OpenAI's first inference chip posts 1.9x Nvidia's throughput per watt, on its own numbers
OpenAI published first benchmark results for its custom inference chip Jalapeño, claiming 1.5-1.9x more throughput per kilowatt than Nvidia’s GB200 and GB300 racks from a 700W part. The numbers were supplied by OpenAI and only partly verified independently, and Nvidia spent the same 48 hours putting Vera Rubin NVL72 into full production with a claimed 30x throughput per megawatt over GB300 on agentic workloads. Three papers landed on the cost rather than the benefit of adding structure to agent systems: injected Agent Skills lowered coding Pass@2 by 1.3-4.2% while raising token cost 72-394%, and two independent groups measured what LLM pairs lose to coordination.
Infrastructure #
Jalapeño’s first results show industry-leading speed and efficiency in AI inference #
OpenAI / SemiAnalysis / TechCrunch
OpenAI published the first public benchmarks for Jalapeño, the inference ASIC it co-developed with Broadcom: 1.5-1.9x more throughput per kilowatt and 1.7-3.6x lower end-to-end latency than Nvidia GB200 and GB300 rack systems on SemiAnalysis’s InferenceX suite, rising to 2.1-4.1x on short interactive workloads of the kind agent execution generates, from a 700W part against accelerators rated at 1,200W and 1,400W. Tests ran on open-weight models — GPT-OSS 120B, DeepSeek R1 670B and Kimi K2.5 1T — with SemiAnalysis reporting over 700 tokens/sec/user on DeepSeek R1 at single-user concurrency and roughly 1,400 on GPT-OSS. The caveats are load-bearing and SemiAnalysis states them plainly: the numbers were provided by OpenAI with in-person verification of only the InferenceX runs, the silicon is A0-stepping engineering samples nine months into the program, OpenAI ran single-token prediction while competitor configurations used MTP, and the workload tested was 8k1k rather than the harder AgentX profile. Hardware chief Richard Ho says volumes are very small through end-2026 with real deployment in 2027, by which time the Blackwell comparison will be a generation stale.
Up to 30x more work per watt: Vera Rubin NVL72 sets a new efficiency standard for AI agents #
NVIDIA
Nvidia says Vera Rubin NVL72 is in full production and delivers 30x higher throughput per megawatt and 35x lower token cost than GB300 NVL72 on agentic workloads, measured on SemiAnalysis’s AgentX benchmark, which replays coding-agent sessions whose accumulated context reaches hundreds of thousands of tokens. The framing is explicitly about agents rather than chat: Nvidia’s own figure is that an agentic session burns 15x more tokens than a conversational one because of multi-step reasoning, tool calls and sub-agent spawning. Read alongside the Jalapeño results, the two announcements are not comparable — OpenAI benchmarked pre-production silicon against Nvidia’s shipping generation, while Nvidia is quoting its next generation against that same baseline on a harder benchmark.
Apple introduces M6 and M5 Ultra #
Apple / Ars Technica
The M5 Ultra is a quad-die UltraFusion part with up to 36 CPU cores, an 80-core GPU with neural accelerators per core, a 32-core Neural Engine, and up to 512GB of unified memory at 1.2TB/s — Apple’s stated pitch is running models with hundreds of billions of parameters entirely on device. The M6 is Apple’s first 2nm chip, with a 12-core CPU, dual 16-core Neural Engine claiming 2x peak compute over the previous generation, and up to 32GB at 170GB/s. The 512GB/1.2TB/s figure is the number that matters for local inference: it puts a desktop within reach of weights that currently require multi-GPU servers, at the cost of bandwidth well below what an HBM accelerator provides, so this buys capacity rather than speed.
Funding & Business #
Hugging Face reportedly in talks to be acquired for $13B #
Business Insider / TechCrunch
Hugging Face has been approached about a sale at a valuation of $13 billion or more and has retained banks to evaluate bids, per Business Insider; no counterparty has been named and no deal has been reached. That price is roughly triple the $4.5 billion post-money it last raised at in 2023, in a round led by Salesforce Ventures with Alphabet, GV and IBM Ventures participating. The strategic question is not the multiple but the asset: the Hub is where open-weight distribution happens for most of the industry, and an owner with a model business of its own would control the default publishing surface for its competitors’ releases. CEO Clément Delangue has repeatedly framed the platform as holding a long-term responsibility to the community that trusts it with models and data, which is the strongest available signal that a sale is not a foregone conclusion.
Stability AI raises $76 million in fresh funding #
TechCrunch
The Series B brings Stability AI to $232 million raised in total, and the investor list is the story: Universal Music Group, Sony Music Group, Warner Music Group and Electronic Arts alongside AMD Ventures and Pacific Alliance Ventures. Three of the four largest music rightsholders funding a generative model company is a different posture from the licensing-and-litigation stance the industry held two years ago, and it follows Stability’s earlier co-development deals with Universal and EA in October 2025 and Warner in November 2025. Stability prevailed in the UK copyright case Getty Images brought against it; the parallel US litigation is unresolved, so the rightsholder détente here is commercial rather than legal.
