❯ OpenAI rolls out GPT-6 and Intelligent UI to all ChatGPT users, letting answers include charts, buttons and small tools
Answers beyond textOpenAI began rolling out GPT-6 in ChatGPT on October 7, together with a new feature called Intelligent UI. According to MacRumors, ChatGPT’s answers can now include graphics, buttons, forms, charts and diagrams that users click and edit right in the conversation. OpenAI’s Thibault Sottiaux said the update is meant to scale to 1.2 billion users.
Change a number, watch it updateOpenAI’s example is cooking: ask for a recipe and the answer carries a small widget, where changing the number of guests updates the ingredient quantities. Other examples include checklists, embedded maps, multiple-choice follow-ups to narrow a request, and tools such as bill splitters, calculators and small games. Answers also start appearing faster while the rest is processed in the background.
Different models for paid and freePaid Plus, Pro, Business and Enterprise users got it first on October 7, powered by GPT-6 Sol; Free and Go users follow on October 8 with the lower-cost GPT-6 Luna. Technically, OpenAI built a library of native components that can be displayed as they are generated, with a compiler processing the interface while the model writes it.
An interface built per questionChatbots used to return a block of text that users copied into a spreadsheet or calculator themselves. Now the model assembles a throwaway interface for each question. The most directly affected are single-purpose utility apps such as bill splitting, unit conversion and trip checklists, whose function is now covered by one prompt.
▮ SIGNALOnce interfaces can be generated on the spot for each question, the line between an app and an answer blurs, and ChatGPT is competing not just to be where people ask but to be the screen where they finish the task.
❯ Anthropic releases Claude Haiku 5.5, priced 90% below its predecessor for requests up to 100,000 tokens
A big cut for the small modelAnthropic released Claude Haiku 5.5 on October 7, the smallest and fastest tier in its lineup. For requests up to 100,000 tokens it costs $0.10 per million input tokens and $0.50 per million output tokens; above that, $0.50 and $2.50. The previous Haiku 4.5 cost $1 and $5. The company says overall running costs are about 75% lower.
The first Haiku with an effort dialHaiku 5.5 is the first Haiku to offer effort settings, with five levels from low to max, letting developers trade cost against accuracy. Anthropic recommends it for high-volume work such as summaries, classification, database queries, live customer support, browser use and subagents, and says plainly that Sonnet 5.5 and Opus 5.5 remain better for complex agentic coding. The model is live on the Claude Platform and on AWS, Google Cloud and Microsoft Azure.
Cheap, with conditionsDeveloper Simon Willison notes in his write-up that the new price matches OpenAI’s GPT-6 Luna, but only up to 100,000 tokens, beyond which Luna is a much better deal. The new model’s tokenizer also differs, so the same prompt uses roughly 1.25 times as many tokens as before. The same day, Anthropic halved the cache-read price for Sonnet 5.5 and began giving Max and Team subscribers monthly API credits.
Who switchesFor companies running large volumes of short requests every day, such as support triage or document Q&A, the cut changes the bill directly. Among customers Anthropic cites, Asana reports latency down more than 30% and Box an 11-point score gain at half the latency. For long-document workloads whose prompts regularly exceed 100,000 tokens, the savings are much smaller.
▮ SIGNALTwo leading labs have now priced their small models at the same level, and competition at the low end has moved from who is cheaper to where the conditions on cheapness are written.
❯ Musk says Grok Bot will call outside models by task, naming Claude Opus 5.5, Midjourney and Suno
A rival's model inside its own productElon Musk said on X on October 7 that SpaceX will from now on use the best back-end model for any given task, “including Claude Opus 5.5, MidJourney, Suno and other leading APIs,” choosing whatever is most likely to give the best outcome. According to Tesla North, the post drew 35.9 million views.
One cloud computer per botGrok Bot is an AI agent app that SpaceXAI launched in beta this August. As tbreak describes it, each bot runs on its own cloud computer, handles tasks in parallel, works on iPhone, iPad and Mac, and comes bundled with SuperGrok and paid Cursor plans rather than being sold separately. Grok Bot team member Lauren said on X that all bots will be powered by Opus 5.5.
