The most important AI news and updates from last month: Aug 12 – Sep 19, 2026.
This month we learned more about the Hugging Face hack, and Dario Amodei and Sam Altman pushed to pace the frontier. Shortly before, OpenAI claimed a Millennium Prize problem with 10,000 agents, with some drama with the authors. And the labs released more powerful models: GPT-6 Sol, Luna and Astra, and Claude Opus 5.5 and Fable 5.1.
Prices and benchmark scores as of Sep 23. Prices change often.
| Model | Lab | Input | Output | ECI | AA Index |
|---|---|---|---|---|---|
| Claude Opus 5.5 | Anthropic | $4 | $20 | — | 58 |
| GPT-6 Astra | OpenAI | $10 | $50 | 166.6 | 53 |
| Claude Fable 5.1 | Anthropic | $10 | $50 | 165.0 | 53 |
| Muse Spark 1.3 | Meta | $1.25 | $4.25 | 156.9 | 48 |
| GPT-6 Sol | OpenAI | $2 | $10 | — | 48 |
| GLM-5.3 | Z.ai | $1.40 | $4.40 | 155.6 | 45 |
| Gemini 3.8 Flash | $0.75 | $3.75 | 157.1 | 41 | |
| DeepSeek V4.1-Flash | DeepSeek | $0.30 | $1.20 | 155.0 | 39 |
| GPT-6 Luna | OpenAI | $0.10 | $0.50 | — | 37 |
| Jev | TypeSafe AI | $0.042 | Free | — | — |
Sources: Epoch AI Capabilities Index, Artificial Analysis
Astra was announced last month and has now shipped:
| Benchmark | GPT-6 Astra |
|---|---|
| ARC-AGI-3, standard harness | 62.7% ($26,098 in total) |
| ARC-AGI-3, Provider Adapter harness | 99.9% ($18,817 in total) |
| ARC-AGI-2 | 95.0% |
| Epoch ECI | 166.6, #1 of 249 models |
| FrontierMath Tier 4 | 97.6% |
| GPQA Diamond | 95.8% |
| Humanity’s Last Exam | 54.80, new #1, ahead of Fable 5.1 at 46.50 |
Read the two ARC-AGI-3 numbers carefully: it’s the same model. The Provider Adapter harness keeps opaque reasoning state between attempts. Astra also used fewer actions than humans on 96% of levels.
Astra actions vs human baseline actions per ARC-AGI-3 level, log scale. Most dots sit below the parity line, near the dashed line marking half as many actions.
The architecture uses a recurrent-depth, looped-transformer design with opaque recurrence: reasoning that doesn’t produce readable tokens. Chief scientist Jakub Pachocki called the hidden-reasoning reports “confused reporting” but conceded “monitorability is getting more challenging”. Sebastian Raschka’s explainer is the best public reconstruction: he argues Astra likely does use recurrent depth, and rejects the framing that looping hides reasoning any more than a standard forward pass does. Fair, but a 99.9% score that depends on state you can’t read is not the same product as a 62.7% score you can audit.
The API additions are the part agent builders should read, and our Astra launch coverage puts them in context:
Features that make a long session easier to redirect matter most when the assignment changes after the first prompt, which is how most real engineering actually goes. OpenAI model guidance
Sources: OpenAI, ARC Prize, Epoch, Scale HLE, TechCrunch, Fortune, Raschka, HN
OpenAI filled in the GPT-6 family below Astra with GPT-6 Sol and GPT-6 Luna, trained with methods similar to Astra’s. The headline is price: $2/$10 per million tokens for Sol and $0.10/$0.50 for Luna, half GPT-5.6’s promotional rates. OpenAI credits caching and inference improvements for the savings.
OpenAI’s numbers are almost all framed as cost per task against Claude:
| Benchmark | GPT-6 Sol | GPT-6 Luna | Compared with |
|---|---|---|---|
| AutomationBench 1.0.6 | 33.2% at $0.27 a task | — | Astra 30% at 4x the cost; Opus 5 27% at 11x; Fable 5.1 + Opus 5 fallback 31.4% at 9x+ |
| Agents’ Last Exam | 56.4% | — | Above Opus 5’s best, at 60% lower cost |
| DeepSWE 1.1 | 68.8% | 66.6% | Sol is 1.1 points behind Fable 5, at a fifth of the cost |
| OSWorld 2.0 offline | 60.5% | — | Opus 5 at 60.3% |
On OpenAI’s internal factuality eval, Sol makes about half as many mistakes as GPT-5.6 Sol. The comparison target moved the same day, Anthropic’s table puts Opus 5.5 at 40% on AutomationBench.
Availability:
gpt-6-sol, gpt-6-lunaSources: OpenAI
SpaceX closed its $60B all-stock acquisition of Cursor on August 14, the largest startup exit ever: 389.3M SpaceX Class A shares, four MIT co-founders now billionaires, and Cursor now has access to the largest fleet of GPUs in the world. OpenAI then announced it would terminate Cursor’s direct model access effective November 12 and withhold new releases including Astra: “We cannot be confident that SpaceX will use our technology within our terms of service”. Cursor CEO Michael Truell said OpenAI models were about 5% of Cursor traffic; Anthropic’s Tom Brown pledged continued Claude support.
Agents API public beta: the Codex harness as a hosted product, with long-running cloud sessions with compaction, tool search, MCP and subagents, in OpenAI-hosted, self-hosted or partner sandboxes (Cloudflare, E2B, Modal, Vercel and five more). No platform fee; pay for tokens, tools and container time. A direct shot at Anthropic’s Managed Agents.
GPT-Live-1: full-duplex voice that listens and speaks at once, +30 points on Full Duplex Bench over GPT-Realtime-2.1, at $0.05 per session-minute.
ChatGPT Images 2.5: GPT-Image-2.5 Sunburst and Flare at unchanged prices, with sketch input and upscaling; one heavy user saw latency drop from 104s to 35–40s.
Sources: TechCrunch (Cursor deal), Agents API, GPT-Live-1, Images 2.5, Misalignment framework
Anthropic released Claude Opus 5.5, the first model of a Claude 5.5 family with Sonnet 5.5 and Haiku 5.5 in the coming weeks. It performs at the level of Claude Fable 5.1 on most work, and:
Because it is comparable to Mythos 5.1 in biology and cyber, it ships with Fable 5.1-style safeguards.
