❯ Meta and Microsoft rein in employee use of Claude, with Claude Code users inside Meta down from about 60,000 to about 30,000

Two big customers cut backAccording to an exclusive from The Information, Meta and Microsoft are reducing employees’ use of Anthropic’s Claude and steering them to their own AI tools. The number of Meta employees using Claude Code has fallen from about 60,000 earlier this year to about 30,000. Microsoft has cut its projected internal Anthropic spend, previously more than $1 billion a year, by more than a third.

The substitutes are in-houseAs relayed by AI Weekly, Meta’s own coding tool MetaCode has passed 30,000 users and Muse Code has more than 6,000 internal users. The report says Meta’s proprietary tools and its Muse Spark models have largely displaced the need for Claude. At Microsoft, leadership told staff to curb Claude use, and engineers in the Experiences and Devices group have moved to GitHub Copilot CLI.

Layoffs explain part of the dropMeta’s halving is not all deliberate substitution. It laid off about 10% of its workforce this spring, which accounts for part of the decline. The report also says Meta had projected spending of up to $10 billion a year on Anthropic. Microsoft’s $5 billion investment in Anthropic is unchanged, and Anthropic’s enterprise bookings through Azure and Bedrock continue to grow.

What it means for Anthropic before its IPOAnthropic’s draft prospectus shows nearly a quarter of revenue coming from two unnamed customers, and many large customers have no long-term contracts. Big tech companies are both its customers and its competitors, and once their own tools are good enough they have reason to keep the money in-house. Whether enterprise bookings make up for lost internal usage will show in its next revenue disclosure.

▮ SIGNALA model company’s largest customers are often the ones most able to build models themselves, so this revenue arrives fast and needs no reason to leave.

❯ SemiAnalysis testing: at the same subscription price, Claude Opus 5.5 offers about five times the usage of GPT-6.1 Sol

What $20 buysResearch firm SemiAnalysis limit-tested AI subscription plans from nine companies and concluded that at the same price, Anthropic subscriptions provide about 5x the usage of OpenAI’s when converted at API list prices. A $20 Claude Pro plan on Opus 5.5 is worth about $1,200 of API usage, against about $280 for a $20 ChatGPT Plus plan on GPT-6.1 Sol. At the $200 tier the figures are about $12,000 and $1,400.

How it was measuredSubscription plans do not publish exact allowances and show only a usage meter. According to Implicator, SemiAnalysis fed chunks of War and Peace and technical essays into paid accounts, recorded how many tokens each tick of the meter consumed, and converted the totals into dollars using each lab’s public API prices and its own September mix for agent workloads, to within plus or minus 5%.

OpenAI just halved its limitsThe gap widened with recent changes. OpenAI halved the allowances for every model tier in its $200 plan, with existing subscribers keeping old limits until October 29. Its new $500 plan offers only 21% more Astra usage than the old $200 plan. Anthropic raised Opus allowances while cutting Opus 5.5 pricing. Limits on the two flagships, Fable 5.1 and GPT-6 Astra, are similar, and the fivefold gap is in the mid-tier models.

Generous limits mean thin marginsThe report cautions that the figure measures allowances at list prices and does not prove a subscriber finishes five times as many tasks. It also estimates that if users consumed their full Opus 5.5 allowance, Anthropic’s gross margin would be minus 369%, turning positive at 6% only at an average 20% consumption. Subscriptions are about a tenth of Anthropic’s revenue but consume more than 40% of its inference compute.

▮ SIGNALSubscriptions all cost $20 and $200, so the real price war is fought over allowances nobody can see, and whoever gives more is trading gross margin for users.

