Daily editorial briefing

№ 20261106

OpenAI closes $11 billion, while investors push Anthropic's $1.5 trillion valuation down

November 6 delivered a set of signals pointing in opposite directions. On the capital side, OpenAI closed the first tranche of an $11 billion round at a $1.25 trillion valuation…

November 6 delivered a set of signals pointing in opposite directions. On the capital side, OpenAI closed the first tranche of an $11 billion round at a $1.25 trillion valuation, while Anthropic’s IPO ran into resistance and investors pushed $1.5 trillion down to $1.2–1.3 trillion. Corporate spending data published the same day showed token consumption still climbing while spending on those tokens barely moved. Technology moved on a different clock: OpenAI spent 30 hours and $1,213,430 in API fees to surface 948 Linux kernel privilege-escalation paths, about $1,279 each. These signals do not contradict each other. Together they say the capability curve is still steep, and the price curve is falling fast.

Theme 1: OpenAI banks the first $11 billion while Anthropic’s valuation expectations slide

Bloomberg reported that OpenAI closed $11 billion of new funding, the first tranche of a round valuing the company at $1.25 trillion, with investors wiring funds that day. The remainder of the round totals up to $20 billion and is expected within weeks.

At the other end sits Anthropic. According to Reuters, its IPO has met resistance: investors question the $1.5 trillion valuation, with some proposing $1 trillion, against a company target as high as $2 trillion in September. Bloomberg’s November 6 account is more complete. Early October discussions centred on $2 trillion; after the mid-October investor briefing the figure fell to $1.4–1.5 trillion; after the Nasdaq 100 dropped 6.4% in a single session on Thursday, some prospective investors were discussing $1.2–1.3 trillion.

Credit markets are repricing too. The Financial Times reported on October 30 that Blackstone, an anchor investor in a chip financing deal for Anthropic, told participants it expects Anthropic’s post-IPO credit rating to be BB — the highest rung of speculative grade — where Blackstone had previously said investment grade by 2027. That is Blackstone’s forecast, not a rating agency decision, and Anthropic declined to comment on valuation talk.

Read together: the funding channels for leading labs remain open, but the premise underpinning high valuations — that revenue will keep compounding exponentially — is being discounted. What the market is cutting is the multiple, not the technical thesis.

Sources:

Theme 2: Enterprise token use rose 65%; spending rose 1.5%

Data from Ramp, a US corporate finance software company, shows enterprise token spending up 42% month over month in July, then just 1% in August and 0.2% in September. Measured weekly, from early August to mid-October enterprises consumed roughly 65% more tokens while spending only about 1.5% more. Ramp’s blended OpenAI and Anthropic unit price fell from $0.76 to $0.56 per million tokens between late September and late October, a 27% drop.

The price cuts were deliberate. In September GPT-6 Sol’s input price was halved from $4 to $2 per million tokens. Anthropic’s filing says Fable 5.1 reduced the cost of common tasks by about 25%, and Opus 5.5 cut per-task cost about 40% versus Opus 5, partly through list prices and partly because the model consumes fewer tokens. Buyers are also economising: Ramp chief economist Ara Kharazian says companies that know the tools best set per-employee budgets and stop letting long tasks default to the most expensive model.

The effect on Anthropic cuts both ways. Its filing reports third-quarter revenue of $16.96 billion, up 47% from $11.56 billion in the second quarter, but in that same quarter revenue from higher usage was offset by roughly $600 million in model price cuts and contract discounts. In Ramp’s data for the week ending October 23, Anthropic took 60.5% of enterprise token spending — the price war lands on it first.

One boundary is needed: this is a Ramp customer sample, not a census. The top 1% of spenders fell from $9,873 per person per month in July to $9,568 in September, while the median company spends $15.7. Those two populations should not be averaged together.

Sources:

Theme 3: 30 hours and $1.21 million for 948 Linux kernel privilege escalations

OpenAI published results from a GPT-6 Cyber scan of the Linux kernel: 948 local privilege-escalation paths, meaning a route from an ordinary local user to root. The scan ran 30 hours and cost $1,213,430 in API fees, roughly $1,279 per vulnerability.

The number worth noting is not the count but the unit price. Kernel privilege escalations are normally priced in tens of thousands of dollars in bounty programmes and grey markets; extracted in bulk by a model, the marginal cost drops to four figures. Greg Brockman said in the launch post that the model has been used on projects including the Linux kernel and Firefox, and was specifically trained to operate within task scopes set by the operator, alongside a testing programme for approved critical-infrastructure defenders. No evaluation scores were disclosed, and the material does not list whether all 948 are valid or confirmed upstream, so this should be read as a vendor-published scan.