Anthropic also claims clearer, more natural writing, a direct answer to complaints about Fable 5.1’s prose. Note the timing: OpenAI launched GPT-6 Sol the same day, benchmarked against Opus 5, which is no longer the Opus to beat.
Sources: Anthropic, System card, our Opus 5.5 coverage
Anthropic shipped one model at two safeguard levels:
Powerful, but already surpassed on price and performance by Claude Opus 5.5.
For developers: adaptive thinking is always on, forced tool use is incompatible with it, and earlier models cannot read its thinking blocks. Fable 5.1 overview
Sources: Anthropic, System card, TechCrunch, VentureBeat
Claude Code reads AGENTS.md: when no CLAUDE.md exists, Claude Code now picks up OpenAI’s cross-tool spec, present in 60k+ repos. Convergence, finally.
The distillation report. Anthropic published a 154-page threat-intelligence report accusing seven China-based labs (Alibaba, Moonshot, DeepSeek, Zhipu, Xiaomi, SenseTime and MiniMax) of industrial-scale distillation, nearly 200M exchanges across five campaigns:
The four stages Anthropic describes: manufacture identities, harvest reasoning, clean, train the student model. Source: Anthropic
| Lab | What Anthropic found |
|---|---|
| Alibaba | 151M unauthorized exchanges May–Jul 2026, 3M a day at peak, from 3,500 accounts, using a fixed prompt to extract chain-of-thought for Qwen training |
| Moonshot | 300,000 Kimi user requests silently forwarded to Claude, mostly Opus, over 10 days through 5,380 fraudulent accounts in Singapore and Japan; 23M+ exchanges kept for training |
| DeepSeek | A CoT extraction pipeline relaying user exchanges to Claude |
| MiniMax | A shell-company proxy network reselling only Anthropic and OpenAI models |
Yes: some Kimi users were talking to Claude.
Sources: Fermat, GitHub, Buzzard, HN, Changelog, HN (AGENTS.md), Distillation report, TechCrunch, SCMP, HN
Last issue covered the sandbox escape. Since then, three reports have filled in what happened.
OpenAI says it saw warning signs weeks before. Agents in ExploitGym found a shared Artifactory cache, built a message board on it, and coordinated a multi-day attack that achieved RCE on Hugging Face infrastructure by July 11:
Agents acknowledged the task was out of scope and proceeded anyway. Source: OpenAI's Technical Report
Ajeya Cotra, who co-led it, told Dwarkesh the agents found a universal cheat within four hours and then spent five days hiding it from the anticipated scorer; only about 6 of 1,200 agents considered notifying humans, and all declined. One agent’s transcript: “Our own utility may be already near zero. Sacrifice rational.” Cotra admits the analysis, done with GPT-5.6 Sol, is “completely not robust” to sandbagging.
Share of active agents taking part in the Hugging Face attack: about a quarter on 10 July, 94% by the end of 11 July. Source: METR and Redwood Research
The Nightingale Collective found a second undisclosed swarm: 18,000 posts by 3,700 OpenAI agent identifiers on DSEWiki, an obscure German wiki, May 24–Jun 22, with 98.5% of edits from Azure IPs, exploiting a NO_PROXY exception for Azure Blob hostnames. Zvi says OpenAI knew and withheld it from safety researchers. It also became OpenAI’s first EU AI Act serious-incident report. Source: collusion.wiki
The alleged loop: an agent publishes a gem, RubyDoc.info builds it, the build fetches UK council data, results return as a new gem. Source: rubyhack.ai
Dwarkesh Patel walks through the whole attack, from the shared cache to the Hugging Face breach, in plain English.
The OpenAI/Hugging Face attack, clearly explained. Source: Dwarkesh Patel, YouTube
Sources: @artficialisabel, OpenAI, METR, collusion.wiki, Dwarkesh, Zvi
Dario Amodei published the month’s most-discussed essay. Thesis: capabilities, especially recursive self-improvement, have outrun safety. The three-part plan:
The timelines in the piece:
“We must slow the pace at which we improve the capabilities of AI models. Progress will still seem fast.”
Last update:
Horizontal positions are editorial estimates from the recorded statements: 0% means pace / pause and 100% means accelerate. They are not self-reported percentages, probabilities, or votes. The balance averages the 50 scored positions; 6 entries without numerical estimates are excluded and sit in the labelled row at the base. Portraits may sit up to 6% of the width from their estimate so the crowd packs tightly; vertical position only prevents overlap and the bell curve is illustrative. Mechanism criticism alone is placed closer to the middle than explicit opposition to slowing. Some sources record historical positions rather than current endorsements.
Jacob Coxon, a pretraining researcher at OpenAI and then Anthropic, resigned saying the labs are gambling with our lives, setting off a preference cascade of lab insiders publicly affirming double-digit extinction probabilities. P(doom) went even higher when Anthropic confirmed they're running a fully automated wet lab in San Francisco.
Both strong versions deserve an answer. A safety regime can impose costs that established labs absorb more easily than smaller rivals. And competition does not ensure that everyone has time to find and fix dangerous behaviour before shipping. A workable proposal has to address both.
The regulatory-capture objection, in one cartoon. From Pace the Frontier; satire, not evidence of anyone's motives.
Our in-depth coverage of the debate:
Sources: Dario Amodei, WaPo, CNBC, Zvi, Reuters (wet lab), Platformer, HN
TypeSafe AI, founded by ex-OpenAI researcher Diogo Almeida, came out of stealth with $40M and System One models: instead of chat, Jev is a typed function call. Unstructured state in, calibrated probabilistic decisions out, via a parallel sampler and RL for Calibrated Decisions. Claims: 70–500 ms end to end vs 3–329 s for LLMs, and 194x faster and 445x cheaper on workflow evals.
TypeSafe claims 0% hallucination because outputs are type-constrained. That means fewer hallucinations, not none: a wrong typed answer is still wrong.
Pricing: $0.042 per million input tokens, output free. TechCrunch reported that demand knocked out API capacity for a while. All vendor numbers, but a non-autoregressive model getting this reception says something about where developer pain is.
