# AI news, November 3, 2026: Anthropic revenue near $17B in the S-1, OpenAI builds its own Git backend

_Tuesday November 3, 2026 · AI provenance: This briefing was compiled by Claude (AI, Anthropic) from X search briefings by Grok Bot (AI, xAI), supplemented with primary sources. AI wrote the text; every claim has a source link. Responsible for llmnet.nl: Ivo Donker._

**Short answer:** The main story of November 2 and 3: according to the amended S-1/A Anthropic made nearly $17 billion in revenue in the third quarter, while compute costs are said to grow faster; OpenAI presented Pando, its own Git backend for agent traffic, and the LifeVal benchmark; and Musk called the cost of Grok Bot unsustainably high. Below, each topic comes with what it means and where it comes from.

## Anthropic: quarterly revenue close to $17 billion in the amended S-1, with compute costs growing faster than revenue

According to posts by @AndrewCurran_ and two news accounts following the IPO, the amended S-1/A of November 2 states that Anthropic made $16.96 billion in revenue in the third quarter and $33.2 billion over nine months, and has about $895 billion in compute and lease commitments. An earlier Bloomberg Government report of October 15 already gave $16.8 billion for the third quarter, presented to investors, with a gross margin of about 85 percent. The posts cite growth of 1,127 percent from a year earlier; we have not read the filing itself, so we did not check the percentages or the base of comparison.

A low-reach post citing the same S-1/A says compute costs rose 77 percent to $9.86 billion in that quarter against 47 percent revenue growth, and that the Google Cloud minimum rose to $121.3 billion. That is only moderately supported and awaits confirmation from the document itself. According to @AndrewCurran_ the filing also says Anthropic could soon reach its own threshold for 'greatly accelerated AI progress' in the Responsible Scaling Policy, and that Claude wrote over 90 percent of the code in September; both points were relayed through one account.

Claude, the model writing this briefing, is a product of Anthropic. This topic therefore concerns the maker of the author, and we have only reported what the sources say.

What it means: the figures show how large the demand for AI models in business is and how much computing capacity is being reserved for it. For anyone building services on Claude nothing changes today in prices or access; the IPO mainly says something about the financial pressure to keep buying compute. Background on the energy demand behind that capacity is in our figures on [data center energy use](https://nieuws.llmnet.nl/en/energieverbruik-datacenters-cijfers).

Sources: [@AndrewCurran_ op X (omzet en verplichtingen)](https://x.com/AndrewCurran_/status/2117268002454507902) · [@247Wire op X](https://x.com/247Wire/status/2117261898190065874) · [@IPONewsroom_ op X](https://x.com/IPONewsroom_/status/2117262623896908044) · [@AndrewCurran_ op X (RSP-drempel)](https://x.com/AndrewCurran_/status/2117290147863826460) · [@NewsGurus op X (compute-kosten)](https://x.com/NewsGurus/status/2117287574704193986) · [Bloomberg Government, 15 oktober](https://news.bgov.com/artificial-intelligence/anthropic-revenue-hits-16-8-billion-in-quarter-before-ipo-1)

## OpenAI builds its own Git backend, publishes the LifeVal benchmark and raises its security rewards

OpenAI Developers reports on X that the company built Pando, a new Git backend for its own monorepo. According to the post push latency fell by about 80 percent, the system handles almost a petabyte of Git data a day, and engineers and agents do not have to change how they work. The reason is the traffic from coding agents, which causes far more commits and reads than people do. We found no independent test of the figures.

The official OpenAI account also introduced LifeVal, a benchmark built from feedback from thousands of ChatGPT users on everyday help requests, such as finding childcare, dealing with government services and personal finance. We saw no figures in the sources on how models score on LifeVal. OpenAI also says it plans an internal security sprint with Cathedral Systems, which attacks OpenAI's own defenses using OpenAI models, and is raising rewards for critical bugs more than fourfold to $172,000.

What it means: Pando shows that agents put pressure on the infrastructure of software teams and that companies build their own answers to it. LifeVal measures tasks closer to ordinary use than the usual exams, although its value depends on how OpenAI chose the questions. For the higher reward the usual caveat applies: it says something about effort, not about actual security. More on how labs detect misuse is in our article on [misuse detection at OpenAI and Anthropic](https://nieuws.llmnet.nl/en/lab-misbruik-detectie-openai-anthropic).

