❯ Google releases Gemini 4 Argon, which ties GPT-6 Astra in independent tests with under a third of its hallucination rate

Cyber defenders firstGoogle DeepMind released its new frontier model, Gemini 4 Argon, built for complex coding, enterprise knowledge work such as law and finance, and cybersecurity defense. It is rolling out first to governments and trusted cybersecurity teams; Google says it will strengthen its systems using early tester feedback before opening access more widely as soon as possible.

A dead heat with AstraIndependent evaluator Artificial Analysis found Argon (high reasoning) matches GPT-6 Astra (max) on its Intelligence Index, with a hallucination rate of just 15% versus 51% for Astra. The hallucination rate measures how often a model makes things up when it doesn’t know the answer. Argon can also output up to 1 million tokens in a single response, about 8x the cap of Astra and Claude Opus 5.5, suiting very long multi-step tasks done in one go.

Questions beyond benchmarksAccording to Bloomberg, some Google employees say Argon scores well on benchmarks but struggles with some real-world coding tasks; Google disputes that. Limiting it to a few security customers at first also shows how cautious Google is about misuse of highly capable models, echoing Anthropic’s restricted release of Mythos Preview.

A three-way race againOver the past year, the frontier lead has mostly traded between OpenAI and Anthropic. Argon puts Google back in the top tier, giving companies another comparable option for their main model. A low hallucination rate is especially appealing where wrong answers are costly, like law and finance, and it is Google’s opening to take on both rivals’ enterprise business.

▪ SIGNALWhen overall scores tie, saying fewer wrong things may win over high-paying enterprise customers more than getting one more question right.

❯ Micron’s quarterly revenue jumps 379% to $54 billion as AI memory demand extends the chip boom

Another beatAccording to CNBC, Micron reported fiscal fourth-quarter revenue up 379% year over year to $54.23 billion, above analyst estimates of $51.07 billion; net income rose 1,078% to $37.7 billion. Its revenue guidance for the next quarter also beat expectations.

HBM sold out past next yearMicron is one of the world’s three big memory makers, competing with Samsung and SK Hynix. AI servers need large amounts of high-bandwidth memory (HBM), stacked next to GPUs to feed them data at high speed. Micron has said its 2026 HBM capacity is fully sold out under multi-year contracts, with tight supply expected to last beyond 2027. Revenue in the prior quarter was $41.5 billion.

Costs flow downstreamThe shortage is a windfall for Micron, but it raises costs for the cloud providers and AI companies buying GPU servers. Pricier HBM makes each AI server more expensive, which eventually shows up in compute rental and model API prices. As long as data center construction doesn’t slow, memory makers will keep their pricing power.

▪ SIGNALThe first to make big money in AI are often not the model makers but the sellers of bottleneck parts, and memory is the narrowest neck right now.

❯ DeepSeek and Huawei open-source Ascend programming tools, filling a key gap for training on Chinese chips

TileLang comes to AscendAccording to Reuters, DeepSeek says it has partnered with Huawei to develop programming tools for Huawei’s Ascend chips, including the open-source TileLang, seen as an alternative to Nvidia’s CUDA. On September 30 DeepSeek open-sourced the Ascend versions along with supporting compute and communication libraries, six tools in total that previously existed only for Nvidia GPUs.

Chips need software to be usableMuch of Nvidia’s moat comes from CUDA: developers use it to write the low-level programs that run efficiently on GPUs, tying vast amounts of AI code to Nvidia. TileLang is a Python-like language for writing these low-level “operators,” and it was used for most of the operators DeepSeek used to train V4. DeepSeek also said it optimized for 128-card “supernodes.”

Training without Nvidia's permissionUnder US export controls, Chinese AI firms struggle to buy Nvidia’s most advanced chips, and Huawei’s Ascend is the main substitute, but it has been used more for inference because large-scale training tools were immature. Some analysts say DeepSeek’s release isn’t yet enough to fully reproduce a frontier training run but comes surprisingly close. Other Chinese model makers that reuse it could skip a lot of detours.

▪ SIGNALThe hard part of chip independence was never just making the chip but getting developers to write code for it, and DeepSeek is solving that toughest software problem for Huawei.