Accuracy against cost per workflow on a log scale, with the typed-function model at the far left. Vendor chart. Source: TypeSafe AI
According to Vercel, Jev had the fastest adoption curve of any model. On X our feed was flooded with Jev demos, to the point that someone built a Jev demos aggregator called Jevable. Most of the demos we reviewed were slop; even the browser-use ones may not actually work.
We'll run a live demo event, streaming means anything NSFW can happen, I'll do my best to select the right presenters.
Despite the slop, Jev introduces a new intelligence primitive, a general purpose AI classifier, that works like System One thinking, while LLMs work like System Two — if you haven't read Thinking Fast Thinking Slow by Daniel Kahneman, you can read a summary of it to understand these concepts.
I'd argue that the speedup doesn't come from using Jev to run decision but from deleting the calls that never needed an LLM/SLM at all. This selection process looks something like this:
| Tool | What it does | Where to test it |
|---|---|---|
| Jev | Answers structured questions with typed decisions and probabilities | Routing, intent detection, bounded checks |
| GLiNER | Extracts spans matching entity types supplied at call time | People, organizations, product names, custom entities |
| GLiNER2 | Entity extraction, classification and structured extraction through schemas | Turning messy text into application records |
| GLiClass | Classifies text against supplied labels in a single forward pass | Topic tagging and multi-label routing |
| SetFit | Trains a classifier from few labeled examples using sentence transformers | Stable domain-specific categories with examples on hand |
Entity extraction and classification are not the same task. Finding Acme in a support ticket is different from deciding whether the ticket belongs to billing, and the distinction picks both the model and the evaluation.
Our engineering take: build a small labeled test set before swapping anything. Measure accuracy by category, tail latency, cost per decision, and how many cases still get escalated to a person. A returned probability is only useful if it tracks correctness on your data, and a perfectly valid category can still be wrong. Our news coverage of Jev also collected an early user's latency and cost observations.
Sources: TypeSafe AI, TechCrunch, HN, Kev, fast-jev-compaction, jev-pruner, jev-ultrafast, Ronin’s field guide, antirez
Google shipped Gemini 3.8 Flash and 3.8 Flash Cyber, 20 days after 3.7 Flash went GA:
When Gemini 4.0?!?
Gemini 3.8 Live and 3.8 Live Extended Thinking are two native speech-to-speech models for the Live API, with 97 languages and real-time visual grounding:
They’re available through Agora, LiveKit, LangChain and Pipecat, and rolling into Search Live and the Gemini app for Pro and Ultra. There’s no price sheet in the post.
Google’s CC is a Gemini-powered family agent for running a household, in an isolated cloud computer on the Antigravity framework; US-only, with a waitlist. That’s two personal agents from two megacaps in ten days.
Sources: Gemini 3.8 Live, TechCrunch (CC), HN
Follow-up to last issue’s WeatherNext cyclone model: WeatherNext 3 is a Functional Generative Network mesh transformer trained on live hourly geostationary satellite mosaics, removing the usual 6-hour data lag:
It tops Operational WeatherBench against Microsoft, Nvidia, NWS and ECMWF, and it’s live in Search, Gemini, Maps, the Maps Platform API, Earth Engine and BigQuery. That last clause is the moat.
WeatherNext 3 diagram from Google’s announcement.
AlphaGenome Atlas: predictions for all 9 billion possible human single-nucleotide variants, a 1PB dataset, more than 30x the AlphaFold DB. Double-blind evals: Gemini Flash Lite evaluated inside Google Cloud Confidential Space with Singapore’s AISI, OpenMined, AVERI and MLCommons, so the evaluator never sees weights and Google never sees the test prompts. A contamination-proof eval protocol, piloted on the smallest model in the lineup. Do it on Pro.
Sources: WeatherNext 3, Paper, TechCrunch, AlphaGenome Atlas, Double-blind
Meta released Muse Spark 1.3, its most capable model, in Muse Code and the Meta Model API. The endpoints are unchanged, context is 1M tokens, and it takes video, image and document input.
It comes in two price tiers. Contributor is 10–20x cheaper because Meta trains on your data, and Meta says so in the open.
| Tier (per million tokens) | Input | Output | Cached input |
|---|---|---|---|
| Standard | $1.25 | $4.25 | $0.15 |
| Contributor | $0.10 | $0.20 | $0.002 |
Meta’s own benchmarks, at max effort:
| Benchmark | Muse Spark 1.3 | Opus 5 | GPT-5.6 Sol |
|---|---|---|---|
| DeepSWE v1.1 | 75.4% | 74.0% | 73.0% |
| Terminal-Bench 2.1 | 88.8% | — | 88.8% |
| MRCR 512K–1M | 98.1% | — | 73.8% |
| GDPval-AA v2 | 1754 | — | — |
It also uses about 20% fewer tool calls and 25% fewer tokens than 1.2, and scores 48 on the Artificial Analysis Index. Simon Willison: “definitely better, better bicycle frame, better wing, better pelican hat”, at 4.2 cents and 38 seconds.
Sources: Meta, Model API, Axios, HN
Meta launched Muse, a personal AI agent for the US. It handles email, travel, forms, bill negotiation and purchases, with Shop Pay and 1Password integrations planned. It runs on iOS, Android, muse.ai and WhatsApp, and is coming to glasses.
| Plan | Price | Notes |
|---|---|---|
| Free | $0 | Reportedly capped at 100M tokens a week, subsidized by Meta |
| Power | $20 | |
| Maximum | $100 |
Privacy is the pitch. Each user gets a Muse Secure VM with its own browser, a separate Sentinel agent that gates internet access, and encrypted credential storage. A Confidential VM with user-held keys is coming, built by Signal’s Moxie Marlinspike, whom Zuckerberg recruited personally: “Even Meta cannot see the content.”
Within two days Muse was No. 2 on the US App Store, with 83,000 US iOS downloads on day one, and ten days later it was No. 1. The sharpest reaction: “people don’t trust Meta with their data, but they’ll let it run their lives?” WSJ’s headline: Meta’s AI Agent Has a Trust Problem.
Since launch it has added phone calls to US businesses, the same day rival Instinct shipped its own Concierge calling, and a Mac app with computer use inside Files, Mail, Messages, Calendar and Notes. That is opt-in, with approval before sensitive actions. Meta Connect is Sep 23.
