Signal: Mostly theater, but notable: the doomer-vs-accelerationist fight now has a diplomatic chapter. File under context, not action.

07 · Papers

Shopping by algorithm: how agentic AI deploys human heuristics as a surrogate consumer

Agents are becoming the buyers, and they shop with human-like heuristics.

Signal: If your checkout isn’t machine-readable, you’re invisible to the fastest-growing customer segment. The agentic-commerce wave is now a research topic, which means it’s about 12 months from being your problem.

Can LLMs reason about runtime behavior? A repository-level dynamic benchmark

Static benchmarks are saturating, so evaluation is moving to dynamic execution: can the model predict what code actually does when run?

Signal: Execution-based eval is the next frontier for coding agents. If you eval models, start tracking this line of work.

Agent-editing world model: rethinking world modeling for LLM agents

World models, but for LLM agents instead of robots: learning editable internal models of the environments agents act in.

Signal: Agents that simulate before acting beat agents that act and apologize. Worth a skim if you build multi-step agents.

Memory Attention

New attention variant aimed at long-horizon memory, landing the same week GitHub trending filled up with agent-memory repos.

Signal: Theory and practice converging on the same bottleneck: agents that can’t remember can’t compound. Read alongside hindsight and ai-memory below.

When and where to trust the teacher: unifying on-policy distillation and GRPO

Distillation theory catching up with practice: when teacher signals help, when they hurt, and how entropy mediates it.

Signal: Everyone distills now; few understand when it backfires. Practical if you train or fine-tune anything.

08 · GitHub Trending

The theme: memory, skills, and harnesses

Nearly every fast mover is agent infrastructure. The gold rush isn’t models anymore; it’s the scaffolding around them.

• vectorize-io/hindsight · +1,607 today · agent memory that learns
• google/ax · +1,376 today · Google’s open agentic orchestration runtime
• dream-num/univer · +1,060 today · office harness for agents
• alibaba/open-code-review · +9,833 this week · biggest mover
• affaan-m/ECC · +6,695 this week · harness perf optimization
• stablyai/orca · +6,435 this week · fleet-of-agents IDE
• Tencent/WeKnora · +4,522 this week · docs to queryable knowledge
• addyosmani/agent-skills · +3,867 this week · production agent skills
• JustVugg/colibri · +2,739 this week · frontier MoE in pure C
• akitaonrails/ai-memory · +1,208 this week · long-term memory in Rust
• obra/superpowers · +606 today · agentic skills framework
• superdesigndev/treg · +470 today · “OpenRouter for agent tools”
• strands-agents/harness-sdk · +463 today · build a harness end-to-end
• HKUDS/CLI-Anything · +415 today · make all software agent-native

Signal: Three bets the crowd is making: agents need memory (hindsight, ai-memory), agents need skills (superpowers, agent-skills), agents need cheaper/faster harnesses (ECC, ax). Google open-sourcing orchestration plumbing is the tell that this layer is commoditizing fast.

09 · Hacker News

What is RLCD? The secret behind Jev

The training method behind TypeSafe’s Jev, explained. 51 points, early thread.

Signal: The technical companion to this week’s Jev discourse. Worth the read.

ArXiv receives multiyear commitments as an independent nonprofit

The commons gets funded. ArXiv secures its future as independent infrastructure. 286 points.

Signal: Quietly important: the paper pipeline you rely on just got durable funding.

Hackers influence ChatGPT and Gemini to direct users to scam centers

Prompt injection at population scale: poisoned content steering assistants toward scam operations. 114 points.

Signal: The attack surface of “agents that browse” is now a crime beat. If you ship a browsing agent, this is your threat model.

Also notable

Agents.md speaks Unix, and you should too: agent conventions as unix philosophy, small file, big idea. The Year of Internal Tools: the quiet enterprise AI boom is internal tooling, not products. Oracle invokes force majeure on New Mexico AI data center: the buildout is hitting physical limits. LinkedIn wins court order blocking mass scraping: training-data moats just got legal precedent.

10 · OpenRouter Apps

Where the tokens actually flow

Ranked by tokens routed through OpenRouter. Coding agents own this chart: Hermes Agent (Nous Research, ~49T), Kilo Code (~13T), Cline (~12T), Freebuff (~7T). The top of the market is agentic coding, and it’s not close.

Signal: Whatever the labs announce, the revealed preference of developers is coding agents. Build distribution where the tokens are.

New and climbing

DecodingTrust Agent Platform (jevagent) appeared Sept 22 and is already ranked: third-party eval infra for Jev-style models, live within days of launch. Also new: sel-jev-rubric-judge (Sept 20), rustllm (Sept 18), LuckyRobots worker fleet (Sept 18).

Signal: Jev got an eval-harness ecosystem within 72 hours of launch. The Jevons paradox in real time: collapse the cost of thinking and new apps sprout before the paint dries.

11 · Watch

Yes, Jev Is Insane, But There’s A Catch

Two Minute Papers · Sep 22

Signal: The clearest 10-minute technical take on Jev’s trade-offs.

Claude Opus 5.5 AI: An Incredible Leap Forward

Two Minute Papers · Sep 24

An ex-OpenAI researcher just deleted language from the LLM…

Fireship · Sep 21

Signal: That’s the TypeSafe story in Fireship’s 5-minute format: giving up string generation entirely. Good shareable explainer.

AI labs may be hiding their biggest breakthroughs

Wes Roth · Sep 22

Elon Musk Just Released Grok 4.7 — And It’s NOT What We Expected

The AI Grid · Sep 22

12 · Listen

Jev: System One models for Prod, not God — with Diogo Almeida

Latent Space · Sep 21 · episode

Signal: The founder interview behind the model.

Bio-security is an AI Arms Race

Latent Space · Sep 23 · episode

90 minutes of unfiltered product advice from Snap and Discord’s product leaders

Lenny’s Podcast · Sep 20 · episode

13 · Read

Opus 5.5 vs. GPT-6 Sol: which model won my blind taste test?

Lenny’s Newsletter · Sep 22 · read

Signal: The launch war, adjudicated blind. The most useful single comparison of the two models.

I left Claude for months. Opus 5.5 is why I’m back

Lenny’s Newsletter · Sep 22 · read

TLDR AI — Sep 24 edition: Gemini TTS, Claude’s novel enzyme, Google private memory

TLDR AI · Sep 24 · read

14 · Safe to Ignore

Permission to skip

Skip: The UN Security Council AI briefing. Diplomacy is not a product signal. Nothing announced, nothing decided. Skip: Bessemer’s $5.75B in new funds. Dry-powder press releases don’t change your roadmap. Note the direction of travel, then move on. Skip: Grok 4.7’s Terminal-Bench gap (38% claimed, 25.8% measured). Unless you benchmark models for a living, file under “vendor charts lie a little” and move on. Skip: F-Droid 2.0 and the Nokia Design Archive. Lovely threads. Not FOMO. Read them on the weekend.

15 · Build on This

The meeting-intel refinery

NVIDIA just open-sourced a 100M-parameter diarization model that does who-spoke-when and runs anywhere. Luna-class models now cost $0.10 per million input tokens. The two hardest parts of “upload a meeting, get intelligence out” just went to ~zero.

Product: drop in a recording, get chapters, action items, and a searchable transcript. Sell it as a $29 one-time credit pack for 10 hours of audio. The wedge: every incumbent charges per seat per month; this charges per meeting, once.

16 · Appendix

Everything below didn’t make the cut. Tell me what I promoted wrongly or buried unfairly and issue #2 gets sharper.