❯ OpenAI ships 20-plus updates at DevDay, led by always-on Dots agents, as ChatGPT becomes a workplace and app store

Agents that keep workingOpenAI held its annual developer conference, DevDay, in San Francisco on September 29, announcing more than 20 updates. The headliner was Dots: always-on agents powered by GPT-6 Astra. Each dot gets its own cloud computer and browser, connects to 4,000+ apps and keeps working after the chat window closes. Users can assign tasks in ChatGPT, Slack and Teams. OpenAI says Dots are available to Pro (including Pro 100), Business Premium and Enterprise users.

A faster tier, a pricier planOn models, GPT-6.1 Sol offers near-Astra capability at one-fifth the price (see next story). On speed, a new paid tier called Ultrafast generates about 300 tokens per second, up to 8x faster in Codex. Using it in Codex requires the new Pro 500 plan. The Pro 200 plan, previously closed to new sign-ups, reopens, but according to The Decoder usage per dollar is halved, upsetting some paying users.

Codex moves to the cloudDeveloper tools were the second theme. The new Codex Cloud provides configurable cloud environments, so tasks keep running after a programmer closes the laptop, and the Codex CLI now takes voice commands. Codex Security Cloud scans entire GitHub repositories and prepares fixes. On the API side, the Decisions API lets a model choose among predefined answers within a few hundred milliseconds. OpenAI says Responses API token traffic is up 100x from a year ago.

People and dots editing one docFor teams, OpenAI introduced ChatGPT Space, where colleagues and their dots collaborate in one place. Its Pages are live documents both people and AI can edit, with embedded charts, checklists and dashboards. ChatGPT is no longer just a chat box but a workspace for running projects.

Another shot at an app storePlugins were overhauled again: third parties can build native apps with interactive panels and file viewers inside ChatGPT, which OpenAI will recommend in conversations to its 1.2 billion+ weekly users. Plus and Pro subscribers can also sign in to 16 partner products such as Devin and Notion with their ChatGPT allowance, and enterprises can spend OpenAI commitments on software from 32 partners such as Adobe. The launches came a day after OpenAI scrapped GPT-6.1 Astra over safety concerns.

▪ SIGNALOpenAI is no longer selling just a stronger model but a whole environment for AI that works on your behalf; whoever holds users’ tasks, apps and budgets is hard to dislodge with a single better model.

❯ OpenAI releases GPT-6.1 Sol, offering near-Astra capability at one-fifth of the price

Five times cheaper, one point behindOpenAI released GPT-6.1 Sol, saying it nearly matches flagship GPT-6 Astra on agentic coding and professional work at one-fifth of Astra’s price: $2 per million input tokens and $10 per million output tokens. Independent evaluator Artificial Analysis found it scores just 1 point below Astra on its Intelligence Index at less than a quarter of the cost per task.

The predecessor lasted seven daysIn OpenAI’s lineup, Astra is the strongest and most expensive tier, while Sol is the workhorse for everyday, high-volume work. GPT-6 Sol was replaced after just 7 days. OpenAI says the new version is clearly better at writing and debugging code, understanding documents and running multi-step business workflows, and moves closer to Astra in alignment evaluations, being more open about its limits and more respectful of user intent. It is available from today to Plus, Pro, Business, Enterprise and Edu users in ChatGPT Work, Codex and the API.

Head-on with Sonnet 5.5The $2/$10 pricing exactly matches Claude Sonnet 5.5, which Anthropic rolled out this week, putting the two companies in a direct value fight over mid-tier workhorse models. Much of the coding and automation work at enterprise dev teams used to force a trade-off between “expensive and strong” and “cheap but not enough”; now near-flagship capability runs at mid-tier prices, changing both model bills and how teams choose. One enterprise vendor said on X that customers already ask which model to budget for next year, while the models themselves change almost weekly.

▪ SIGNALWhen the gap between flagship and mid-tier shrinks to a single point, the price sheet rather than the leaderboard starts deciding how much work companies hand to AI.