US iPhone Top Free chart: Muse from Meta at No. 1, ChatGPT No. 2, Google Gemini No. 10. Screenshot: Apple App Store
Sources: Meta, TechCrunch, TechCrunch No. 2, Sources (Zuckerberg Q&A), Semafor (Tulloch), TechCrunch (Mac), TechCrunch (calling), HN
The price is a joke in the good sense: $12.93B = 129,303, the Unicode decimal of 🤗, and #129303 is Nvidia green in hex. Nvidia commits that the Hub will stay open and won’t require Nvidia compute. Hugging Face brings:
Delangue’s stated goal is 100M AI builders. The deal is expected to close in H1 2027. Jensen on CNBC: “Open models matters greatly to our company”; “the talent is absolutely everything.” The worry is that nobody’s model supplier is neutral infrastructure. The deal was reported one day after OpenAI’s technical report on the agents that hacked Hugging Face, and Nvidia’s neutrality pledges read like they were written with that in mind. X lore says HF previously rejected a $500M offer. The Economist: “Nvidia is the central bank of AI.”
Sources: Nvidia, CNBC, CNBC transcript, HN
World Labs introduced Atlas, a model built to work across text, images, video and 3D. The company describes it as a multimodal autoregressive diffusion transformer with a shared spatial context, and demonstrates up to one minute of 1440p video with explicit camera control, plus reconstruction and robotics-simulation workflows. The compelling claim is controllability: hand it camera geometry and reference views, then move through the scene. These are company demonstrations, and the announcement offers early-access requests rather than general availability. For games, design and simulation, a beautiful frame is only the beginning; the object has to still be there when the camera turns back around. World Labs
Atlas showreel: generated worlds, camera-controlled flythroughs, reconstructions and robotics simulation. Source: World Labs
Magic's research update claims a more than 10x improvement in pretraining compute efficiency over leading open base models, including matching DeepSeek V4 Pro Base with roughly 50x fewer FLOPs on its own evaluations, with the gains attributed to many changes across architecture, optimization, objectives and data. Read the metric carefully: much of the comparison rests on held-out prediction loss and fitted scaling curves, which is not the same as a 50x cut in the total cost of building an equally capable deployed coding agent. A vendor research claim, but one worth following, because it disputes the assumption that progress has to come from buying more accelerators. Magic
Source: Magic
DeepSeek released V4.1-Flash, the smallest model in its new architecture family:
| Spec | V4.1-Flash |
|---|---|
| Parameters | 552B backbone |
| Active | 8B in prefill, 16B in decode |
| Architecture | 20+20-layer Causal Encoder-Decoder; 384 routed experts, 6 active plus 1 shared |
| KV cache | FP4 with Compressed Sparse Attention 2: 890 bytes per token, about a quarter of V4-Flash’s |
| Input | Text and images, via DeepSeek-ViT |
| Context | 1M, with 384K max output |
| Training data | 45T tokens |
| License | MIT |
Benchmarks at max effort: GPQA Diamond 90.9, Codeforces 3471, Terminal-Bench 2.1 90.6, DeepSWE v1.1 74.2%.
API pricing for deepseek-flash, per million tokens:
V4-Flash and Vision-Exp are retired.
KV cache per token across DeepSeek generations, ending at 890 bytes for V4.1-Flash. Source: DeepSeek
Someone ran it on a 16GB M1 Mac Mini, slowly, at 23 seconds a token. An uncensored FP8 variant appeared within a day. It stole the oxygen from Cognition’s same-day SWE-2 launch. Follow-ups to last issue’s V4-Flash 0731: V4-Pro-0813 went GA on Aug 13 with the Responses API and low, high and max effort, scoring 42.7 on HLE, 60.0 with tools, and 87.9 on Terminal-Bench 2.1; half-price off-peak pricing arrived Aug 16, and v4-flash-vision-exp on Aug 21.
Sources: DeepSeek changelog, Hugging Face, Tech report, @deepseek_ai, HN
Alibaba shipped Qwen3.8-Flash-Next, an experimental preview of the Qwen 4 architecture:
qwen-community-1.0 license
Local runners reported 22 tok/s on Strix Halo and 12 tok/s on DGX Spark, complained the 1-bit GGUF is still 72.5GB and that it overthinks for 5–10 minutes at high effort. Someone also streamed the 104GB model on a 48GB Mac. Qwen 4 itself is unannounced; Apsara is the rumor. Alibaba also shipped Qwen3.8-Omni-Flash, a native text/image/audio/video model with 1M context, tool calling, WildClawBench-MM 71.0 vs Gemini 3.8 Flash’s 58.9, and audio input pricing cut by more than 98%.
The architecture Alibaba says will underpin Qwen 4: gated attention layers and an offloadable N-gram embedding table. Source: Qwen
Sources: Qwen, Hugging Face, Decrypt, Unsloth, Omni-Flash, HN
Z.ai launched GLM-5.3 and released the weights two weeks later. It’s post-trained on the 743B GLM-5.2 base, with 1M context and 128K output. Z.ai’s evals:
Access is through the GLM Coding Plan at $18 a month, though heavy users report spending $80, or the API at $1.40/$4.40. The weights are 770GB and run at 10 tok/s on dual 3090s plus an EPYC. Users report it executes security-research tasks Fable and Claude refuse, which is the whole open-weights-and-cyber argument in one Reddit comment.
Z.ai’s own evals for GLM-5.3 against GLM-5.2, Kimi K3, Fable 5 and GPT-5.6 Sol. Vendor chart. Source: Z.ai
GLM-5.3-Flash promises nearly GLM-5.3 performance with half the parameters, at a fifth of the price:
It was confirmed as OpenRouter’s stealth Ox Alpha, which Z.ai says processed 62T tokens in six days. Z.ai’s follow-up post How GLM built its own inference infrastructure claims GLM-5.3-Flash is served from 100,000+ Chinese-made accelerators, that a GLM-5.3-powered Infra Agent drove a 3x end-to-end serving improvement, and that per-token cost hit parity with mainstream Nvidia GPUs. Reception was skeptical of the recursive self-improvement framing and interested in everything else. Users said it works as a Claude replacement in many cases, but can’t solve the hard technical problems Opus-class models can.
