❯ Google unveils Gemini 4 Carbon, starting with selected cyber defenders
A phased releaseGoogle announced Gemini 4 Carbon on October 28 and is initially giving selected cybersecurity defenders access this week. Its announcement says access will expand through trusted partners, while availability for developers, businesses and consumers is still being prepared.
Engineering work that lasts for daysCarbon scores 59 on Artificial Analysis Intelligence Index v5.0. Google highlights improvements in coding, tool use and sustained task execution, saying internal teams have used it for days to diagnose and patch problems in large codebases. Google also highlights evaluations of vulnerability discovery, validation and remediation, with cybersecurity a central use case for the initial release.
Introductory prices halvedThe introductory price is $2 per million input tokens and $10 per million output tokens. A 95% cached-input discount puts cached input at $0.10 per million tokens; regular input and output prices will be $4 and $20 after the introductory period. Google also reports zero successful attacks in the evaluated Gray Swan indirect prompt-injection benchmark.
▮ SIGNALLong-running tasks call models repeatedly, making cache discounts and sustained execution jointly important to the cost of engineering work.
❯ OpenAI raises Pro and Business Premium limits and cancels the Pro 200 cut
Higher paid allowancesOpenAI product leader Thibault Sottiaux said in an October 29 update, U.S. time, that usage limits for Pro 100, Pro 200, Pro 500 and Business Premium were increasing immediately. In day 25 of the company’s 28-day improvement series, he also confirmed that the planned Pro 200 reduction was canceled, preserving the higher existing allowance for ChatGPT Work and Codex.
Preserving older allowances and restoring newer onesNew Pro 200 subscribers after September 29 had a lower allowance, while existing subscribers were in a transition period. The change doubles the lower allowance to match the previous higher level. Subscribers already on that higher level retain their existing allowance. Pro 100 and Pro 500 receive higher Work and Codex allowances than before.
Task choices still affect usageOfficial OpenAI documentation confirms that Work and Codex share usage and that Pro currently has no five-hour limit, although other usage constraints remain. Context size, model choice and speed affect consumption, and Ultrafast has separate eligibility requirements. For people using these tools throughout a coding or office workday, the immediate benefit is more work within the same subscription and the removal of the expected Pro 200 cut.
▮ SIGNALHigher allowances at the same subscription price give frequent coding and office users more capacity, strengthening the appeal of premium plans.
❯ Anthropic renames its top subscription Ultra and removes verified accounts’ five-hour limit
Two renamed tiersAnthropic announced on October 30 that the $100-a-month Max 5x plan is now Max, while the $200-a-month Max 20x plan is now Ultra, with prices unchanged. Its official announcement removes the five-hour usage limit for verified Ultra accounts, beginning with most existing subscribers. Newer accounts retain a temporary limit until verification.
Weekly allowances remainThe Claude pricing page lists Max at five times Pro usage both per five-hour window and per week. Ultra offers ten times Pro’s weekly usage and no five-hour limit for verified accounts. Max and Ultra also include $100 and $200 in monthly Claude Platform API credits, respectively, separate from regular Claude usage.
Fewer interruptions during sustained workA five-hour window constrains how quickly subscribers can spend their allowance. Removing it lets eligible Ultra subscribers concentrate their weekly budget into longer coding, research or office sessions. Max supplies a larger allowance; Ultra additionally relaxes the pace at which that allowance can be used. The weekly budget continues to constrain total usage.
▮ SIGNALPremium subscriptions increasingly sell uninterrupted working time; spending cadence matters alongside the size of the weekly allowance.
❯ Claude users report a possible model test, without confirmation of Fable 5.5
Users share changed experiencesAn October 30 report carried by Phoenix collects tests from Claude users who noticed changes in animation, drawing and longer tasks after selecting Opus 5.5. They suspect their requests were routed to a newer model. The report links these experiences to Fable 5.5 or a future Opus, but Anthropic has not officially confirmed the model’s identity.
From fixing a bug to understanding the goalThe report also cites The Information’s interviews with testers, who said Anthropic was preparing a model better at understanding user intent. A request to fix a bug, for example, could prompt it to identify the cause of similar errors and ask whether to address those too. Testers valued its ability to identify related problems and make a fix more comprehensive.
Animation and drawing provide visible examplesThe report also shows outputs involving Three.js animation, SVG drawings and longer tasks. Animation and vector drawing require a model to translate written instructions into structure, positioning and interaction details, making differences in the output easier to see. An official Fable 5.5 version, API, price and release schedule have not been announced.
▮ SIGNALThe improvement these users value most is identifying problems related to the task, which can reduce the extra instructions needed for complex work.