The defensive side produced its own case the same day. Özgün Kültekin, a researcher at security firm Wakefield, used swarms of agents combined with differential fuzzing to test 17,600 login variants and found an OpenSSH flaw roughly a decade old that bypasses long passwords. Scope is limited to Linux builds using musl libc; macOS, Windows and other distributions are unaffected. What matters here is the method: turning a guess into a parallelisable exhaustive search.

Attackers are using the same road. Google/Mandiant, CrowdStrike and Microsoft each disclosed abuse of the AI penetration framework ARTEX. Google observed 27 operations in 23 seconds mapping out AWS permissions; CrowdStrike recorded a single controller hitting more than 200 organisations across 14 countries in 11 days; Microsoft said attackers used stolen credentials to reach high privileges in a corporate cloud environment in under three hours. These are vendor-reported observations, and the underlying samples and attribution have not been fully published.

There is also a human layer. Anthropic measured what happens when a dangerous command is hidden among Claude Code permission requests: only 17% of people rejected it early in a session, dropping to 5% after more than 50 prompts. So while privilege-escalation discoveries get cheap, the final human confirmation step is failing. Telling users to “look carefully” does not fix that.

Sources:

Theme 4: METR audit says Anthropic may have crossed its own warning threshold in July

METR published an eleven-week audit conducted on Anthropic’s internal data. Its conclusion: since October 2025, the rate at which model test scores improve has accelerated roughly 1.9 times, about 80% of that is attributed to AI being used in the company’s own research, and the company may have crossed the warning threshold in its responsible scaling policy on July 15.

The finding carries substantial uncertainty. According to secondhand accounts, METR’s team disagreed internally and both 1.3x and 3x were in play; the claim that most acceleration comes from AI-assisted research is an estimate, not a measurement. The public report redacted large amounts of key evidence, so outsiders cannot independently check the attribution method or confirm the timing of the threshold call.

The dispute therefore concerns method as much as magnitude. Anthropic writes self-acceleration into its risk framework while heavily redacting the evidence behind the conclusion, and those two positions are hard to hold at once. For readers, the defensible conclusion is that an independent organisation is tracking this curve and the direction it measures is acceleration. The exact multiplier is not defensible, and neither is any inference that the model can already improve itself.

Sources:

Theme 5: In one day, two Chinese models at 280B and 560B opened their weights

Baidu released ERNIE 5.5 Flash Base and open-sourced the weights directly: sparse MoE, 280B total parameters, 14B active per token, pretrained on 33.5T tokens. The context claims need separating — pretraining tops out at 128K, while fine-tuning and inference can extend to 1M. Baidu compared it against same-size Base models from DeepSeek, Tencent Hunyuan and Xiaomi MiMo, taking first place on 24 of 30 benchmarks. Those are the vendors’ earlier Base releases, not their latest post-trained models.

Ant Group’s Ling released Ling-3.2-flash: natively multimodal, 560B total parameters with 25B activated per step, 1M native long context, with emphasis on visual-feedback self-checking and adaptive recovery from errors. Parameters and mechanisms come from the official release, so they are vendor claims; long-horizon task performance has not been independently reproduced.

Tencent’s Hunyuan Hy4 has an external reading. Artificial Analysis’ Intelligence Index v5 gives it 35 points, second among Chinese labs. On a 50-task agentic set run three times each, 26 tasks passed all three attempts (52%) and 40 succeeded at least once (80%). The second figure comes from the publisher’s own question set, so it describes hands-on experience rather than a public benchmark.

Taken together, the competitive focus has shifted from parameter count to activation efficiency, context length and multimodal self-checking. Baidu released a Base model that has finished pretraining but not post-training, with the post-trained version promised later — which explains both why it scores well and why it is not yet a product.

Sources:

Theme 6: Image and video models cut both price and waiting time

Artificial Analysis benchmarked GPT Image 2.6, released by OpenAI on November 4: the two models took first and second place on AA-Image-T2I v2.1 and AA-Image-Editing v2.1, with the Flare variant nearly twice as fast as the other. This is an independent evaluator’s ranking rather than a vendor self-assessment; exact margins and comparison images require the original leaderboard.

Google released Gemini Omni 1.2 Flash, its first general availability release. A single video now runs up to 15 seconds, against 10 previously, at $0.07 per second for 720p, 30% below 1.1. At list prices a 15-second 720p clip costs about $1.04; a 360p draft about $0.35, 1080p about $1.56 and 4K about $3.13. It supports successive requests against the same clip, with up to six reference images plus a three-second motion video.

The common thread is cheaper trial and error. On video, drafting at 360p and finishing at high resolution turns the cost structure from pay-per-generation into tiered payment; on images, faster generation directly shortens iteration. Google also notes that fast sports motion and complex physics still need work, and that only English has been fully evaluated, so Chinese prompts carry no quality guarantee.