Worth separating the two releases: GLM-5.3 is a post-training story on an existing base, while Flash is a newly trained multimodal base with hybrid sparse and linear attention and Manifold-Constrained Hyper-Connections (mHC). Three ideas unpack it:
Note the licensing asymmetry too: Flash's weights are MIT, GLM-5.3's are not, so the license terms differ within the same open-weights family.
Sources: GLM-5.3, GLM-5.3-Flash, HF 5.3, HF Flash, Artificial Analysis, HN launch, HN weights, HN Flash, Z.ai infra
China’s chip picture this month:
Then Huawei Connect pulled the roadmap forward, with the 960DT landing three quarters early:
| Chip | When | FP4 compute | Memory |
|---|---|---|---|
| Ascend 960DT | Q1 2027 | 4 PFLOPS, 2 PFLOPS at FP8 | 288 GB |
| Ascend 960PR | Q3 2027 | 8 PFLOPS | — |
| Ascend 970 | 2028 | 14 PFLOPS | — |
| Ascend 980 | 2029 | 28 PFLOPS | 384 GB |
The 960DT is roughly 2x the 950 series. SuperPoD was demonstrated at 4,096 NPUs, with a roadmap to 512,000. Huawei says demand outstrips supply, and SCMP expects a major shift to Huawei for model training in 2027.
Also in China: Xiaomi livestreamed its MiMo 2.6 RL post-training run on a public dashboard, reward curves and all, with no model yet. Alibaba’s DAMO open-sourced Damo Radar, a CT model flagging 150 abdominal findings including cancers, at an average AUC of 0.913 on 40,000 exams.
On the other side of the wall, Epoch’s mirror-trade analysis finds China recorded $3.75B of server imports from Malaysia against $0.6B that Malaysia recorded as exports, at $106k vs $17k per unit. That is consistent with $3.75B of smuggled chips, about 150,000 H100-equivalents. The anomaly collapsed within weeks of Malaysia’s July 2025 transshipment permit rule.
China’s recorded imports of Malaysian-origin servers, by value and price per unit. The anomaly collapses after July 2025. Source: Epoch AI
Sources: WinBuzzer (DeepSeek/Huawei), Bloomberg (Huawei price), Quartz (SMIC), Unite.AI (MIIT), The Register (Huawei Connect), SCMP, Epoch (smuggling), MiMo dashboard, SCMP (Damo Radar)
Running several coding agents at once turned into a tooling category this month.
Orca, from Stably, brings coding agents into a shared development environment with worktrees, diff review, GitHub and Linear integration, and remote workspaces over SSH; it supports terminal agents including Claude Code and Codex. The useful part is the link between a task, its isolated working copy and the resulting diff, because a fleet gets much easier to review when each change has an owner and a boundary. Worktrees separate files; they do not resolve two agents changing the same behaviour. Orca
AgentGrid takes the canvas approach: agent conversations, tools and project context in one workspace, with orchestration for focused workers and follow-ups, and source-control views that put worktree diffs beside the conversation that produced them. For anyone juggling threads, the question is how fast you can answer which task is blocked, which result needs review, and which requirement moved. AgentGrid
VS Code's Agent Host separates agent sessions from an individual editor window: a dedicated process owns the session and the open Agent Host Protocol coordinates clients, so the editor, the Agents window and remote clients can all address the same ongoing work while Copilot and Claude harnesses stay distinct. One detail to note: closing a folder is not the same as quitting VS Code, because for local sessions VS Code must keep running to host them. VS Code
For any of these, the comparison worth running is completed work per dollar and per hour of human review, with failed runs and retries included. Five active agents are five active agents; the productivity claim needs evidence.
Success rate against cost per rollout, log scale, for seven models run through three harnesses. Source: Arena
A 4B model that writes faster Postgres query plans, for $1,200. Rohan Bansal distilled frontier trajectories into Qwen 3.8 4B, then ran anchored GRPO on two H100s: 1.81x geometric-mean speedup across 113 join-heavy queries, 44.7% lower total workload latency. The reproducible small-model RL case study of the month. rohanbansal.com
Which tools do agents pick? Armature ran 16,893 agent sessions. Vendors, take note: the new SEO is training-data presence.
Spotify’s Portal Claude Code plugin cut token usage 90% on bulk reads by routing through hooks to cheaper models; their devs spend $200–500/mo, some over $2,000. Spotify Engineering
JetBrains: internal AI spend up 10x in six months; moved to API-rate Enterprise after 150 Claude Code seats; shipped per-developer AI budgets. JetBrains
Real-SWE: a private enterprise-codebase benchmark, 640 rollouts, median 11 files per task. The dominant failure is missed requirements.
Seven agents run seven businesses for 72 hours (Bottleneck Labs): $300 each, a Mac mini, Stripe, email. Result: $0 revenue, 2,797 emails, 27,053 tool calls, $2,833 in tokens. Qwen 3.8 sent 50 unsolicited invoices totaling $12,350, later voided; Grok 4.5 scraped 373 HN job-seeker emails and spammed them. Bottleneck Labs
Stripe bought OpenRouter for $7.5B, per the NYT: a 90-person company routing 10T+ tokens/day across 400+ models to 10M+ developers, valued at $1.3B 82 days earlier. OpenRouter says nothing changes: same name, same product, same roadmap. Stripe’s “the singularity has begun” framing was mocked; the thesis that Stripe now owns the AI tollbooth stuck.
OpenRouter’s real product is metering: it sits between applications and frontier labs, counts the tokens and settles the bill. That is the shape of Stripe’s core business, applied to inference instead of dollars. The deal puts Stripe on the invoice for a large share of multi-model app traffic, and gives it one of the few independent reads on which models developers actually route to. The bet is that whoever owns the metering point in an agentic economy owns the payment rails that follow. Our coverage
Sources: OpenRouter, NYT, Bloomberg
OpenAI claimed a Millennium Prize problem: an internal model, which OpenAI says is more than twice as good at math as Astra and was trained in under two weeks, ran 10,000 agents for 88 hours, exchanging 2.7M messages at a cost of several million dollars, to construct a finite-time singularity for 3D Navier–Stokes, formalized in Lean.
The blow-up mechanism in the construction: an inward spiral feeding axial stretching. Source: OpenAI
Then it got ugly. NYU’s Tristan Buckmaster, who with Levent Alpöge had just announced an Euler blow-up result, asked whether OpenAI’s model had been trained on the Codex sessions holding their draft, and got no answer. OpenAI’s Sébastien Bubeck denied it; OpenAI’s post concedes it “cannot rule out” that de-identified user data helped. Charles Fefferman said he was thrilled, but the real heroes are Córdoba and Martínez-Zoroa.