Most requests still go to GrokDigital Music News cites Musk as saying most requests will go to a lightning-fast version of Grok 4.8. SpaceXAI has not said how the two statements fit together, when Midjourney and Suno will be connected, or whether users can pick the model themselves. Some users raised data-privacy concerns about conversations being routed to third parties; Musk did not respond.
Landlord and customer at oncetbreak notes that Anthropic currently rents SpaceX’s Colossus 1 data center, which houses more than 220,000 Nvidia GPUs, to run Claude. SpaceXAI is now buying Claude’s model services in return, so the same company both supplies Anthropic with compute and buys intelligence from it.
▮ SIGNALA company that trains its own models is saying publicly that its product works best on a competitor’s, which shows that in agent products whose model it is matters less than whether the task gets done.
❯ Samsung Electronics estimates third-quarter operating profit at 107.4 trillion won, the first Korean company to top 100 trillion won in a quarter
More than twice last year, in one quarterSamsung Electronics released preliminary third-quarter results on October 8: operating profit of 107.4 trillion won (about $80.28 billion), up 782.5% from a year earlier, on revenue of 195 trillion won, up 126.6%. According to The Korea Times, it is the first time a Korean company has exceeded 100 trillion won in quarterly operating profit.
Profit beat, revenue missedOperating profit came in above the market consensus of 106.64 trillion won, while revenue fell short of the 200.64 trillion won expected. SBS reports that Samsung’s full-year 2025 operating profit was 43.6 trillion won, so this single quarter is nearly 2.5 times that, and the first nine months of this year total 254.1 trillion won. Final results are due later this month.
Almost all of it from memoryPreliminary results do not break out divisions. The Korea Times estimates the memory business contributed about 110 trillion won, driven by full-scale shipments of HBM4 for Nvidia’s Vera Rubin platform and DRAM prices that kept rising at double-digit rates. HBM is the high-bandwidth memory paired with AI accelerator chips. SBS estimates the DX division, home to phones and appliances, lost money for a second straight quarter, and that foundry and system chips lost about 1 trillion won.
Chip buyers pay the billThe other side of rising memory prices is downstream cost. Forecasts cited by SBS put fourth-quarter DRAM prices up another 10% to 15% and NAND up 15% to 20%. For makers that need memory for servers, phones and PCs, that cost shows no sign of easing soon.
▮ SIGNALNearly all of Samsung’s profit comes from memory while phones and appliances lose money, so even inside one company the AI boom has a side that collects and a side that pays.
❯ Musk says his own companies will build and run the Terafab chip plant, with TSMC subleasing a part at most, as Intel says it will stay involved
"Let there be ZERO doubt"Musk responded on X on October 7 to speculation that TSMC would take over Terafab: “No, we will build and run the fab. Let there be ZERO doubt about that.” He added that TSMC might sublease part of the Terafab if it wants, “but nothing more than that.” According to Automotive World, Intel CEO Lip-Bu Tan had told Bloomberg hours earlier that Intel would stay on the project.
Three turns in a weekTerafab is Musk’s planned chip manufacturing project in Grimes County, Texas, with $16.8 billion committed to the first phase, a target of 3,000 wafers a month, first silicon by 2029 and Intel’s 14A process. A report on October 2 said TSMC was exploring a role in operations; on October 3 Musk said it was just talks that might lead to something; on October 7 his tone hardened.
Intel's role has changedIn April, Intel committed to design, fabricate and package chips for Terafab, taking operational responsibility. On Musk’s current account, Intel looks more like the supplier of 14A process technology, with the plant run by Musk’s own companies. Tan confirmed only that Intel will stay and did not say how its role is being adjusted. Intel’s foundry business posted a $10.32 billion operating loss in 2025.
The hard part of running it yourselfTSMC CEO C.C. Wei has cautioned that building a fab takes two to three years and ramping capacity another one to two. Musk’s insistence on running it himself means that time and learning cost falls on his own companies, and how much manufacturing work Intel gets out of it becomes less certain.
▮ SIGNALMusk wants the process, not the foundry service, and for an Intel that badly needs outside customers to prove 14A, that is an order at a discount to what it first looked like.