❯ OpenAI aims to raise at least $30 billion at a ~$1.4 trillion pre-money valuation, a private bridge in place of an IPO this year

Up 60% in six monthsAccording to Bloomberg (via Techmeme), OpenAI plans to raise at least $30 billion at a pre-money valuation of about $1.4 trillion. The round is framed as a bridge round, private capital to cover needs ahead of an eventual listing. The raise is still being planned; investors and timing have not been disclosed.

Altman doesn't want Wall Street yetOpenAI was valued at about $852 billion when it raised in March, so the new figure is more than 60% higher. Sam Altman has ruled out a 2026 listing, telling media he did not want the “additional pressure” from Wall Street. Meanwhile rival Anthropic has filed its prospectus targeting a valuation of around $2 trillion, and major OpenAI shareholder SoftBank just completed an $11 billion junk bond sale in which investors asked about OpenAI’s listing timeline and data center plans.

Burn rate sets the paceTraining new models and building data centers demand continuous, massive funding, and by staying private OpenAI must rely on private markets to keep it going. It can keep disclosing less and avoid quarterly scrutiny, but every round needs buyers willing to come in at a trillion-dollar-plus valuation, and such buyers are few and far between.

▪ SIGNALDelaying an IPO doesn’t mean OpenAI needs less money; it just swaps the public market’s verdict for the judgment of a handful of mega-investors for another year.

❯ OpenAI’s annualized revenue nears $70 billion as business revenue doubles since July

Up 70% in a quarterAccording to Axios, citing people familiar, OpenAI’s annualized revenue run rate is nearing $70 billion, up more than 70% since the start of Q3. A run rate extrapolates current revenue over a full year; it is not revenue already booked.

Business is the engineGrowth came mainly from companies: OpenAI’s B2B revenue more than doubled over the period. Consumer growth also accelerated, with new consumer revenue added in Q3 exceeding the amount added in all of 2025. Behind this are the rollout of ChatGPT Enterprise, API usage and the Codex coding tool, as well as OpenAI’s more flexible pricing and allowances for business customers; The Information reported this week that Anthropic has begun ending discounts of about 15% once customers use up their purchased tokens.

Where the valuation comes fromThese figures emerged in the same week OpenAI is planning a $1.4 trillion raise and unveiling a wave of enterprise products at DevDay. Doubling business revenue is the most direct evidence behind a high valuation. Conversely, the harder OpenAI and Anthropic fight for deals, the more leverage corporate buyers gain on price and contract terms.

▪ SIGNALThe second race among model makers is no longer about whose model is smarter but who can turn enterprise budgets into renewable revenue faster.

❯ Anthropic’s prospectus shows nearly half of 2025 sales flowed through Amazon and Google clouds, costing ~$351 million in fees

47% via the cloudAccording to Reuters’ analysis of the prospectus, Anthropic routed 47% of its 2025 sales, about $2.16 billion, through cloud partners Amazon AWS and Google Cloud. Reuters estimates Anthropic paid the two about $351 million in distribution fees.

Shareholders that are also resellersMany large companies already have budgets committed on AWS or Google Cloud, so calling Claude inside those platforms is easier than signing a separate contract with Anthropic. Amazon and Google are also major Anthropic shareholders. According to earlier Reuters reporting on the prospectus, Anthropic’s 2025 revenue grew about 12x to about $4.6 billion, with a net loss of about $42 billion, most of it non-cash accounting charges; about a quarter of revenue came from two customers.

Scale bought, margin sharedUsing the cloud giants’ sales networks lets Anthropic reach large enterprises quickly, but each sale gives up a slice of revenue and depends on platforms that also push their own models. Prospective investors must therefore answer a key question: how much of Anthropic’s growth rests on its own customer relationships, and how much on two partners who are shareholders, resellers and potential rivals all at once.

▪ SIGNALSelling models on a cloud platform is like opening a store in someone else’s mall: the foot traffic comes fast, but the rent and the landlord’s moods come with it.