❯ Codex delivers on day one of its 28-day pledge: GPT-6 Astra and GPT-6.1 Sol go from 30 to 50 tokens per second for subscribers

Day one delivered speedThibault Sottiaux, who runs Codex at OpenAI, posted on X on October 5: “Day 1,” default speed is about 50% faster for subscribers on GPT-6 Astra and GPT-6.1 Sol, across OpenAI’s products and partners using Sign in with ChatGPT, including OpenCode, Pi, Amp and Devin. He added that speed went from 30 tokens per second to 50. The post has more than 1.6 million views.

The promise made a day earlierThis is the first delivery on the pledge he made the day before: for 28 days, the team will each day either ship a clear improvement relevant to most Codex and Work users or give users a usage reset. The speed-up requires no change on the user’s side and, by his account, takes effect within two hours.

No third-party dataRuntime Wire points out that going from 30 to 50 is actually an increase of about 67%. OpenAI supplied no independent measurements, the figure describes only how fast text streams and not how quickly a whole task completes, and it does not change how much work a user can do before hitting usage limits.

Mixed reactionsEthan Mollick commented on X that the reset approach is weird and makes sense only to a narrow slice of people, while most users expect to plan their usage rationally instead of getting random prizes of tokens. A speed-up is felt by every subscriber and is easier for ordinary users to understand than a reset.

▮ SIGNALDelivering speed first while allowances are being questioned answers a question only heavy users calculate with a metric everyone can feel.

❯ Reflection unveils its first open-weight model, Beam, saying it needs a third to a quarter of GLM-5.2’s inference compute

501 billion parameters, 23 billion activeNvidia-backed Reflection AI unveiled its first open-weight model, Beam, on October 5. According to MarkTechPost, it uses a sparse mixture-of-experts architecture with 501 billion total parameters and only 23 billion active at a time, targets coding, reasoning and agentic tasks, and supports a context of up to 1 million tokens. Full weights are due in late October under Apache 2.0, with an early version available through a waitlist.

Trained in four weeksIn a mixture-of-experts design the model contains many expert modules and calls only a few for each answer, so total parameters are large while actual computation is small. Reflection says Beam was pretrained in under four weeks on 6,144 Nvidia GB300 GPUs with 23.8 trillion tokens. The reinforcement learning stage ran more than 100 million rollouts and used about 1.3 billion sandboxes.

Claims to approach China's leading open modelsReflection says Beam is comparable to GLM-5.2 on most tasks and approaches Qwen 3.8-Max on coding and agentic tasks, while acknowledging that Kimi K3 has stronger raw capability. The claim of three to four times less inference compute is, by the company’s own description, an approximate compute comparison and not a measured inference cost, leaving out prompt prefill and serving overhead. Benchmarking firm Artificial Analysis says it has been given access and is testing independently.

Where American open models standResearcher Nathan Lambert commented on X that Reflection joins Nvidia and Thinking Machines in releasing its strongest model and coming up behind Chinese counterparts, and that the clearest takeaway is that Chinese teams are very good at building LLMs. Companies that want to run models on their own servers gain an option from a US company, but whether it really saves compute awaits third-party checks once the weights are out.

▮ SIGNALAmerican makers of open models no longer claim to beat their Chinese peers and instead pitch more efficiency at equal capability, a quiet change in the frame of reference.

❯ OpenAI is in talks for a new $30 billion round, with UAE funds discussing up to $10 billion in total

$30 billion at a $1.4 trillion valuationAccording to Bloomberg, as relayed by The Next Web, OpenAI is in talks with several investment funds from the United Arab Emirates, including Abu Dhabi’s MGX, to anchor a $30 billion funding round. The funds have discussed putting in as much as $10 billion altogether. The pre-money valuation is about $1.4 trillion.

OpenAI sets the price itselfThe approach is unusual. Normally investors submit terms and a lead negotiates the valuation with the company. This time OpenAI is presenting a fixed price and has chosen no lead. BlackRock is in talks to take part, and the University of California’s endowment and existing investors Thrive Capital and Andreessen Horowitz are also in discussions. The talks are not settled and details could change. OpenAI and BlackRock declined to comment, and MGX did not respond.