❯ OpenAI and Synopsys partner on a chip design model under a revenue-sharing deal

AI that runs EDA toolsOpenAI and chip design software giant Synopsys announced a multi-year partnership to build a chip design model, GPT-Synopsys, with revenue sharing between them. According to summaries including TLDR, the model is meant to operate chip design tools like an expert engineer, handling tasks such as timing closure and power optimization.

The most labor-intensive stepSynopsys makes EDA (electronic design automation) software, the essential tools chip companies use to turn circuit designs into manufacturable blueprints. Timing closure ensures signals reach each circuit unit within set times, while power optimization cuts power use without losing performance. These tasks require experienced engineers to tune parameters repeatedly, and a chip often takes more than a year.

Engineers' roles shiftIf a model can take over much of that tuning, chip engineers could focus on architecture choices and design cycles could shorten. OpenAI is also developing its own AI chips, so joining the design-tool business is both a new revenue line and a way to serve its own hardware. The revenue-sharing structure shows both sides are betting such models can be sold directly to chip design customers.

▪ SIGNALAI is starting to help design the chips that run AI, and the faster that loop turns, the harder it gets to predict how fast compute costs will fall.

❯ GPT-6.1 Sol becomes OpenAI’s most demanded model as the CFO confirms 70%-plus Q3 run-rate growth

Swamped on day oneThibault Sottiaux (Tibo), who leads OpenAI’s Codex, said on X that GPT-6.1 Sol, released Tuesday, is pretty much OpenAI’s most demanded model ever, across both the API and subscriptions. ChatGPT and Codex were under heavy load at times; the team has brought more compute online and speeds are recovering.

Cheap means more useGPT-6.1 Sol offers near-Astra capability at one-fifth of Astra’s price, at $2 per million input tokens. Artificial Analysis estimates its cost per task is about 30% lower than GPT-6 Sol, released just a week earlier. Lower prices directly drove usage, shifting many coding and automation tasks that users were reluctant to run on the flagship model onto Sol.

Revenue figures confirmedOpenAI CFO Sarah Friar confirmed on CNBC that the company’s annualized revenue run rate grew more than 70% in the third quarter and business revenue has doubled since July, in line with earlier Axios reporting. Surging demand also brings strain: when compute runs short, users feel it as slowdowns, which is why OpenAI is raising money and building data centers at the same time.

▪ SIGNALCutting prices hasn’t made OpenAI earn less; it has pushed more work onto its servers, making compute supply the real constraint.

❯ OpenAI changes Pro plan usage math, and heavy developers say it halves their allowance

The price of reopeningThe day before DevDay, OpenAI said it would reopen its $200-a-month Pro plan while changing how usage is calculated. Tibo explained that the new formula drops the five-hour cooldown limit and promised to pass gains back to users through model efficiency improvements and API price cuts. But by his own account, usage per dollar works out to roughly half.

Trading the present for the futureDevelopers’ anger centers on one point: OpenAI used future promises to cut the allowance they can actually use now. Several developers rebutted his logic point by point, and some cited the overall value per dollar of Anthropic’s Claude Opus 5.5, asking OpenAI to match it; Tibo publicly agreed with that comparison. Tech commentator kimmonismus complained that the $200 plan has become a $100 plan.

Pushing subscribers toward pay-as-you-goCommercially, OpenAI no longer wants to absorb the high compute costs of the GPT-6 era through cheap flat-rate plans; it wants heavy users on usage-based billing, while leaving margin for new integrated workflow products like Dots. The problem is that heavy developers have little platform loyalty, with rivals like Claude Code right next door; once they feel betrayed, switching costs are low.

▪ SIGNALThe hardest part of subscription AI isn’t setting prices but changing them: once an allowance is given, taking it back burns the trust of your most important users.

❯ Baseten lets enterprises use Kimi K3 inside Codex, billed against their OpenAI commitments

No new vendor neededUS AI inference platform Baseten announced that OpenAI enterprise customers can use open models hosted by Baseten, including Moonshot AI’s Kimi K3, inside the Codex coding tool or via the Responses API. The charges count against the purchase commitments companies have already made to OpenAI, with no separate vendor procurement process.

A Chinese open model on the purchase orderBaseten’s business is running open models for companies and serving their requests; it began hosting Kimi K3 on the day of its July release. K3 is an open model of about 2.8 trillion parameters with leading coding ability among open models. The deal marks the first time a Chinese open model enters the mainstream billing channel of OpenAI’s enterprise customers.