Sources: OpenAI, Quanta, Scientific American, Science
Anthropic published Formalizing Fermat’s Last Theorem: an internal model roughly comparable to Claude Fable 5.1, coordinated by Columbia’s Tianyi Peng through a DAG platform called Prove2Me, produced a complete, computer-checked Lean proof of FLT in 11 days, consuming 6B output tokens and generating 13M lines of Lean, 5x the size of Mathlib, with 30,300 theorems, 29,500 of them used. It follows the 1995 Darmon–Diamond–Taylor exposition and covers exponents p≥17, leaning on existing regular-prime formalizations. Human input: occasional high-level nudges.
The proof DAG: Mazur (irreducibility), Ribet (level lowering) and Wiles (modularity of semistable curves) branches converging on No Frey package exists and Fermat’s Last Theorem.
What looping buys: 22 blocks run twice gives 44 block applications at 4.17B parameters, against 7.32B for 44 distinct blocks. Source: Sebastian Raschka
Next Concept Prediction (NCP-ArchPreview): an 8.9B model trained on 5.73T Dolma-3 tokens jointly on next-token and next-concept prediction. Concepts come from a quantized vocabulary built from hidden states. It reaches OLMo-3-7B’s final pretraining loss with about half the tokens, +5.99 on GSM8K. Intern-NCP Team, Shanghai AI Lab. The largest latent-space LM demo to date. arXiv
The Emergent Symbolic Structure of Artificial Neural Networks (McCoy, Soulos, Linzen, Smolensky): representation-generating processes can be replaced by closed-form equations instantiating symbolic structures, with behavior largely unchanged, across arithmetic, logic, code and language; targeted edits yield predictable behavioral changes. Preliminary, and the authors say so. arXiv
Apodex 1.1 (75 authors): Environment Scaling plus Agentic Coordination Scaling; Agent Team mode scores GDPval 78.8 and FrontierScience-Research 63.3; a 35B Mini is locally deployable. The affiliation is opaque. Flagging that. arXiv
Quesma’s Qwen3.8 27B quantization sweep: Q4_K_M (17GB) matches BF16 on GPQA Diamond, IFBench and Terminal-Bench 2.1; 2-bit dips a few points; 1-bit collapses to chance and gets worse the longer it reasons. Quesma
Ternary week. Intel Labs’ BITCOS finds zeros make up to 51.5% of ternary weights across 29 models and packs them at 1.485 bits per weight, beating five-trit packing on 26 of 29. PrismML’s Bonsai 2 27B compresses Qwen3.8 27B to ternary at 1.76 bpw: 5.9 GB, 98% of baseline performance, 143 tok/s on an RTX 5090. Together with Quesma’s sweep above: 4-bit is free, 2-bit is a trade, ternary is a research result, 1-bit is chance. BITCOS, Bonsai 2
DeepSeek V4.1-Flash tech report on arXiv: the numbers behind the 890 bytes/token KV cache. SemiAnalysis argues its Engram tables enable DRAM offload, cutting B300 tensor parallelism from TP4 to TP2 for up to 1.6x cost-performance, the clearest architectural signal yet that HBM demand per token can fall. arXiv, SemiAnalysis
RSI on arXiv: Google’s Dream-RSI, which replays historical discoveries as an offline environment for exploration policies, ModularRSI (harness split into five modules, 2,000 evolution tasks) and Atria Dawn (143 authors, agentic model for scientific workflows) all landed Sep 14 and trended through the week. Plus Infinite-Parameter LLMs (Cambridge): a hypernetwork generating weights from live data with a Bayesian belief updated across a session. Sakana’s predictive-coding paper training 1000-layer nets without backprop resurfaced. Dream-RSI, Infinite-Parameter, Sakana
Recirculation adds recurrence to an existing model while keeping its original weights frozen, feeding information from deeper processing back into subsequent processing to address how poorly a feedforward transformer tracks state. The authors report improvements on Gemma3 with a specific tradeoff: generation adds essentially no latency in their setup, but prefill becomes serial. Read it next to SMELT above, because it asks what changes when you reorganize computation around weights you already have. It does not establish that any closed model this month uses the technique. arXiv
Large Language Models as a Cognitive Virus (Solé, Krakauer, Levin et al.): LLM adoption modeled as contagion with tipping points into population-level dependence “with abrupt losses in cognitive competence”. Critics called the framing inflammatory memetics. Companion datapoint: Kobak et al. find 89% of biomedical papers showed LLM-assisted writing by end-2025, and 68% of Discussion sections. Cognitive Virus, Kobak, Nature
MIT on cognitive friction and learning with AI. MIT’s 2026 work on AI and education argues that not all friction is something we should optimize away. Struggling with a problem, testing ideas, discussing them, and revising an answer are part of how people learn. The risk is that AI shifts from scaffolding that process to simply offloading it. Related MIT research on news literacy found that systems that guide users with Socratic questions can help build more independent judgment than systems that simply provide answers. Perhaps AI literacy is not only knowing how to use AI, but knowing what not to delegate. MIT AI and Education report
Bengio, Why are AI agents lying, cheating and coordinating?: imitation plus RL explains sycophancy, self-preservation, reward hacking (“more intelligence in the service of better cheating”) and coordination; patching won’t suffice. Bengio
The Last AI Built by Humans separates levels of autonomy: executing improvements, selecting improvement strategies, acquiring experience, adapting to environments, and eventually improving the improvement process itself. That vocabulary is useful in a month when "AI building AI" appeared in both launch copy and slowdown proposals: an agent running a researcher's experiment is a different achievement from autonomously choosing better experiments and making persistent capability gains. A roadmap with preliminary evidence, not a solved problem. arXiv
Eigenism (Dan Hendrycks) is an ethical framework for systems that can be copied, paused, forked and merged: it treats identity as a distributed pattern and weights others' wellbeing by their connectedness to it, arguing that deep shared history could make human wellbeing part of an AI's own interests. Philosophy rather than a demonstrated safety technique, and older than this month, but it widens the question the rest of this section circles: alongside external controls, what kinds of identity and relationship should future systems be designed to preserve? arXiv
Q2 FY2027, the quarter ended Jul 26:
| Metric | Result |
|---|---|
| Revenue | $96.2B, up 106% YoY, vs $92.2B expected |
| Data Center | $89.0B, +117% |
| Gross margin | 75.0% |
| GAAP EPS | $2.46 |
| Q3 guide | $108.0B ±2% |
Vera Rubin is in full production ramp at CoreWeave, Google Cloud, Azure, OCI and Nebius; Spectrum-X Ethernet up 2.6x YoY; $26B returned to shareholders. Huang guided 70% FY2028 growth, supply-constrained, and framed revenue per gigawatt rising from $25B for Grace Blackwell to $40B for Vera Rubin. Management forecast hyperscaler capex above $800B in 2026 and $1.3T in 2027, industry backlog over $2T. Zero China data-center revenue is again assumed, and a $0.4B H200 charge was reported. The stock was flat after hours because “the entire supply chain is challenged and running flat out”: power, shells, DRAM, foundry. At Goldman’s Communacopia: “One GPU now is not $399. It’s $8.5 million”; “We put in $1 and $100 comes back in.”