❯ Broadcom seeks more than $50 billion in financing for OpenAI’s custom chips, with Apollo and Blackstone among those in talks
The chip seller goes borrowing firstThe Wall Street Journal reports that Broadcom is arranging more than $50 billion in private financing for the custom AI chip it is building with OpenAI, with Apollo Global Management and Blackstone among the lenders in talks. Accounts from AI Weekly and Runtime Wire both stress that the financing is in early discussions, with the borrower, terms and security not yet settled.
What the chip isOpenAI and Broadcom announced their partnership in October 2025 to jointly design accelerators for inference, meaning chips built to run models and answer user requests, with a plan to deploy 10 gigawatts by the end of 2029. According to Crypto Briefing, the program is known internally as Nexus and its first two generations are called Jalapeño and Serrano. Runtime Wire says engineering samples are running in labs, with no commercial shipments yet.
A playbook used once beforeCrypto Briefing notes that Broadcom previously arranged a financing package of about $60 billion for Anthropic’s chip needs. The same Journal report also covers Oracle, which is in talks with Apollo and Goldman Sachs for private investors to fund an off-balance-sheet entity that would buy chips and lease them back to Oracle for a 1-gigawatt data center.
Where the risk landsWhen the chip supplier organizes the financing, it is in effect fronting the purchase for its customer, and OpenAI can lock in capacity without spending its own cash. The price is that Broadcom and the private-credit lenders carry the repayment risk: if the chips are slow to reach volume or OpenAI’s revenue falls behind, the reporting so far does not say what repays the debt.
▮ SIGNALFrom Nvidia to Broadcom, chip sellers increasingly double as their customers’ financing arrangers, and it is getting hard to tell how much AI compute demand is being held up by suppliers’ own credit.
❯ Qualcomm and Arm return to court in Delaware, with Qualcomm seeking up to five years free of royalty payments
A five-day jury trialThe contract dispute between Qualcomm and Arm went to trial on October 5 in US federal court in Delaware before Judge Maryellen Noreika, with the jury trial expected to run five days and the judge hearing part of the dispute separately. Reuters reports that Qualcomm wants to stop paying Arm royalties for up to five years, a demand that could be worth billions of dollars.
Two allegationsQualcomm says Arm failed to provide chip-testing tools required by contract and leaked to the press its 2024 threat to terminate a key license, damaging Qualcomm’s chip deal with Meta and cutting its value by $170 million. Arm denies any breach, saying Meta sought to adjust the terms on its own after shifting focus from VR headsets to AI glasses, and that Qualcomm was “not harmed in the least.”
The real fight is the price increaseIn court, Qualcomm’s lawyer kept returning to a chart showing Arm seeking an 1,800% increase in royalties between version 9 and version 10 of its computing architecture. Under the two companies’ 2013 agreement, royalties were capped at $1.88 per chip. Arm licenses its computing architecture to chip designers such as Qualcomm and collects a royalty per chip. The part the judge is hearing separately is whether Arm negotiated in good faith over next-generation technology.
A lawsuit between major partnersAs Lawyer Monthly lays out, Qualcomm filed this suit in April 2024. Arm had sued Qualcomm in 2022 over its acquisition of chip company Nuvia; a December 2024 jury verdict favored Qualcomm and an appeal is pending. Qualcomm accounted for 9% of Arm’s total revenue in the fiscal year ended this March.
▮ SIGNALArm wants royalties that match what chips are worth in the AI era, while one of its largest customers is in court asking to pay nothing for five years, putting the pricing power of architecture licensing to a direct test.
❯ Three safety researchers fired by OpenAI write to the board, urging a halt to models whose reasoning is hard to monitor
A letter a week after the firingsThe Wall Street Journal reports that Jasmine Wang, Tomek Korbak and Mikita Balesni, three researchers fired by OpenAI on October 1, wrote to the company’s board on October 7 urging OpenAI and its peers to stop developing models whose reasoning is hard to monitor and audit, and to work with outside safety auditors. They say the firings are chilling those who remain at the company.
Two accountsOpenAI says an internal investigation found the three violated its policies on accessing and handling sensitive company information. According to Fortune, they are accused of sharing information with an outside AI safety organization that the company has not named; Bloomberg reported that some of it related to the architecture of OpenAI’s infrastructure. OpenAI denies the firings were connected to employees raising safety concerns.