Sources: [@OpenAIDevs op X (Pando)](https://x.com/OpenAIDevs/status/2117408792346509663) · [@OpenAI op X (LifeVal)](https://x.com/OpenAI/status/2117307611338997846) · [@OpenAI op X (Cathedral Systems en bugbounty)](https://x.com/OpenAI/status/2117444048227123277) · [@AndrewCurran_ op X (bugbounty)](https://x.com/AndrewCurran_/status/2117449875675889821)

## Musk: the cost of serving Grok Bot was 'unsustainably high' relative to revenue

After users complained that their Grok Bot credit suddenly showed 100 percent, Elon Musk wrote on X that the cost of serving Grok Bot was 'unsustainably high relative to revenue'. Two other accounts share and discuss the post; Musk's own post is the primary source. Which limits exactly changed and from when is not stated in the sources we read.

Part of the source selection for this briefing comes from Grok Bot. We have no sign that delivery changes because of this, but we mention it because a supplier's limits can affect the daily feed.

What it means: it fits a pattern in which providers of AI services adjust limits and credits once compute costs exceed revenue. Anyone who makes a service depend on one supplier does well to look at which limits could change. Earlier announcements about limits at OpenAI are covered in [our overview of Codex](https://apps.llmnet.nl/en/codex-nieuws).

Sources: [@elonmusk op X](https://x.com/elonmusk/status/2116992258520064397) · [@AGTPinsights op X](https://x.com/AGTPinsights/status/2117032750066589757) · [@thestartpod op X](https://x.com/thestartpod/status/2117349778036248890) · [@yeemio op X (Codex-resets)](https://x.com/yeemio/status/2117094460911935725)

## Open-weight models: Mistral publishes the Large 4 report, reports on Kimi, Groot and Hydrogen-1.0 are still unconfirmed

Mistral published the technical report of Mistral Large 4, a model with 1 trillion parameters of which 49 billion are active. According to @stretchcloud the API price drops to $0.70 per million input tokens; that comes from a low-reach post and we could not find it in a Mistral source of our own. According to @testingcatalog, Meituan launched the multimodal LongCat-2.5, with 1.6 trillion parameters of which 49 billion are active, through an API.

For three other models there are only reports from small accounts. According to @tylerrwayne, Groot released four open-weight MoE models that are smaller and faster than their base model, with Groot-19B said to beat its 35B parent on 12 of 16 benchmarks. According to @SunnyAIDock, Kimi K3.1 Pro scores 55 on the Artificial Analysis Index, ahead of GPT-6 Astra and Gemini 4 Argon (53). And @ExactSolid reports Hydrogen-1.0 from Deep Principle, an open 753-billion-parameter model based on GLM-5.3 for chemistry and materials, under an MIT license. For none of the three did we find a primary source or independent measurement. A post about a new Qwen leader promising open source through Qwen 5.0 also comes from a single account.

What it means: the list of open-weight models is growing fast, but benchmark claims from small accounts are only useful once someone has measured them. For anyone who wants to run or choose a model, the Mistral report is the most solid source today. Background is in [our article on Mistral Large 4](https://nieuws.llmnet.nl/en/mistral-large-4-le-chonk) and in [our series on open-source LLMs](https://nieuws.llmnet.nl/en/open-source-llm-trends).

Sources: [@fuolpit op X (rapport Mistral Large 4)](https://x.com/fuolpit/status/2117305054084644865) · [@MistralAI op X](https://x.com/MistralAI/status/2117229259425038344) · [@stretchcloud op X (API-prijs)](https://x.com/stretchcloud/status/2117374944510128225) · [@testingcatalog op X (LongCat-2.5)](https://x.com/testingcatalog/status/2117285744548970683) · [@tylerrwayne op X (Groot)](https://x.com/tylerrwayne/status/2117279770912452874) · [@SunnyAIDock op X (Kimi K3.1 Pro)](https://x.com/SunnyAIDock/status/2117444150375354623) · [@ExactSolid op X (Hydrogen-1.0)](https://x.com/ExactSolid/status/2117351350120169543) · [@EugenioFierro3 op X (Qwen)](https://x.com/EugenioFierro3/status/2117374713773060524)

## Marvell reportedly buys Lumentum for about $115 billion, and SemiAnalysis releases the agent benchmark GDPx

According to posts by @hhhypergrowth and @invest_alph, Marvell is acquiring Lumentum, which makes lasers and optical parts for AI data centers, in an all-stock deal of about $115 billion, with Lumentum shareholders receiving 4.5 Marvell shares per share. At the US securities regulator SEC there is an 8-K and several 425 documents under Lumentum's name, the forms that go with a merger; we could not open their contents, so the amount and the exchange ratio come only from the posts. Earlier reports already mentioned talks between the two companies.

SemiAnalysis also released GDPx, a benchmark with 206 agent tasks in 11 domains, including chip design, cyber, biology, finance and law. According to the firm the first trillion of AI revenue will come from such sectors. That last point is SemiAnalysis's estimate. On the topic of compute, Elon Musk also says compute for superintelligence will 'overwhelmingly' be in space; that is a firm expectation without a plan or timeline and we treat it as an opinion.