❯ Anthropic discloses up to $84.5 billion in SpaceX compute payments through 2029, mostly cancellable on 90 days’ notice

Nearly double SpaceX's figureAccording to The Information, Anthropic’s prospectus shows agreements to pay SpaceX up to $84.5 billion through 2029 for its Nvidia-based compute. Most of the agreements can be cancelled with 90 days’ notice.

From $45 billion to $84.5 billionSpaceX’s IPO filing in May disclosed that Anthropic would pay about $1.25 billion a month through May 2029, roughly $45 billion in total; Elon Musk then said on X that the deal was really a 180-day lease, contradicting the filing. Anthropic’s new figure is nearly double, indicating the partnership has grown. Reuters also reported that Anthropic’s infrastructure commitments with six partners exceed $518 billion over 10 years, about 80% of them non-cancellable.

The cancellation clause cuts both waysThe 90-day exit gives Anthropic an escape hatch if demand slows, which matters for a company still losing money. But it also discounts the revenue stability of this big contract for SpaceX, making it harder for its investors to judge what the revenue is really worth. Compute lessors and model companies are forming a new kind of supply relationship: enormous in scale, but adjustable at short notice.

▪ SIGNALAI compute contracts keep getting bigger, but the number that matters is whether they can be cancelled: that decides whether the risk ends up with the model company or the compute provider.

❯ Anthropic says Zhipu’s open GLM-5.3 can autonomously build full cyberattacks but lacks robust safeguards

An open model nears MythosAnthropic published an assessment saying Zhipu’s (Z.ai) openly downloadable GLM-5.3 can, like Claude Mythos Preview, autonomously build end-to-end cyber exploits, from finding a vulnerability to writing working attack code, yet it was released without robust safeguards against misuse, and existing protections are easy to bypass. In the tests, GLM-5.3 achieved full control-flow hijacks in 4% of trials, versus 6% for Mythos Preview.

Mythos is still locked awayA control-flow hijack lets an attacker make a target program run code of their choosing, a key step in breaking into a system. In April, Anthropic limited Mythos Preview to a small group of tech and security companies over concerns about its vulnerability-hunting and exploit skills, refusing a public release. Now a model anyone can download and run offline is approaching that threshold, and open weights, once released, cannot be recalled or restricted after the fact.

Less time for defendersAutomated attack skills once limited to a few top models may soon reach ordinary attackers, further squeezing the window enterprise security teams have to patch holes. The report also has a vantage point: Anthropic is a closed-model company heading into an IPO, and the findings support its argument that stronger capabilities need tighter controls. Still, concrete figures like 4% give regulators and the security industry something tangible to debate.

▪ SIGNALClosed labs can lock up their most dangerous models, but not stop others from building something similar months later; the pace of open-source catch-up now sets how much time cyber defenders get.

❯ Meta launches an enterprise unit and hires MongoDB’s CEO to sell its Muse agent to businesses

A CEO jumps, a stock dropsMeta has created a new business unit, Meta Enterprise Platform, led by CJ Desai, who had been MongoDB CEO for less than a year, as chief enterprise platform officer reporting directly to Mark Zuckerberg. According to CNBC, MongoDB shares fell more than 18% on the news and Meta shares fell about 5%; former CEO Dev Ittycheria is MongoDB’s interim chief.

Muse goes firstThe new unit will first push the Muse agent, a business customer-service agent and a coding tool, packaged for companies and developers. Muse has taken off with consumers, but ordinary people use it infrequently and are hard to convert to paying. This week Meta also launched Muse for Small Business, connecting it to 15 business apps such as Slack, QuickBooks, Shopify and Stripe for multi-step tasks. Desai was previously president and COO of ServiceNow and president of product and engineering at Cloudflare, with long experience selling and delivering enterprise software.

Meta needs a new businessMeta’s capital spending on AI has been enormous, and beyond advertising it needs large, high-ticket enterprise revenue renewed annually to spread those costs. That is exactly the market OpenAI and Anthropic are already fighting over; Meta’s edge is its relationships with vast numbers of small merchants through ads and shops. The hiring focus is shifting too: two years ago labs fought over top researchers, now they pay up for leaders who know enterprise sales and operations.