Research & Papers #
Injecting Agent Skills into coding sessions lowered Pass@2 on 31 of 31 public skills tested #
arXiv
WebDev-Skills-Bench ran 31 public WebDev Agent Skills against 50 Web-Bench projects and 1,000 ordered tasks across four models, with a length-matched irrelevant control to separate skill content from prompt-length effects. Target skill injection reduced mean Pass@2 by 1.3% to 4.2% and increased token cost by 72% to 394%, with gains appearing in only 17% to 36% of skill-project pairs. The control isolates two distinct failure modes: some models are length-distracted, where an equally long irrelevant skill reproduces most of the loss, while others are content-misled, where length is neutral but the skill’s content still costs 1.1-1.4% Pass@2. Losses concentrate on easy early tasks and skill rankings transfer weakly across models, which makes a skill a hypothesis about one skill-project-model triple rather than a portable asset — inject conditionally, and audit per model.
Two LLMs coordinating lose to one LLM working alone, and the loss shrinks with capability #
arXiv
The paper defines a “collaboration tax” as the team-decentralisation loss of a two-player cooperative game with private information, then measures it on 32 tasks each model can already solve alone, across 11 models from 7 providers. The tax orders identically across every model by task category and decreases monotonically with capability, and the mechanism is not reasoning failure but a four-stage conversational cascade: agents assert without grounding, fail to query the partner, skip integrating both views, then accept the answer without re-deriving it. A prompt intervention targeting all four stages closes a substantial fraction of the gap, and in mismatched pairs the result is pulled toward the stronger partner rather than the average — so pairing a weak model with a strong one costs less than the midpoint would predict, but pairing at all still costs something.
When agents read each other’s full outputs, their proposals converge within one round #
arXiv
Testing 11 verifier-scored optimization tasks under matched budgets, this group finds that different model families reach structurally different solutions, but exchanging complete outputs collapses that diversity in a single round — erasing the exact property that motivated using multiple models. They call it the interaction tax, and their conclusion is that full-solution interaction is a weak default: independent proposal generation avoids the collapse, and critique loops help only when the violated rule is one the model can easily locate and fix. Landing the same day as the collaboration-tax paper, it points at the same design error from the other side — what multi-agent systems exchange matters more than how many agents they run.
Prompt rewording moves agent benchmark scores 11-58x more than rerunning the same prompt #
arXiv
An audit of three native tool-calling endpoints across two providers on BFCL’s multiple and parallel categories finds temperature-0 reruns are close to deterministic — ever-flip fractions of 0.7%, 2.0% and 2.7%, with run correlations of 0.997, 0.966 and 0.961 — while semantics-preserving prompt perturbations produce median paired standard deviations 11x to 58x larger. Failure character also diverges under the same headline accuracy: malformed output accounts for 30%, 7% and under 1% of task failures across the three endpoints. The operational consequence is that any tool-calling comparison reported without a perturbation sweep is reporting a number smaller than its own noise floor, and that two endpoints scoring alike can be failing in entirely different ways.
Employment for 22-to-25-year-olds in AI-exposed occupations is 19% below their less-exposed peers #
Stanford Digital Economy Lab / Ars Technica
Brynjolfsson, Chandar and Chen released a revised “Canaries in the Coal Mine?” with an extended ADP payroll panel: since 2022, employment among workers aged 22-25 in the 40% of occupations most exposed to AI has fallen roughly 11%, while the same age group in the least-exposed 60% rose about 10%, leaving a 19% relative gap. Older workers in the same exposed occupations held steady. The authors attribute the decline primarily to reduced hiring rather than layoffs or voluntary exits, which is a materially different mechanism from the displacement story usually told — the effect shows up at the entry gate, not in the termination queue, and exposure is measured partly using Anthropic’s Economic Index rather than an independent task taxonomy.
Developer Tools #
AWS publishes Agentic Resource Discovery, an open specification for finding agents across environments #
AWS
ARD is an Apache-2.0 specification, published at agenticresourcediscovery.org with reference implementations on GitHub, that standardizes how agentic resources are described so local registries can federate without bilateral agreements or proprietary connectors — AWS’s analogy is DNS resolution across networks. The stated problem is that every cloud, on-premises system and SaaS platform currently describes its agents in a different format, so a publisher must re-describe per consumer. What the announcement does not specify is transport, wire format or endpoint structure, which is most of what determines whether this interoperates with MCP’s server discovery or duplicates it; AWS’s own Agent Registry is reachable at a remote MCP endpoint, so the two layers are meant to coexist rather than compete.
Claude Cowork finally remembers what you told the app in chat #
TechCrunch
Anthropic merged the memory stores behind Claude Chat and Claude Cowork, which until now kept separate context and forced users to re-brief the agent when moving from research to execution. Memory is now written as conversations progress rather than at session end, and the stored topics are viewable and editable by the user. It ships on Free, Pro and Max across web, desktop and mobile, with sensitive categories — health, race, ethnicity, religion, gender identity — excluded by default and opt-in only, and government IDs and criminal history never stored. The design decision worth noting is mid-conversation writes: it fixes the handoff but widens the window in which a poisoned or mistaken assertion gets committed to durable memory.