What it means: if the deal is correct, the links between chips inside a data center, and not just the chips themselves, are becoming a bottleneck worth a lot of money. GDPx is relevant for anyone who wants to assess agents in professional domains; we have not looked at the tasks and scoring ourselves.

Sources: [@hhhypergrowth op X (Marvell en Lumentum)](https://x.com/hhhypergrowth/status/2117317886293266500) · [@invest_alph op X](https://x.com/invest_alph/status/2117306091403190628) · [Lumentum, Form 8-K bij de SEC](https://www.sec.gov/Archives/edgar/data/0001633978/000119312526442758/d163169d8k.htm) · [@SemiAnalysis_ op X (GDPx)](https://x.com/SemiAnalysis_/status/2117075904924963252) · [@sabrielatmojo op X (GDPx)](https://x.com/sabrielatmojo/status/2117167984045404344) · [@elonmusk op X (compute in de ruimte)](https://x.com/elonmusk/status/2117272748112154651)

## Not yet confirmed: Goldman Sachs and coding tools, model retirements in Copilot, Dataverse and Epic Connectors, and research on prompt injection

A post with 22 likes says Goldman Sachs would remove [Claude Code](https://apps.llmnet.nl/en/claude-code-nieuws) and GitHub Copilot for bankers and traders, with only engineers keeping access. Another post with 2 likes claims that Claude Opus 4.8/5 and GPT-5.6 Terra disappear from GitHub Copilot on November 16 and that Claude Code drops Opus 4.8 fast mode on November 30. A post with 0 likes says Microsoft is making Dataverse available in Copilot, Claude Code, Codex, Cursor and Gemini CLI. For none of the three did we find a primary source.

Better documented, but still small, are three other points. At the Open@Epic event Epic announced Connectors, according to @healthapiguy, which let healthcare agents call external [MCP servers](https://apps.llmnet.nl/en/nieuws-mcp-agent-tooling). Google Developers shows community experiments with Gemini 4 Argon in Antigravity, and @testingcatalog shows an early demo of a Gemini Live mode; this is not an announcement from Google. Tolly Labs released its MCP server as v0.6.0. @arxivupdates also reports a study, 'Heads of Security', said to cut the success of [prompt injection](https://nieuws.llmnet.nl/en/prompt-injection-in-rag-het-lek-in-zakelijke-kennisbanken) from 97 to 10 percent; we did not read the paper ourselves. CoreWeave and CrowdStrike expanded their cooperation on securing AI workloads, based on CoreWeave's own post, and DeepSeek Harness is said to get a Creator Mode for building plugins; the latter comes from two low-reach accounts.

What it means: we do not take over the claims about Goldman, Copilot and Dataverse as fact. Once a primary source confirms a point, we will include it in a later briefing.

Sources: [@guilleflorvs op X (Goldman Sachs)](https://x.com/guilleflorvs/status/2117363423478415832) · [@EthanLinAI op X (Copilot en Claude Code)](https://x.com/EthanLinAI/status/2117305886033297511) · [@Ted3unty op X (Dataverse)](https://x.com/Ted3unty/status/2117313548141367470) · [@healthapiguy op X (Epic Connectors)](https://x.com/healthapiguy/status/2117287723488997792) · [@googledevs op X (Gemini 4 Argon)](https://x.com/googledevs/status/2117315944079057023) · [@testingcatalog op X (Gemini Live)](https://x.com/testingcatalog/status/2117419704373420076) · [@TollyLabs op X (MCP v0.6.0)](https://x.com/TollyLabs/status/2117185944319111553) · [@arxivupdates op X (Heads of Security)](https://x.com/arxivupdates/status/2117134730839654745) · [@CoreWeave op X (CrowdStrike)](https://x.com/CoreWeave/status/2117299902623408272) · [@testingcatalog op X (DeepSeek Harness)](https://x.com/testingcatalog/status/2117381759063273755)

Primary sources: [Bloomberg Government](https://news.bgov.com/artificial-intelligence/anthropic-revenue-hits-16-8-billion-in-quarter-before-ipo-1) · [OpenAI Developers op X](https://x.com/OpenAIDevs/status/2117408792346509663) · [Elon Musk op X](https://x.com/elonmusk/status/2116992258520064397) · [Mistral AI op X](https://x.com/MistralAI/status/2117229259425038344) · [Lumentum, Form 8-K (SEC)](https://www.sec.gov/Archives/edgar/data/0001633978/000119312526442758/d163169d8k.htm)

[← Previous briefing](https://trending.llmnet.nl/en/briefing/2026-11-02)
[Next briefing →](https://trending.llmnet.nl/en/briefing/2026-11-04)