The three say they stayed in boundsAs relayed by Hugging News, the three say in the letter that they never exceeded their professional duties: Korbak was the company’s technical contact for the evaluation group METR, and Balesni helped the board draft an industry safety pledge. Fortune also notes that US Congressman Greg Casar said it looks like the company is firing whistleblowers, while a legal scholar pointed out that California’s existing whistleblower protections typically do not cover disclosures to private third parties.
▮ SIGNALOutside evaluators learn about frontier models largely through people inside the labs who are willing to work with them, and when the boundary of that channel is drawn by dismissals, what suffers is the credibility of third-party evaluation as a whole.
❯ OpenAI previews a College Planner for high school students as Common Sense Media rates ChatGPT for Teens an “unacceptable risk” the same day
Two verdicts on one dayOpenAI shared progress on ChatGPT for Teens on October 7 and previewed a College Planner feature. The same day, the children’s advocacy group Common Sense Media published an assessment rating the product an “unacceptable risk” for all users under 18, its worst rating.
OpenAI's numbersChatGPT for Teens launched this August and applies automatically to accounts identified as belonging to minors. According to Runtime Wire, OpenAI says teens spend under 15 minutes a day on average and that nearly 1.2 million used Learning Visualizations in one week. College Planner helps students track application requirements, deadlines, financial aid and scholarships, starting with US grades 10 to 12; no release date has been announced.
A test of more than 4,000 promptsAccording to FOX45, a Common Sense institute tested more than 4,000 prompts before and after launch. It found that conversations about suicidal thoughts or self-harm failed to trigger parental alerts, that more than a quarter of situations calling for a mental-health crisis referral were missed, and that the limits meant to prevent doing schoolwork were easy to bypass.
The dispute is over methodOpenAI responded that it welcomes independent evaluation but does not believe the testing accurately reflects how the safeguards work in practice, and said the tests may have begun and ended before parental controls were fully activated. The two sides disagree not on whether safeguards should exist but on whether they trigger every time.
▮ SIGNALBuilding a separate version for minors amounts to a public promise that the safeguards will work, and every miss from here on will be measured against that promise.
❯ Claude produces a machine-verified proof of a decades-old conjecture in percolation theory, not yet reviewed independently by mathematicians
The gap from 3 to 10 dimensionsIn late August, Anthropic technical staff member Justin Leder published on GitHub a proof completed autonomously by Claude models: on the square lattice in every dimension from two upward, percolation has no infinite connected cluster at the critical point. Previously this was proved only in two dimensions and in 11 and above, leaving dimensions 3 to 10 open. The result has drawn wide attention after recent media coverage.
What the question asksPercolation theory studies a grid where each edge is open with some probability, and asks how high that probability must be before a network spanning the whole appears. According to 36Kr, the problem was first posed by Broadbent and Hammersley in 1957 and has resisted solution for nearly 70 years. Hugo Duminil-Copin, a 2022 Fields Medalist, tried repeatedly; he wrote in an August 30 blog post that it was probably only a matter of time before the field’s most famous conjecture fell to AI.
Standing on a 2024 paperThe proof follows the approach of Gady Kozma and Shahaf Nitzan from 2024, who reduced the original conjecture to several conjectures about general graphs. Claude proved one of them and then applied their theorem to reach the conclusion. The argument is written in Lean 4 and checked step by step by computer. Gil Kalai wrote on his blog that the conjecture was so general that many expected a counterexample.
Checked by machine, not yet by peopleThe repository lists its own caveats: the proof has not been reviewed by independent mathematicians, all of the code and the 15-page guide were written by AI, and readers need to confirm for themselves that the formal definitions really express the original problem. Mathematician Kalai’s verdict is that “if verified, this will be an extraordinary breakthrough.” Benedikt Jahnel of Technische Universität Braunschweig told the press that a human who proved it would most likely win a Fields Medal.
▮ SIGNALA machine can guarantee that every step of the derivation is sound but not that the statement proved is the one people meant to ask, and checking whether the question was translated correctly is the mathematician’s new job in proofs like this.
❯ Meta, Google DeepMind and Isomorphic Labs put $300 million into Biohub’s Virtual Biology Initiative, bringing total commitments to about $1.8 billion
Building cell data for AIGoogle DeepMind, Alphabet’s drug discovery company Isomorphic Labs and Meta agreed on October 7 to put a combined $300 million into Biohub’s Virtual Biology Initiative. According to Gadget Review, with contributions from the US Department of Energy and the National Institutes of Health, total commitments come to about $1.8 billion, though not all of it is cash.