Up more than 60% in seven monthsOpenAI’s previous round was in March, when it raised $122 billion at a post-money valuation of $852 billion. MGX has raised about $50 billion this year, holds stakes in OpenAI, Anthropic and xAI, and runs a $30 billion infrastructure vehicle with BlackRock, Microsoft and Nvidia.

The weight of Gulf moneyFew buyers can absorb a $30 billion raise. MGX invests in several competing labs at once, betting on the industry and not on one company. For OpenAI, setting its own price with no lead suggests it judges demand from buyers to be strong enough.

▮ SIGNALWhen a company can raise $30 billion at a stated price with no lead investor, what is scarce in capital markets is no longer money but assets large enough to hold it.

❯ OpenAI will test visual ads during image generation in ChatGPT this month, starting with free users in the US

Ads while the image rendersOpenAI announced a new visual ad format on October 5: when users generate images in ChatGPT, ads with images will appear on the page. According to BleepingComputer, testing starts later in October in the US with a small initial group of advertisers, for users without paid subscriptions.

Labeled, and separate from answersOpenAI says ads will be clearly labeled, kept separate from the image being created, and will not influence ChatGPT’s responses. Ads can use images to show product inspiration, how a product is used or the experience of a service. The company also says safeguards are designed to keep ads out of emotionally vulnerable, sensitive or otherwise unsuitable conversations.

The ad infrastructure behind itA set of third-party ad technology companies is being connected at the same time: Hightouch, Tealium and LiveRamp for conversion tracking, AppsFlyer, Adjust, Branch and others for attribution, and DoubleVerify and Integral Ad Science for brand suitability. ChatGPT has 1.2 billion weekly users, which is why advertisers want in.

Turning free users into revenueThese ads target users who do not pay, and every image generated consumes compute. Placing ads in the wait for an image is relatively unobtrusive. How large this becomes depends on whether users accept it and whether OpenAI can convince people that ads really do not affect answers.

▮ SIGNALThe chatbot has ended up on the same road as the search engine: gather users with a free product, then sell their attention to advertisers.

❯ Inference chip company Etched fields investment offers at a $40 billion to $50 billion valuation, up from $21 billion two months ago

Offers at double the priceAccording to TechCrunch, citing sources, as relayed by AI Weekly, AI inference chip company Etched is receiving new investment offers valuing it at between $40 billion and $50 billion. The talks are early, no financing has been completed, and terms could change.

Chips built only for inferenceEtched was founded in 2022 by Gavin Uberti, Chris Zhu and Robert Wachen, all Harvard dropouts. It makes chips only for inference, the step in which a trained model answers user requests, and says its chips process tokens faster and at lower cost than Nvidia’s. According to Trending Topics, it splits inference into two phases, reading in the prompt and generating output, and designed dedicated hardware for each.

Valued at $5 billion at the start of the yearEtched was valued at about $5 billion early this year. In July it closed a $300 million Series C led by Sequoia at a $10.3 billion valuation. Its last round, $700 million led by Jane Street, valued it at $21 billion. Jane Street is also an early customer and runs an Etched rack in its own data center. The company said in July it had secured $1 billion in customer orders, and it has about 400 employees.

Valuation ahead of deliveryAI Weekly cautions that the step-up reflects Etched’s roadmap and not its revenue, that customer identities and delivery timelines are unclear, and that delays into 2027 would hit the valuation hard. The price investors are offering is a bet that it can get these chips into customers’ hands on time.

▮ SIGNALA doubling in two months for an inference chip maker buys the possibility of an alternative to Nvidia, one that has not yet been proven by deliveries at scale.

❯ Two former Groq engineers sue the board, saying the $20 billion licensing deal with Nvidia short-changed common shareholders

The target is the deal structureBenjamin Serebrin and Joshua Rubin, two former engineers who hold Groq shares, filed suit in Delaware last Friday against Nvidia and Groq’s board. According to National Technology, the complaint says the board improperly sold core assets and took billions of dollars in benefits for itself, senior management and affiliated funds that it did not share with other stockholders.