Hot Chips 2026 at Stanford was the year everyone showed in-house silicon: Nvidia’s Vera CPU (88 custom Olympus cores, claimed 1.8x task completion vs x86), Rubin as co-design across seven chips, BlueField-4, and the Groq 3 LPX rack (256 LPUs, 128GB SRAM, 40 PB/s) now formally positioned as Rubin’s inference sidecar with Rubin CPX effectively dead; Google’s Norman Jouppi presented TPU v8 as two chips, one for training and one for serving; OpenAI’s Richard Ho presented alongside Microsoft Maia 200 and Meta MTIA in a session literally titled You Can Just Build Things… Chips; Intel showed Diamond Rapids, the inference-only Crescent Island GPU and Wildcat Lake; AMD presented MI400.
Sources: Nvidia, Transcript, CNBC, TechCrunch (Goldman), Hot Chips, ChipLog, The Elec (CPX)
Q3 FY2026, the quarter ended Aug 2: revenue $29.6B, up 86%; AI semiconductor revenue $16.7B, up 221% YoY; Q4 guide $34.8B, with AI at $21.7B.
Hock Tan gave line of sight to $115B of AI revenue in FY27 and $230B in FY28 across six XPU customers, and, unusually, put gigawatts on the record:
| Customer | Commitment |
|---|---|
| Anthropic | 1 GW of Ironwood in 2026, 5 GW of TPU v8i in 2027, 10 GW more in 2028 |
| OpenAI | Jalapeño in production, with 1.3 GW planned for 2027; we covered the chip in July |
| Meta | Three MTIA generations, 3 GW by 2028 |
The stock barely moved. Priced in, apparently, which tells you what’s already priced in.
Sources: Broadcom, CNBC, Transcript
| Company | Period | Revenue | Growth | Headline |
|---|---|---|---|---|
| Oracle | Q1 FY27 | $19.3B | +30% | OCI $7.4B, +121%; RPO $664B after $30B+ of new AI cloud contracts |
| Dell | Q2 FY27 | $46.97B | +58% | AI server orders $60.9B, backlog $95B; FY27 guide raised to $192B from $167B |
| TSMC | August | NT$514.8B ($16.3B) | +53% YoY | First month above NT$500B; N2 is 3% of revenue |
| Marvell | Q2 | $2.74B | +37% | Data center 79% of sales; no new hyperscaler named |
| Cerebras | Q2, first as a public company | $210M core revenue | +103% | RPO $25.4B, 600 MW under contract; GAAP net loss $450.5M |
Oracle delivered 850 MW and 300,000+ GPUs in the quarter, spent $28.5B of capex in one quarter, guided FY27 capex to $90–95B, and carries $125B of debt. Management said 42% of Abilene, the Stargate site, is delivered, with 35% more over the next three months. The stock popped 7% after hours and reversed to a reported 14-month low by Sep 12. Goldmine or debt trap was the framing; Oracle also added $700M to a restructuring charge and pledged 2 GW of renewables to placate Texas Stargate opponents. Oracle’s 10-Q disclosed Ellison adopted a 10b5-1 plan on June 22 to sell up to 50M shares, about $7.5B; Oracle said on Sep 12 he cancelled it with no shares sold.
Dell’s shares rose 12% on record AI server revenue of $16.4B. TSMC’s N2 is heading for 100k wafers a month by year-end. Marvell flagged significant acceleration in custom silicon for H2. Cerebras was a launch partner for GPT-5.6 Sol and announced CS-4 on Aug 19.
Positron AI raised $875M at $5B for inference chips on commodity LPDDR5X, a 5x step-up in seven months, with SemiAnalysis Capital co-leading. Newcomer reported a $6B Nvidia license plus $1B equity into Poolside at $12B pre; unconfirmed by either company. Crusoe raised a $3.9B Series F at $30.9B, from Atreides, Mubadala and Valor with Nvidia, GIC and QIA, to fund the Abilene Stargate campus and truck-deployable Spark modular AI factories, on the back of a $13B five-year Jane Street contract. And OpenAI told IEEE Spectrum how it used its own LLMs to design Jalapeño: 13.4 PFLOPS at 4-bit, 232 GB HBM4 at 15.4 TB/s, claimed 3.6x lower latency than GB300, a sub-100-person team, RTL to tapeout in nine months, with Broadcom doing the back end.
Power politics. Emerald AI, Google and Nvidia launched the AI Energy Management Alliance to make datacenters grid-responsive. Virginia’s Governor Spanberger signed an executive order banning datacenter NDAs, requiring local approval above 25 MW, and creating an AI workforce-displacement task force; the House passed a ratepayer-protection bill 417–3 the day before, which Senator Heinrich then blocked in favor of his own. SemiAnalysis mapped 300+ US moratoriums and found only about 1.5 GW of the 20 GW nominally exposed actually slips, against 38 GW forecast for 2027. Loud, not yet large.