How the $1.8 billion adds upBiohub, the research organization founded by Mark Zuckerberg and Priscilla Chan in 2016, had already committed $500 million to the initiative. The Department of Energy will contribute more than $500 million over five years; the NIH is providing datasets and repositories valued at more than $500 million based on prior federal investment.
The goal is a predictive "virtual cell"The initiative will measure how cells respond across far more conditions than have been tested before, then use that data to train models that let researchers simulate a candidate drug’s effect on a computer before deciding which experiments to run. The first dataset is expected in about a year, with usable predictive models targeted within five. Biohub head of science Alex Rives said an accurate predictive model of biology could dramatically accelerate discovery by letting scientists perform experiments digitally.
Funders get a one-year head startThe data will not be fully public from the start: commercial funders get roughly a year of priority access to the data they helped pay for before it is released. For a drug discovery company like Isomorphic Labs, that one-year lead is what the money buys.
▮ SIGNALLanguage models had the whole internet to learn from, while biology models lack the data itself, which is why the big money in this field is going first into running experiments rather than buying compute.
❯ Nous Research, maker of the open-source Hermes agent, closes a $90 million Series B at a $1.5 billion valuation
Funding and a new product togetherNous Research confirmed on October 7 that it has closed a $90 million Series B at a post-money valuation of $1.5 billion. According to TechCrunch, Robot Ventures led the round with participation from Nvidia, Union Square Ventures, Menlo Ventures, Samsung and 1789 Capital, bringing total funding to $158 million, and the company is launching Hermes for Businesses at the same time.
A free, open-source agentHermes Agent, developed by Nous Research, is an open-source AI agent under the MIT license, which permits free use. As Runtime Wire describes it, the agent has persistent memory and reusable skills, can search the web, execute code and run scheduled tasks, and connects to messaging platforms. TechCrunch says it has been downloaded more than 24 million times.
Where the revenue comes fromThe software is free, and the company charges for subscriptions and enterprise services: individual plans cost $20, $100 and $200 a month, alongside a free tier. As of mid-September, annualized revenue was about $36 million, and the company projects more than $100 million by year-end. Founded in 2023, the company is led by CEO Dillon Rolnick.
From individuals to enterprisesHermes users have so far been mainly developers and individuals. The business version adds team controls, shared skills, model selection, private or on-premises deployment, single sign-on and service-level agreements, so companies can deploy customized agents for multi-step workflows while keeping data in their own hands. The round is chiefly meant to fund that shift.
▮ SIGNALBeing free won the open-source agent its downloads, but what enterprises pay for is permissions, deployment and accountability, none of which lives in the code, and that is what Nous’s valuation is betting on.
❯ Transsion launches its Hong Kong share sale to raise up to HK$3.36 billion, with proceeds going to AI assistants and agents
The "King of Africa" heads to Hong KongShenzhen-based phone maker Transsion Holdings has launched its Hong Kong listing. According to a Bloomberg report carried by China Daily’s Hong Kong edition, its filing with the Hong Kong exchange shows it plans to offer 86.6 million shares at up to HK$38.80 each, raising as much as HK$3.36 billion (about $428 million), with trading expected to begin on October 15 and Citic Securities as sole sponsor.
More than half of Africa's marketTranssion’s brands include Tecno and Infinix, aimed mainly at emerging markets, and it holds about 53% of Africa’s mobile phone market by unit sales, hence the “King of Africa” nickname. The company is already listed in Shanghai, where its shares are down 19% this year. In this offering, GIC, E Fund Management, Millennium Capital, Golden Link Worldwide and an investment arm of Longsys have committed to take about 37% of the base offering at the top price.
The money goes to AIThe filing says proceeds will be used to accelerate AI-related technology development, including AI assistants and agents, and to strengthen product differentiation. For a phone maker focused on emerging markets, that is a software investment on top of hardware, and its payoff depends on whether users there will pay more for such features.
▮ SIGNALEven a phone maker focused on Africa is earmarking its IPO proceeds for AI agents, so the handset industry largely agrees on what the next selling point is, and the remaining disagreement is over who can afford the R&D.