A license in name, with 90% of staff goneGroq is an AI chip company. In December 2025, Nvidia announced a “non-exclusive” technology license with it worth about $20 billion, with Groq nominally remaining independent. As laid out by Crypto Briefing, $17 billion was a cash licensing fee and another $3 billion was a pool of Nvidia stock for the roughly 200 engineers who joined Nvidia. About 90% of Groq’s workforce moved to Nvidia, including founder Jonathan Ross.

Common holders got lessThe plaintiffs say common stock was sold at a discount while engineers who left for Nvidia received separate compensation. The lawsuit is aimed at exactly this structure: no traditional acquisition, a licensing fee plus stock for the team that joins, and proceeds divided unevenly among classes of shareholders. In September, the US Department of Justice opened an antitrust investigation into whether the deal sidestepped merger notification requirements.

ResponsesGroq called the allegations “meritless” and said the license delivered “exceptional value.” Nvidia declined to comment. If a court finds the board breached its duty, other deals built the same way would face shareholder claims, and substituting “license plus hiring” for an acquisition would become more expensive for large companies.

▮ SIGNALThe acquisition that is not called an acquisition avoided antitrust review and the common shareholders’ vote, and now both are coming to collect.

❯ Huawei and Qualcomm sign a multi-year patent cross-license that gives Qualcomm rights to Huawei’s LogicFolding chip technology

Cross-licenses, plus a patent purchaseHuawei and Qualcomm announced a multi-year patent license agreement on October 5 covering 5G, AI, computing and networking technologies. According to The Next Web, each gains rights to the other’s patents, and Qualcomm will also buy some of Huawei’s US patents in computing, AI, networking and other areas, subject to regulatory approval. Financial terms and royalty rates were not disclosed.

LogicFolding recognized by a rivalAccording to Bloomberg, the technologies licensed to Qualcomm include Huawei’s recently unveiled LogicFolding chip architecture and optical packaging for high-speed interconnects. Without access to ASML’s extreme ultraviolet lithography machines, Huawei has not pursued further transistor shrinking and has instead raised performance by improving data transmission inside the chip. LogicFolding is the product of that approach.

From payer to collectorHuawei first paid Qualcomm patent fees in 2001 and received its first licensing income only in 2011, from Motorola. Its licensing business has been revenue-positive since 2021. Huawei says that once this deal completes, the total value of its patent licensing agreements will exceed $6.9 billion. US trade restrictions since 2019 limit Huawei’s access to advanced chips and software and do not automatically block patent licensing.

A channel outside the sanctionsQualcomm taking a license to Huawei’s chip technology amounts to acknowledging that the technology has value. Export controls can stop equipment and chips flowing into China but do not reach intellectual property flowing the other way. Whether the deal clears review will show how US regulators view this kind of cooperation.

▮ SIGNALA company barred from buying advanced equipment is now selling technology licenses to a US chipmaker, so controls have changed where goods flow without stopping technology from creating value.

❯ SpaceX stock rises 7.6% in a day to its highest since mid-June, returning Musk to trillionaire status

Up 7.6%SpaceX shares closed up 7.6% on October 5 at their highest level since mid-June. According to CNBC, as relayed by Techmeme, the immediate cause was Morgan Stanley calling the stock “cheap,” which also returned Elon Musk to trillionaire status.

A roller coaster since listingSpaceX listed on Nasdaq on June 12 this year at $135 a share, raising $86 billion in the largest IPO ever. It closed its first day at $161 with a market value of about $2.1 trillion, briefly making Musk the world’s first trillionaire. By late July the shares had fallen about 50% from their peak. He controlled about 42% of shares after the IPO.