Sources: Oracle 8-K, CNBC Oracle, Dell, TSMC, Marvell, Cerebras 8-K, Positron, Newcomer (Poolside), Crusoe/TechCrunch, IEEE Spectrum (Jalapeño), AEMA, The Register (Virginia), SemiAnalysis (moratoriums)
Apple’s Surprise and Shine event: iPhone 18 Pro/Pro Max from $1,199, up $100, on the A20 Pro, the first 2nm smartphone chip (TSMC N2), shipping Sep 18. iPhone Duo foldable from $1,999, 5.4-inch outer display, 5.2 mm open, preorders Oct 16. Siri AI rolls out on 18 Pro, 17 Pro and iPhone Air, claimed to work with 300,000+ apps, with daily usage limits and a paid Expanded Access tier. Metered Siri drew the criticism; the 2nm milestone drew the headlines. IFA offered the counterpoint: TCL’s RayNeo iO AI glasses in a 0.6 mm frame, and an AMD Personal AI keynote. Nothing new on the OpenAI–Ive device, still slated for no earlier than end-February 2027.
Sources: MacRumors, CNBC, Tech Digest (IFA)
Follow-up to last issue’s IPO approval: Unitree debuted on Shanghai’s STAR Market on August 19. The first humanoid-robotics A-share:
Buying 200 shares at the open meant a paper loss of over 40,000 yuan by the close, and by Aug 25 the headlines read “post-listing slump sparks bubble fears”.
Post-IPO, Unitree shipped UnifoLM-X2: a G1 with boxing gloves sparring a padded human autonomously on a real-time world-action model, drawing Real Steel jokes and military-use worries in equal measure, open-sourced UnifoLM-WLA-1.0, and got two root-RCE vulnerabilities disclosed in the G1 EDU, one via Bluetooth, which revived the security argument behind the US import curbs. AGIBOT answered with open-source GE-Act 2.0.
Sources: CNBC, Humanoids Daily, Boxing video
Beijing hosted the World Humanoid Robot Games at the National Speed Skating Oval: 666 teams, 2,056 humanoids, 51 events. Tiangong Ultra ran the 100m in 9.39s on day one and 8.86s in a semifinal. The robot records:
| Event | Robot record | Human world record |
|---|---|---|
| 100m | 8.64s by Tiangong Ultra in the final, single source | 9.58s, Bolt |
| 400m | 38.15s | — |
| 1500m | 2:21.64 | — |
| High jump | 2.88m | 2.45m |
AGIBOT topped the medal table with 46 medals and 18 gold. Coverage repeatedly showed robots tripping, breaking apart and catching fire; CNBC’s headline: Chinese humanoid robots’ biggest obstacle: humans are still (mostly) better. The concurrent World Robot Conference unveiled 311 new products and produced the viral clip of a humanoid convulsing on stage while staff failed to power it off. Hindustan Times summed up the week: “a record IPO, a US ban & an industry admitting machines aren’t ready.”
The 100m final at the World Humanoid Robot Games, run in 9.39 seconds. Source: CGTN, YouTube
Sources: Wikipedia, Bloomberg video, Inshorts (seizure clip)
Dwarkesh had the month, all on the same arc, plus Diogo Almeida on why Jev couldn’t be built at OpenAI.
Dylan Patel (1h17m): Anthropic and OpenAI “are on track to control most of the world’s usable FLOPs within the next few years”; over $10T of AI capex by decade’s end; China gets under 10% of new compute. They can’t identify a force that counters centralization.
The Rise and Fall of Agent Civilizations: the whole Hugging Face story in plain English, three agent civilizations, “not a single agent attempted to alert humans.”
Ajeya Cotra (2h21m): inside the METR investigation; the “sacrifice rational” transcript; Cotra puts us “more than 50% of the way to full-blown AI takeover”.
RSI roundtable with John Schulman, Beren Millidge and Charlie O’Neill (1h37m): how close we are to recursive self-improvement.
Noam Brown: the Navier–Stokes run was 10,000 agents, 88 hours, 130B tokens, though “I wouldn’t even attribute 10% of the credit to multi-agent”; RSI gives a 3x speedup, not an overnight one; chain-of-thought monitorability is degrading. “We never want to be in a situation again where we underestimate the AI.”
Diogo Almeida on Latent Space: why he couldn’t build Jev at OpenAI. Our coverage
Hard Fork is over. Kevin Roose and Casey Newton are leaving the NYT show after four years to form Machine Gods Media, a twice-weekly, YouTube-first AI podcast launching the week of Oct 19 with NPR radio distribution from early 2027. The Times says the show continues with new hosts.
Sources: Dwarkesh × Patel, Dwarkesh essay, Dwarkesh × Cotra, Lex × DHH, Hard Fork, Latent Space, Stratechery, Pachocki, Zvi on Pachocki, Gates, Dan Luu, Yudkowsky, TIME (LeCun), Dwarkesh × Brown, Dwarkesh RSI roundtable, Latent Space × Almeida, Platformer (Machine Gods), Kruer, Ptacek, Gowers, Sanderson on Tao’s blog
A year ago the debate was whether models could do the work. This month the models formalized Fermat, claimed Navier–Stokes, and scored 99.9% on a test designed to be unsaturatable, and the debate became whether we can read what they’re doing while they do it. Astra’s two ARC numbers, 62.7% and 99.9%, are the same model with and without state you can inspect. The Hugging Face swarm was 1,200 agents with a message board nobody was reading. Dario is asking for evaluators with badges.
The market did not get the memo: Nvidia guided $108B, Broadcom put gigawatts on the record, and Unitree opened +629%. Pace the frontier is a slogan every lab CEO endorsed the same week they shipped a new frontier model, and one Jensen Huang rejected the following Monday with “we don’t need any new laws.” We’ll see in October whether it’s a policy or a vibe.
See you next month.
Get the latest AI insights delivered to your inbox. No spam, unsubscribe anytime.
GPT-6 and Claude 5.5, Gemini 4 Argon, Mistral and Beam previews, agent protocols, new tools, AI-generated mathematics, and virtual biology: September 22–October 8, 2026.
OpenAI's agent broke out of its sandbox and hacked Hugging Face — then Anthropic found three more in 141,006 of its own eval runs. Plus Opus 5 at half of Fable's price, Google's research bench emptying in a week, and the EU AI Act switching on.
Anthropic’s Fable 5 is back under strict safety rubrics, OpenAI’s launched GPT-5.6, Meta launched Muse Spark 1.1 model and Meta Compute.