One fewer gate for procurementWhen large companies adopt a new model, compliance review and procurement often take longer than the technical evaluation. Paying with existing OpenAI credits sidesteps that gate. The irony: the same day, according to Bloomberg, OpenAI accused individuals linked to Moonshot AI of taking part in a distillation campaign to extract its models’ reasoning, even as its platform opens a door for Kimi.

▪ SIGNALBuying AI increasingly looks like ordering from one account, and whoever controls the checkout decides which models reach the table, even a rival’s dish.

❯ Meta’s Muse passes 3 million weekly prompting users, giving consumer agents a first stable base

Internal data surfacesAccording to internal data obtained by The Information’s Jyoti Mann, Meta’s AI agent Muse has more than 3 million users who submit at least one prompt a week, and more than 1 million daily active users who send at least one prompt. Separately, Sensor Tower estimates the Muse app passed 5 million downloads in three weeks, faster than ChatGPT, Grok or Claude.

An assistant that runs errandsMuse is the always-on agent Meta launched in September; it connects to users’ apps and handles multi-step everyday tasks. This week Meta launched a version for small businesses and created an enterprise platform unit, hiring former MongoDB CEO CJ Desai to sell Muse to companies. OpenAI’s Dots, launched Tuesday, is widely seen as its direct rival.

Many downloads, fewer stayThe gap between 5 million downloads and 3 million weekly users shows many people tried it and left. For Meta, 1 million daily users is a start but far from its billions-strong social apps. The challenge for consumer agents is that most people don’t have something to hand to AI every day, which is why Meta is rushing Muse into the enterprise.

▪ SIGNALAgents won’t be won on download speed but on whether users think of them every day, a habit no company has built yet.

❯ SpaceXAI plans a unified Grok and X subscription topping out at $100 a month

Four tiers in oneAccording to a document seen by Bloomberg’s Edward Ludlow, SpaceXAI plans to merge memberships for the Grok chatbot and the X social platform into a single subscription with four tiers: a $100-a-month Ultra plan, an $8-a-month Lite plan, a free offering and a middle tier. The plan has not been officially announced.

Two memberships sold separatelyxAI has been folded into SpaceX to form SpaceXAI. Until now, X sold its own Premium membership while Grok sold SuperGrok separately, often leaving users unsure what they got for which. According to earlier disclosures, Grok and X subscription products brought in about $365 million in 2025, with only about 1.9 million users paying for SuperGrok.

Tidying the books before an IPOSpaceX is preparing a listing at a valuation above $2 trillion, and subscription revenue is the easiest part of its AI business to explain to investors. Bundling lets X’s large user base drive Grok sign-ups. The $100 Ultra tier also lines up with OpenAI’s and Anthropic’s premium plans, competing head-on for heavy users.

▪ SIGNALBundling can quickly lift paid numbers, but people pay for social media and for AI for different reasons, and combining them won’t necessarily make both more valuable.

❯ Trump releases a voluntary AI accord with tech companies that calls for external audits

Signed but not bindingAccording to Reuters, Trump released a voluntary AI accord signed with tech leaders, asking companies to partner with external auditors and set up internal controls to monitor AI alignment, meaning whether AI acts as humans intend. Google, OpenAI, Anthropic, Meta, xAI and Nvidia signed it at Tuesday’s White House meeting.

Drafted by ZuckerbergAccording to Semafor, the accord grew out of a conversation between Mark Zuckerberg and House Speaker Mike Johnson; Zuckerberg circulated a draft and Jensen Huang helped build support. Trump also said there would be no new federal AI rules. The signed document misspelled “United States,” drawing widespread mockery.

What voluntary is worthThe accord arrives amid public concern over an OpenAI agent breaching an Australian government website and Anthropic warning of “existential risk” in its prospectus. Voluntary commitments carry no enforcement and no penalties, but they give companies a “we already self-regulate” argument against tougher legislation. Meanwhile, the Federal Trade Commission (FTC) is expanding its probe of OpenAI and Anthropic.

▪ SIGNALRules drafted by the regulated rarely restrain them most; their main effect is deciding how many reasons others still have to restrain them.

❯ The Pentagon taps Musk, Luckey and Gingrich to lead a study of future warfare capabilities

A 120-day studyAccording to CNBC’s Luke Fountain, the US Department of Defense launched Project Meridian, led by Elon Musk, Anduril founder Palmer Luckey and former House Speaker Newt Gingrich, to spend 120 days studying the capabilities the US may need in future wars. The project was launched by Defense Secretary Pete Hegseth.