▪ SIGNALBy poaching a public-company CEO, Zuckerberg is buying the capability Meta lacks most: turning AI into contracts businesses renew every year.

❯ Moonshot’s Kimi K3.1 surfaces on its API platform, reportedly with a 1M-token context and a launch next month

A new name in the registrySeveral developers found a callable kimi-k3-1 identifier in the model registry of Moonshot AI’s API platform, and a preview of K3.1 models then appeared on Kimi’s official developer platform. According to details shared by developers, K3.1 supports a context window of up to 1 million tokens, offers three reasoning-effort levels, and may add agent mode, multi-agent collaboration, and search and batch task modes. Moonshot has not officially launched it; it is expected next month.

The first upgrade after K3Moonshot open-sourced Kimi K3 in July, a sparse mixture-of-experts model of about 2.8 trillion parameters, also with a million-token context, which many in the industry called another “DeepSeek moment.” The context window determines how much material a model can read at once; a million tokens can hold an entire codebase or several long books. Reasoning-effort levels let users choose how long the model “thinks” by task difficulty: more compute for hard problems, faster and cheaper answers for easy ones.

Pricing may not be finalThe K3.1 API price currently shown matches K3, though it may be a placeholder. If the effort levels map to different compute configurations and billing, the cost of calling Kimi will depend more on which level developers choose than on the model name. For developers, it also means another long-context model with open-source roots joining the fight, pushing prices down further for long-document and coding work.

▪ SIGNALChinese model makers are moving from chasing a benchmark score to selling one model at several price points, making pricing design as important as capability.

❯ DeepSeek Harness v0.2 preview ships desktop installers, taking the agent tool off the command line

Install and goDeepSeek released a preview of DeepSeek Harness v0.2 on September 29, offering ready-to-use macOS and Windows desktop installers for the first time, downloadable from its website. The new version comes preloaded with common office and development features, adds a plugin install and management page, and opens up more configuration options so users can extend it as needed.

From terminal to desktopHarness is DeepSeek’s open-source agent runtime, built on the idea that “everything is a plugin”: tool calls, file access and task execution are all added as plugins. Previously users had to run a terminal command and then open a local web interface in a browser, prompting the community to build several third-party desktop wrappers. The new desktop app is built on Electron and reuses the existing web UI and the agent, session, tool and plugin modules. It also bundles an optional plugin for scheduled tasks.

A lower bar for office usersOffice users who don’t code can now install it with a double-click, far friendlier than configuring a command line. That gives Harness a chance to move beyond developers into everyday work and compete with Claude Code and Codex for the desktop. Whether a plugin ecosystem takes off will depend on how many third parties build plugins for it and how DeepSeek manages plugin security.

▪ SIGNALCompetition among agent tools is shifting from whose model codes better to whose installer is easier and whose plugins are easier to find; the desktop is the new battleground.

❯ The Trump administration launches America.gov, an AI chatbot front door to federal services

One site for the federal governmentAccording to The Hill, the Trump administration officially launched America.gov on September 29 as a single, AI-powered access point for federal services and information, where people can ask about procedures and policies in a chat. According to The Rundown, it runs on Google’s Gemini and xAI’s Grok.

Led by Airbnb's co-founderThe project is led by Airbnb co-founder Joe Gebbia, now the first US Chief Design Officer and head of the National Design Studio. The launch event in Washington featured Trump and Vice President JD Vance, with Elon Musk, Jensen Huang and other tech executives in attendance. The same day, the White House and Google, OpenAI, Anthropic, Meta, xAI and Nvidia signed a voluntary AI safety accord; Trump also signed an executive order requiring federal agencies to call AI “Super Intelligence” (SI) in official communications and said there would be no new federal AI rules.