AI as a reason for the valuationMorgan Stanley analyst Adam Jonas has a $300 price target. SpaceX acquired xAI in February, and its AI unit now rents compute to outside customers including Anthropic. Investors valuing the company now have to count this AI business alongside rockets and Starlink.

One company, three businessesVery little of SpaceX’s stock trades freely, with only about 5% floated at the IPO, which amplifies price swings. It is at once a space company, a satellite operator and a compute lessor, and conclusions differ widely depending on which valuation method an analyst applies, one reason the shares move so sharply.

▮ SIGNALA rocket company’s share price now needs its AI business to explain it, which shows markets are ready to reprice any asset that can be tied to AI demand.

❯ Doubao’s team is said to have worked through the Mid-Autumn holiday for six days to rush out a personal assistant, Spell, aimed at Meta’s Muse

An overtime post on social mediaAccording to a October 3 report by PEdaily, posts on social media say ByteDance’s Doubao team launched a rush project over the Mid-Autumn holiday, with integration testing due on September 27 and a trial version due on September 30, about 6 days in all, and several upstream and downstream teams cancelling their holidays. One post read: “First time working overtime at ByteDance on triple pay.” ByteDance has not responded.

What Spell isThe product is codenamed Spell internally, with “Xiaodou” as its planned product name, and is positioned as a personal assistant aimed at Meta’s Muse. According to Huxiu, Spell has been in internal testing since April and comes from the Doubao phone assistant team, so this was an urgent rebuild and not a start from scratch. Huxiu specifically notes that the overtime details are online rumor and should not be treated as settled.

The pressure from MuseMeta’s Muse launched on September 8, focused on doing things for users instead of chatting. It passed 900,000 downloads in six days and 3.4 million by September 24, and topped the US App Store. PEdaily says Tencent has launched a cloud personal agent, LightVela, and Alibaba has plans for a Qianwen personal agent, though none has made the business model work.

Copying may not travelPoe Zhao, an analyst who follows Chinese tech, said on X that he doubts such a product carries over to China, reading Muse as a product of the open web. In China, shopping, food delivery and travel live inside each company’s own app, and how much a personal assistant can get done depends on whether those apps open their interfaces.

▮ SIGNALA big company can rebuild a product in six days but cannot get other companies’ apps to open up to it in six days, so the bottleneck for personal assistants in China is access, not the model.

❯ Google is found to be developing “Superprojects” for Gemini, and a Nano Banana 2.1 label appears in Flow

Two unconfirmed findingsBoth items below come from observation of Google’s app code and interface, and Google has confirmed neither. TestingCatalog, which tracks unreleased features, said on October 5 that Google is developing a feature for Gemini internally named “Superprojects.” It is the next iteration of Projects and will replace Notebook projects in Gemini, with files and chat sessions in these projects accessible across various Google products.

Chats, email and calendar in one workspaceAccording to Crypto Briefing, Superprojects surfaced in app teardowns in September under the internal name “Robin Superproject,” and aims to let users combine chats, emails, calendar events and notes into a single workspace. The Notebook projects it would replace launched only in April, and there is no firm launch date.

Nano Banana 2.1Nano Banana is Google’s image generation model, and the current Nano Banana 2 is officially Gemini 3.1 Flash Image. In late September TestingCatalog found a “Nano Banana 2.1” label in the web build of Google Flow, replacing an earlier “2.5 Flash” reference. On October 5 it said 2.1 appears to be available in Flow, but the output looks similar to Nano Banana Pro and it is unclear whether requests are actually routed to a new model.

What these signs suggestIf Superprojects is real, Google is putting material scattered across its products in one place for Gemini to use, the same direction as the project and workspace features OpenAI and Anthropic have introduced recently. But names and labels in code are often changed or dropped before launch, and until Google announces something, neither the details nor the timing can be relied on.

▮ SIGNALEvery assistant is competing for the “project” container, because whoever holds a user’s files, email and calendar becomes the default place to work.