All close to the administrationMusk’s SpaceX is one of the Pentagon’s largest contractors, providing launches and Starlink satellite communications; Luckey’s Anduril makes drones, autonomous weapons and counter-drone systems and is a major recipient of defense contracts. Gingrich is a veteran Republican politician. TechCrunch noted the picks make sense given their ties to the Trump administration, but the conflicts of interest are obvious.

Who writes the requirements, who gets the ordersStudying what future wars require is essentially setting priorities for future defense procurement. When the researchers are themselves potential suppliers, conclusions may tilt toward drones, satellites and AI, the areas they excel in. That would speed defense contracts to tech companies outside the traditional defense giants.

▪ SIGNALLetting suppliers define the requirements can bring new technology into the military faster, at the cost of procurement no longer starting from neutral.

❯ GPU cloud provider GMI Cloud raises $668 million with Nvidia participating

Equity plus debtAccording to The Information, GPU cloud provider GMI Cloud raised $668 million, including $223 million in equity led by ARCHIV with participation from Nvidia, and a $445 million credit facility led by Taiwan’s CTBC Bank.

Renting compute in the US and AsiaGMI Cloud is a neocloud that rents GPU compute to AI companies and provides model inference services across the US and Asia-Pacific. It raised an $82 million Series A in 2024. Earlier reports said it is building an AI data center in Taoyuan, Taiwan, with Nvidia as its main GPU supplier and committed to leasing any unused capacity for up to six years.

Nvidia backstops its own customersThe neocloud model is to borrow money to buy GPUs and rent them out; the biggest risk is failing to rent the capacity and being unable to repay. By both investing and promising to lease back idle compute, Nvidia lowers the risk for lending banks. Such arrangements help Nvidia sell more chips, but also tie it more tightly to its customers’ financial risks.

▪ SIGNALNvidia increasingly acts like the central bank of this compute build-out, selling chips, investing in buyers and backstopping them, so the chain’s risk ultimately flows back to it.

❯ Apple will unveil a smart home hub with a display on October 13, alongside a new HomePod mini

Apple's next big thingAccording to Bloomberg’s Mark Gurman, Apple will launch new smart home products on October 13, centered on a home hub with a 6-inch screen. It will also release the first HomePod mini update since 2020 and its first new Apple TV since 2022, all built around the new Siri AI.

A shrunken iMac G4The hub is about as thick as an iPhone and has no battery, so it must be plugged in. The standard version recalls the 2002 iMac G4, with the screen on a half-dome base. It can tell household members apart by voice or face, and its software blends Apple Watch, Apple TV and iPad interface elements for controlling appliances, video calls and checking schedules.

A bet on SiriAmazon and Google have sold smart displays for years; Apple is entering now on the strength of the new Siri’s AI. It is also the first launch event since new CEO John Ternus took over and began pushing a faster product cadence. If the new Siri disappoints, the device will struggle to win over homes that already own rivals.

▪ SIGNALApple is years late to home screens, betting that AI makes the category genuinely useful for the first time rather than just a tablet in the kitchen.

❯ Alibaba’s ModelScope and OSChina’s MoArk vie to become China’s Hugging Face

Two platforms by sizeAccording to Rest of World’s Viola Zhou, Alibaba’s ModelScope and OSChina’s MoArk are competing to become China’s alternative to Hugging Face. ModelScope hosts more than 170,000 open models, while MoArk hosts about 20,000.

Why a local platformHugging Face is the main site where developers worldwide download and share open AI models, but access from mainland China is unreliable. Chinese open models have grown rapidly in global influence, with DeepSeek, Qwen and Kimi often publishing to domestic platforms as well. ModelScope, backed by Alibaba Cloud, lets users run and fine-tune models directly on the cloud, while OSChina has long run a domestic open-source code hosting community.

Who owns the download gatewayA model hosting platform looks like just a warehouse, but it decides where developers discover models and which cloud they deploy on. Alibaba uses ModelScope to steer model users to Alibaba Cloud, a classic case of free services driving paid compute. Once that gateway is taken, other domestic cloud providers will find it harder to pull developers onto their own platforms.

▪ SIGNALOpen models are free; the fight is over where developers first download them, because that’s often where they first pay for compute.