❯ Anthropic uses post-response Claude self-assessment to screen for misuse efficiently
A check after the answerAnthropic drew renewed attention this week to self-assessment, a safety technique described in its October 16 research post. After answering, Claude produces a brief assessment invisible to the user, screening the conversation for potential misuse and feeding offline analysis and further review. The method covers a large share of production traffic on models released from Fable 5 onward.
Reusing existing contextA separate reviewing model would normally need to reread the conversation. Self-assessment reuses the responding Claude model and its KV cache, the intermediate computation already available for that context. It adds a short instruction and roughly ten output tokens on average. Anthropic estimates less than 2% additional inference compute across its production API traffic distribution, without charging those tokens to users.
An initial screening layerIn Fable 5 tests, compared with a stronger prompted classifier, self-assessment flagged 78% of potentially harmful conversations and 94% of high-priority conversations at a 1% false-positive rate. Flagged conversations feed offline analysis and further review to identify aggregate risks, cases missed by other detectors and new misuse patterns. Other safeguards continue to handle real-time blocking.
▮ SIGNALReusing computation from the original task can make sophisticated safety screening affordable across a much larger volume of traffic.
❯ Apple sets November 4 beta testing for Apple Intelligence in China ahead of a December release
A confirmed beta dateApple says selected Apple Intelligence features will begin testing in China on November 4, Beijing time, with the iOS 27.3 developer beta. The public release is planned for December, without a specific date. MacRumors reports the schedule, citing Apple’s Chinese newsroom announcement.
A limited initial feature setThe initial features target supported devices from iPhone 15 Pro onward and include Writing Tools, Genmoji, Clean Up, notification summaries, Visual Intelligence and Live Translation with AirPods. Apple is also partnering with Alibaba to integrate Qwen into Siri, Writing Tools and Visual Intelligence. The new Siri AI and Image Playground are outside this first release.
Distribution through the operating systemThe features sit inside everyday text editing, photos, notifications and the system assistant, reducing the need to open a separate chatbot for each task. Operating-system integration also creates a distribution channel for the local model partner. Users still need the relevant software release, compatible hardware and supported language settings before they can use it.
▮ SIGNALAI distribution now extends into operating-system workflows, where the everyday actions covered can matter more than the number of models offered.
❯ OpenAI lowers its fundraising target to about $20 billion at a proposed $1.25 trillion valuation
A smaller targetOpenAI is seeking approximately $20 billion at a $1.25 trillion valuation, down from a previous plan for $30 billion at $1.4 trillion. Reuters, carried by AOL, attributes the figures to the Financial Times. Some investors have reduced commitments, and the first close is expected to fall below $20 billion. The round remains in progress.
Competing for overlapping investorsThe ChatGPT developer had been trying to secure commitments before Anthropic’s expected public-market debut in mid-November, with both frontier-model companies drawing on an overlapping investor base. The adjustment also comes during a U.S. technology selloff and renewed scrutiny of large AI spending commitments. OpenAI did not immediately respond to Reuters’ request for comment.
A tougher market for fundraisingReuters reports that the Nasdaq fell 2.9% on Friday and stood 7.8% below its October 6 closing high. OpenAI had set Friday as the deadline for investor commitments, but some investors reduced their planned allocations. With model training, compute purchases and service operations requiring continuing investment, changes to the round’s size and price add uncertainty to its fundraising plans.
▮ SIGNALReduced investor commitments are pushing OpenAI to adjust both the size and price of its round, showing how capital supply can constrain frontier-model expansion.
❯ Baseten seeks about $2 billion in two tranches valued at $26 billion and $30 billion
Two prices within one financing effortAI infrastructure provider Baseten is arranging roughly $2 billion in new funding, with two tranches valuing it at $26 billion and $30 billion. Bloomberg’s report, together with FourWeekMBA’s account, describes a financing in progress.
Infrastructure for running modelsBaseten supplies deployment and inference services: the systems that keep trained models running and answering real users. Its company funding announcement lists customers including Abridge, Cursor and Notion. Customers pay for response speed, reliability and deployment operations instead of building the entire serving stack themselves.
A rapid valuation increaseBaseten raised $1.5 billion at a $13 billion valuation in June. The newly reported prices are approximately two and 2.3 times that valuation. Investors continue to back demand for running models in production, where reasoning models and agents can consume more compute through multi-step tasks. Delivering that capacity reliably at an acceptable cost underpins the business.
▮ SIGNALAs model applications expand, low latency, reliable operation and deployment efficiency become more valuable, sustaining investor interest in inference infrastructure.
❯ DeepSeek’s second round reportedly reaches RMB105.5 billion with nearly 30 investors
More detail on the second roundOn October 30, IT Home’s account of The Information put DeepSeek’s second financing round at approximately RMB105.5 billion, with a valuation around RMB615.5 billion and nearly 30 investors. Earlier reporting had described the round as completed. The latest report provides a more precise round size and a fuller investor list.