Security #
OpenAI bans a Russian account cluster running a fake think tank #
OpenAI / CNBC / The Register
The operation used VPNs to reach ChatGPT from Russia and worked in Russian to draft social posts for Substack, Telegram, X, Facebook and LinkedIn, explicitly prompting the model to strip stylistic markers that would reveal machine authorship. It fronted as the International Burke Institute, a site registered in February 2025 claiming an Israeli street address and falsely listing Francis Fukuyama and Noam Chomsky among its experts; of 36 articles attributed to those experts between September 2025 and May 2026, 34 were copied from elsewhere, sometimes with incorrect attribution. Its distinctive artifact was a “sovereignty index” scoring countries on a metric constructed to favour Russia. The reusable detail for defenders is the laundering step — the model was asked to remove provenance signals, so stylometric detection was an explicit part of the adversary’s threat model, not an oversight.
Regulatory & Policy #
Taiwan indicts an Nvidia manager and eight others over B300 servers routed to China #
Bloomberg / Ars Technica / Al Jazeera
Taiwanese prosecutors charged nine people, including a senior manager at Nvidia’s Taiwan branch and staff at Supermicro’s New Taipei office, with organizing the export of 74 servers carrying high-end B300 chips to China via Japan and Indonesia; a further 56 servers were seized before shipment. Prosecutors are seeking up to five years for seven of the defendants. This is the Taiwanese leg of the case whose US counterpart was charged in March 2026 under the Export Control Reform Act, tied to roughly $2.5 billion in Supermicro sales since 2024. The escalation that matters is jurisdictional: enforcement is now running through the manufacturing chain’s home jurisdiction rather than only through US export law, which reaches parties a US-only action cannot.
Open Source #
A 4-bit 60B distilled from GPT-OSS 120B beats the bf16 model recovered the usual way #
Hugging Face (Multiverse Computing)
Quantization-Aware Healing distills the compressed, quantized student directly from the original pre-compression teacher instead of from an intermediate recovered checkpoint, using KL-divergence on output logits with chunked computation for long context. Applied to GPT-OSS 120B compressed to 60B parameters at MXFP4, the result beats a bf16-recovered 60B on 7 of 9 benchmarks: +7.4 points on AA-LCR long-context reasoning (42.7 vs 35.3), +5.6 on AIME 2025 (76.3 vs 70.7), +2.7 on Aider agentic coding (40.9 vs 38.2). The claim is that the usual two-step recovery pipeline anchors the student to an already-degraded target and imposes a ceiling that skipping the intermediate removes — worth checking against your own eval set, since the comparison is against the authors’ own baseline rather than a published one, and the post does not say whether the weights are released.
Other #
Anthropic puts $5M behind independent evaluations of AI’s effect on user wellbeing #
Anthropic
The program funds clinicians, psychologists and methodologists to build open-source evaluations and benchmarks measuring how models affect users, with model access and technical support alongside the money; applications close 21 September and decisions land by 5 October. The stated requirements are more specific than the usual grant language: evaluations must state what they measure, involve clinical experts in design and validation, cover both safeguards and harms, run over realistic multi-turn conversations, and validate their graders against subject-matter experts. Two named problems are the useful part — detecting escalating risk across a long conversation, and calibrating between overcompliance and overrefusal — because both are measurement gaps that no lab currently has a public benchmark for.
Threads to Watch #
Vendor-supplied benchmarks are now the primary evidence for frontier hardware claims. OpenAI and Nvidia both chose SemiAnalysis suites this week, which at least fixes the harness, but OpenAI’s figures came from OpenAI on A0 engineering samples with verification of only part of the run set, against a generation Nvidia has already replaced. Nvidia’s own 30x is quoted against its previous product on a harder benchmark than the one OpenAI used. Neither claim is checkable by a buyer today, and the arXiv noise-floor audit published the same day argues that even reproducible harnesses report differences below their own perturbation floor.
Agent research is switching from what structure adds to what structure costs. Three unrelated papers landed on the same shape of result: injecting Agent Skills lowers Pass@2 while multiplying token cost, two LLMs coordinating underperform either alone, and exchanging full outputs collapses the model diversity that justified using several models. All three point the same direction — additional scaffolding is a per-deployment decision requiring a measured control, not a default. Anyone shipping a multi-agent or skills-injection architecture on the assumption that more structure is monotonically better now has three separate measurements arguing otherwise.
Compute is being sold by the watt and the memory envelope, not the FLOP. OpenAI led with throughput per kilowatt, Nvidia with throughput per megawatt, Apple with 512GB at 1.2TB/s. All three framed the pitch around agentic workloads specifically, and Nvidia supplied the reason: an agent session consumes roughly 15x the tokens of a chat turn. When the workload is defined by sustained token volume against long accumulated context, the binding constraint moves from peak arithmetic to power delivery and memory capacity — and the marketing has followed it.