Hallucinations cost more on a government siteUsing a chatbot to simplify dealings with government can spare people hopping among dozens of agency websites. But large models still confidently make things up, and a wrong answer about benefit eligibility or tax rules carries far more serious consequences than an ordinary chat. Outlets including TechCrunch are already testing its accuracy; whether this front door is trusted will depend on how the government labels sources and corrects errors.

▪ SIGNALPutting AI on the front line of public services tests “who is responsible when it’s wrong” more than any industry, and the answer will become the template for other agencies.

❯ General Intuition raises $220 million at a $6.2 billion valuation to train spatial-reasoning agents on game footage

Nearly tripled in two monthsAccording to GamesBeat, General Intuition raised $220 million at a $6.2 billion valuation, bringing total funding to more than $650 million. Earlier reports said new investors including Valor Equity Partners and Point72 Ventures joined the round, nearly tripling the $2.3 billion valuation set about two months ago.

Game clips as textbooksThe New York company spun out of gameplay-clip platform Medal in October 2025. Players share huge volumes of game clips on Medal, along with records of their inputs, and General Intuition trains “world models” on hundreds of millions of hours of this footage so AI learns to understand 3D space and predict how objects and characters move. It aims to apply that spatial reasoning to agents that act in virtual or real environments, with robots as the next target.

A shortcut for the data-starvedWhat robot training lacks most is large-scale, real “see this, do that” data, which is expensive to collect. Game footage offers a cheap substitute, which is why investors are raising the price so quickly. The hard part is that game physics differ from the real world, and how much capability survives the move from screen to robot arms and wheels still needs to be proven in deployment.

▪ SIGNALPhysical AI is first a contest over data sources: whoever finds cheap, massive “action data” can double its valuation faster than rivals.

❯ Dataflow chipmaker Efficient Computer raises a $97 million Series B at a $650 million valuation

Second round this yearAccording to Reuters reporter Stephen Nellis, chip startup Efficient Computer raised a $97 million Series B at a $650 million valuation. The company closed a $60 million Series A just in January, led by Triatomic Capital with participation from Union Square Ventures, Toyota Ventures and RTX Ventures.

Data flows to the computeFounded in 2022 by computer architecture researchers from Carnegie Mellon University, Efficient Computer builds processors based on a spatial dataflow architecture. A conventional CPU fetches data, computes and writes it back one instruction at a time, spending much of its power moving data around; a dataflow design lays compute units out in a grid and lets data flow through the right units as it arrives. Its first chip, Electron E1, is programmable in C, and the company says it is 10 to 100 times more energy-efficient than commercial ultra-low-power CPUs on typical embedded tasks.

AI heads to the edgeDrones, sensors, wearables and industrial equipment all want to run AI locally, but their batteries and cooling can’t support power-hungry chips. That makes efficient processors a key component as AI spreads to devices, and explains why defense and automotive investors appear on its cap table. The hard part for any new architecture is the software ecosystem: whether developers will rewrite code for it matters more than chip specs.

▪ SIGNALThe AI chip battle isn’t only in data centers; on devices where one battery must last for years, compute per watt is the real barrier.

❯ Data center power supplier Accelevation prices its US IPO below range, raising $540 million

Marketed at $20–$24, priced at $18According to Bloomberg, AI infrastructure company Accelevation and its backer Olympus Partners sold 30 million shares at $18 each in a US IPO, raising $540 million, below the marketed range of $20 to $24. The company plans to list on Nasdaq under the ticker ACCV.

Selling power gear for data centersThe Ohio-based company designs, manufactures and installs power distribution and other critical infrastructure for AI and cloud data centers, routing grid power safely to rows of servers. Private equity firm Olympus has owned it since early 2025 and will retain majority voting power after the offering. Its filing shows revenue grew 147% in 2025, with a backlog of about $1.1 billion as of the end of June 2026.

Investors get pickyThe data center building boom has sent revenue soaring at “picks and shovels” suppliers, but investors are growing more cautious on valuations. The same day, smart ring maker Oura postponed its IPO, and Bain published a report saying AI needs $6 trillion in annual revenue by 2031 to justify current data center spending. For AI infrastructure companies still in the IPO queue, high growth no longer automatically buys a high price.