A broader industrial investor baseThe reported investors include Tencent, CATL and its chairman Robin Zeng, CXMT, SMIC, Meituan, vivo and Xiaohongshu. Tencent contributed RMB10 billion, while CATL and Zeng together contributed RMB10 billion. State-backed investors supplied about 27% of the capital in this round, up from 2% in the first.
An expanded fundraising targetDeepSeek supplies foundation models and related services, requiring sustained computing resources for development and usage. According to the report, its first round in June raised roughly RMB51 billion. The second-round target increased from approximately RMB50 billion to RMB100 billion before the more precise RMB105.5 billion figure emerged. Investments from chip, consumer-electronics and internet companies connect the model developer with a wider industrial capital base.
▮ SIGNALInvestment from chipmakers, device manufacturers and internet companies broadens the funding base for foundation-model development.
❯ 01.AI secures hundreds of millions of yuan from Hangzhou investors and plans a Linping headquarters
Investment alongside a headquarters move01.AI announced a partnership with Hangzhou and its Linping district, involving hundreds of millions of yuan from city- and district-level state-backed investors. The company plans to locate its headquarters in Linping and establish an industrial AI transformation base. National Business Daily reported the arrangement on October 30; Zhidx’s account dates the agreement to October 28. An exact investment amount was not disclosed.
Products aimed at business operationsFounded by Kai-Fu Lee in 2023, 01.AI has shifted its emphasis from foundation-model competition toward enterprise applications. Its Wance AI platform connects company data, knowledge, workflows and organizational structures, combining multiple agents with forward-deployed engineers. The offering targets management decisions, sales, production and supply chains, where customers want AI to improve actual operating outcomes.
Preparing for commercialization and an IPOThe company is pursuing a pre-IPO strategy and previously set goals for a 2027 IPO and RMB1.5 billion in orders in 2026. The Hangzhou arrangement brings together capital, a headquarters and industrial use cases. Expanding the business requires turning experience from customized major-customer projects into products that can be sold and delivered repeatedly.
▮ SIGNALEnterprise AI expansion needs both capital and repeatable delivery; connecting a headquarters with industrial customers addresses the latter.
❯ Volcano Engine prioritizes MaaS commercialization after RMB4.8 billion in first-half revenue
Revenue takes priority over token volumeLatePost reports that Volcano Engine’s model-as-a-service business generated RMB4.8 billion in first-half revenue at a gross margin above 55%, making it ByteDance’s largest AI business by revenue. Since midyear, commercialization has become a higher-priority metric, and the company has raised its original RMB15 billion annual revenue target. The revised figure was not disclosed.
Customers pay for actual useModel-as-a-service gives enterprises access to models through a platform, charging by tokens, images or audio and video duration. Volcano Engine supplies ByteDance’s Seed and Seedance models alongside third-party models such as DeepSeek and Zhipu. Equal token volumes therefore need not produce equal revenue: the mix of text, video and complex agent tasks changes both pricing and serving costs.
A market with measurable revenueIT Home, citing IDC, puts China’s first-half 2026 MaaS market at RMB10.63 billion, with Volcano Engine holding a 45.1% revenue share. The category covers language, image, audio and video model services. As models enter coding, office work and content production, the commercial focus increasingly includes repeat customer spending and gross profit on inference, alongside attracting usage.
▮ SIGNALAs customers buy video and complex-task services alongside text, product mix and inference costs jointly shape platform revenue and gross profit.
❯ Intel falls 28.58% as server demand fails to offset a weak fourth-quarter outlook
The post-earnings selloffIntel closed down 28.58% on October 30, U.S. time. CNBC describes its worst one-day performance since the 1970s and an approximately $12 billion reduction in the value of the U.S. government’s stake. The selloff extended from the previous evening’s after-hours trading into Friday’s regular session.
Growth now, pressure aheadIntel’s official earnings release reports third-quarter revenue of $16.5 billion, up 21% year over year. It forecasts fourth-quarter revenue of $15 billion to $16 billion and adjusted earnings of $0.25 per share. Data-center revenue recorded its strongest growth in 15 years, but the company expects the PC slowdown to last at least through the first half of 2027.
Supply and business mix shape the outcomeIntel needs to increase Xeon supply to meet server-CPU demand associated with AI infrastructure. Its substantial PC business remains exposed to a weak market, while manufacturing expansion requires capital. Growth in server orders therefore coincides with pressure elsewhere, and fourth-quarter revenue guidance remains below the third-quarter result.
▮ SIGNALAI infrastructure can lift server-CPU demand, while supply, product mix and manufacturing costs still determine a diversified chipmaker’s near-term results.