▪ SIGNALAI infrastructure orders are still climbing, but public markets have started discounting them as investors sort durable growth from the peak of this build-out cycle.

❯ Oura postpones its IPO at the last minute as some investors balk at a ~$15 billion valuation

Called off the day before pricingSmart ring maker Oura postponed its Nasdaq IPO one day before its planned September 29 pricing, citing market uncertainty, though the company said demand was strong and its business had strengthened during the process. According to Bloomberg, some potential investors held off over a target valuation of about $15 billion and the poor post-listing performance of peers such as Fitbit.

Rings plus subscriptionsFinland’s Oura sells smart rings that track sleep, heart rate and temperature, and charges users a subscription for health-data analysis. It had planned to sell 50 million shares at $40 to $44 each to raise up to $2.2 billion; Eli Lilly and Dragoneer had indicated interest in buying up to $100 million and $300 million, respectively.

Hardware struggles to earn software multiplesInvestors doubt whether Oura can keep selling hardware under pressure from Apple, Samsung and Garmin while its subscriptions justify a near-software valuation. The delay disrupts the funding plans of the company and its existing shareholders and makes other IPO candidates more cautious; a guest on The Information’s show even discussed whether market uncertainty could affect Anthropic’s listing plans.

▪ SIGNALMarkets will pay up for an AI story, but for a ring that must win back its users’ subscriptions every year, they do the hardware math first.

❯ Apple’s new CEO John Ternus begins a shake-up, cutting management layers and speeding up launches

Moving within weeksAccording to Bloomberg’s Mark Gurman, Apple’s new CEO John Ternus is overhauling the company to run faster and leaner with a greater focus on engineering, and to launch products more often, possibly year-round instead of in concentrated spring and fall events. Apple has laid off a batch of engineering program managers in its hardware division, including about six directors, and employees expect more cuts.

An engineer in chargeTernus long ran Apple’s hardware engineering and has been CEO for only a few weeks. The core of the overhaul is to cut layers between engineers and top executives so product development takes fewer detours. Gurman also reported that Apple had planned in recent months to lay off about 5,000 AppleCare customer service staff and replace them with AI agents, but the plan has been put on hold.

Speeding up after falling behind in AIApple is widely seen as trailing Google, OpenAI and Meta in generative AI, and Siri’s upgrade has been repeatedly delayed. A faster launch cadence and fewer layers aim to shorten the path from R&D to market. Middle managers are first in line, and denser launches will also force the supply chain to adjust its stocking rhythm.

▪ SIGNALApple’s problem has never been a lack of money or talent but slowness; by cutting layers rather than products first, the new CEO signals he thinks the drag lies in the organization itself.

❯ Shein’s net income falls 66% in its first report since its Hong Kong IPO, with shares ~30% below the offer price

Revenue barely movesAccording to the Wall Street Journal, fast-fashion e-commerce company Shein reported Q2 revenue up 0.9% year over year to $11 billion, while net income fell 66% to $228 million. It is Shein’s first report since its Hong Kong IPO, and its shares trade about 30% below the offer price.

From rapid growth to standstillShein sells clothing and small goods worldwide at rock-bottom prices with rapid new releases, mainly to customers in the US and Europe. It expanded quickly on cheap parcels for years, but after the US ended duty-free treatment for small parcels and Europe tightened rules on fast-fashion platforms, logistics and compliance costs rose, while price wars with platforms like Temu squeezed margins. After years of trying to list in New York and London, it ultimately chose Hong Kong.

The first report card after listingStagnant revenue plus a profit plunge shows the low-price model struggles to hold past margins under tariff and regulatory pressure. Shein now has to prove to its Hong Kong investors that it can grow by raising basket sizes, expanding its third-party seller marketplace or entering new markets, not just by cutting prices further.

▪ SIGNALCheap cross-border e-commerce grew in the gaps between rules; once those gaps close, the earnings report tells the truth before the share price does.