❯ CXMT’s third-quarter revenue rises more than sixfold, with RMB54.435 billion in attributable profit
A return to quarterly profitCXMT reported third-quarter revenue of RMB118.714 billion, up 613.18% year over year. Profit attributable to listed-company shareholders reached RMB54.435 billion, reversing a year-earlier loss. The quarterly filing carried by CFi also reports total net profit of about RMB88.983 billion. That figure includes profit attributable to non-controlling interests.
Volume, prices and unit costsCXMT makes DRAM, the working memory used in servers, PCs and mobile devices. The filing attributes growth to rising global compute demand and capacity allocation by major suppliers, which tightened DRAM supply. CXMT’s capacity expansion and product improvements helped lift sales volume and prices, while scale and manufacturing management reduced unit costs, jointly supporting profitability.
Revenue grows faster than profit sequentiallyQuarterly revenue increased approximately 19.30% sequentially, while attributable profit grew about 3.01%. IT Home’s account of the filing puts nine-month attributable profit at RMB132.040 billion. Alongside its large year-on-year gains, CXMT continued to grow both revenue and attributable profit from the preceding quarter.
▮ SIGNALTight DRAM supply and added capacity are improving CXMT’s volumes, prices and unit costs, turning compute demand into memory-sector earnings.
❯ Reddit falls 29.27% as AI search traffic losses weigh on its fourth-quarter outlook
Strong results do not prevent a selloffReddit closed down 29.27% on October 30, U.S. time, its largest daily decline since its 2024 IPO. CNBC points to a weak fourth-quarter revenue outlook. Its earnings release reports third-quarter revenue of $895 million, up 53%, including $847 million from advertising. Fourth-quarter revenue is forecast at $1 billion to $1.03 billion.
Answers without a site visitDecrypt’s account of the earnings call says CEO Steve Huffman reported an average weekly decline of more than one million Google search referrals during October. He warned that daily users could decline by a low-double-digit percentage in the fourth quarter if the trend persists. AI search can summarize discussions directly, reducing the need to click through and affecting both discovery and return visits.
Licensing is separate from the advertising funnelReddit has data partnerships with Google and OpenAI, but licensing content is a different business from monetizing visits through advertising, which still supplied most quarterly revenue. The company plans fourth-quarter tests of a more logged-in experience in selected markets outside the U.S. By encouraging more people to log in and use Reddit directly, the company hopes to reduce its reliance on external search referrals.
▮ SIGNALAdvertising remains Reddit’s main revenue source, making direct visits especially important as AI search reduces clicks through to the site.
❯ Musk says Neuralink’s Blindsight is engineering-ready and awaiting approval
An engineering updateElon Musk said on October 30 that Neuralink’s second product, Blindsight, is ready from an engineering standpoint and awaiting regulatory approval. The product has not yet received approval for commercial sale, according to Drive Tesla.
A different goal from thought-based controlNeuralink’s first product, Telepathy, primarily uses brain signals to control external devices. Blindsight aims to provide artificial vision for people with severe vision loss, potentially including people blind from birth. Musk also said he aims to increase Telepathy’s output speed substantially over the next few years, exceeding the speed of human speech or typing.
An earlier breakthrough designationBlindsight received FDA Breakthrough Device Designation in September 2024, providing support for regulatory engagement during development. It remains in development and awaits regulatory approval, with human studies still needed to determine eligible patients, the vision it can provide and its long-term effects.
▮ SIGNALMoving from thought-based device control to artificial vision expands Neuralink’s intended applications, with patient access still dependent on clinical development and approval.
❯ Binance reserves funds after a roughly $6.3 billion theft as trackers identify automated laundering
More than $6.3 billion reserved for compensationBinance says it has set aside more than $6.34 billion in company funds, including about $5 billion already held in the same tokens affected by the theft. Reuters, carried by CNA, reports an estimated $6.3 billion stolen and Binance’s statement that partners helped freeze approximately $3 billion. Binance is funding the compensation reserve from its own assets, while recovery of the frozen funds remains in progress.
Concentrated, automated transfersA Techmeme excerpt of the Financial Times report quotes Elliptic as saying almost all the roughly $6.3 billion left in the first 30 minutes. Chainalysis observed many rapid transactions and circular transfers, indicating automation intended to obscure the trail.
Proof of reserves to verify fundsBinance has pledged to cover user losses and says funds will be publicly verified through proof of reserves. The theft exceeds the roughly $1.5 billion Bybit incident in 2025. The rapid outflow of assets puts pressure on the exchange to mobilize its own funds and restore customer balances, testing its reserves and emergency compensation capacity.
▮ SIGNALLarge-scale automated transfers leave less time for tracing and freezing funds, increasing the importance of readily available exchange reserves for emergency compensation.