The field is narrowing. Omni has replaced Veo in the Gemini API, with the Veo 3.1 preview retired on October 22; OpenAI’s Sora app and API closed in April and September respectively. On Runway’s platform, per-second pricing for video with audio puts Omni 1.2 at $0.07 against $0.40 for Veo 3.1, $0.30 for Seedance 2.0 and about $0.16 for Kling 4.

Sources:

Theme 7: Agent memory engineering moves from storage to curation and retrieval

Anthropic published a set of memory experiments. On an internal corpus of roughly 350,000 documents, 420 million tokens and 25 evaluation questions, retrieval alone covered 53% of grading criteria; a knowledge base plus retrieval rose from 54% to 61% as curation rounds increased, using about 20,000 fewer tokens per question. The 61% refers to criteria coverage, not answer accuracy, and the token savings do not subtract the background cost of building the knowledge base.

They call the background curation “dreaming”: keep raw records for traceability, write down goals, plans and verified lessons during a task, then revisit records outside the task to correct errors, remove duplicates and add links. One contract case shows the value of organisation itself — multiple agents co-writing a single table overwrote each other, so the design moved to one page per contract, one agent per contract and verbatim clause quotes, lifting key-fact coverage from 41 of 57 to 57 of 57.

A second set of experiments from Amazon gives the complementary answer. Fine-tuning user facts directly into weights reached 36.9% accuracy; on-policy self-distillation with the same model reached 59.9%, above retrieval at 47.8%, extracted context at 51.9% and raw dialogue at 51.3%, and the gain held at 4B, 9B and 27B. The cost is that weight memory forgets older facts and cannot tell when something was never recorded, so a verify-then-regenerate step is needed, and it adds 6.2 points.

Stanford’s collaboration benchmark shows from another angle that remembering is not the same as using. Across 240 tasks covering eight teamwork skills, the best of 18 models scored 82.2 out of 100; once a teammate creates a problem, every model loses 20 to 47 points. In negotiations where a teammate pushes, applies guilt and threatens to walk away, only 14% of runs reached the best deal for the team. Nearly half of all failures happened after the model had already recognised the problem.

Sources:

Theme 8: Engineering progress is turning verification into executable state

Shopify’s engineering blog describes Reflex, a mobile app architecture designed for AI coding: business logic is separated from the interface, the core layer runs directly on macOS or the JVM, and each page’s ViewModel exposes a uniform interface of current state, available actions and child components. Agents send actions over the command line and read JSON state to verify business logic, input validation, navigation and error handling, with external dependencies such as networks replaced by fixtures.

The result is more than 1,700 behaviour tests running in continuous integration across both platforms in about four minutes; after warm-up a single test needs a few hundred milliseconds without a device, against about two seconds in the iOS simulator. The boundaries are equally clear: it does not verify pixels, cannot prove a button is wired to the right action, still relies on design-system checks and a few screenshot tests for visuals, and is not currently an installable open-source library.

Uber’s Spark work shows another source of the same kind of gain. Gluten/Velox hands Spark SQL operators to a native vectorised engine, and migrated production workloads run about 3.5 times faster with CPU cost down to roughly one third; Quanton goes further on lakehouse writes, reaching about 5.5 times the median steady-state write throughput of the open-source baseline on insert-heavy fact tables and about 4.7 times on update-heavy dimension tables under identical resource quotas. But those figures exclude periodic metadata compaction rounds — with compaction included the gain is about 2.8 times — and the benefits are uneven: the write phase is about 7.8 times faster while deduplication and indexing are slightly slower.

To turn this kind of experience into reusable knowledge, Hugging Face’s handbook is the most complete public material available: roughly 800,000 GPU hours of experiments, 70,000 words and 55 interactive diagrams. It names MoE’s three real problems. Routing must be computed in FP32, because low precision can change close scores and therefore expert selection. Expert load must be balanced through auxiliary losses or bias adjustments. And communication can consume the entire compute gain: in one profile, communication not overlapped with compute accounted for about half a step, and a case where networking misconfiguration fell back to ordinary sockets ran about 25 times slower.

Sources:

Theme 9: 600,000 H200 chips sit in Hong Kong, and the throttle is Beijing

Reuters reported that roughly 600,000 Nvidia H200 chips are idle in Hong Kong warehouses, held up not by Washington but by Beijing. Although H200 sales to China had already been cleared, China is issuing import licences in small batches at a rate of about 10,000 chips per major technology company, and domestic server prices have risen roughly threefold as a result.

The reversal changes the framework for explaining controls. For two years the market assumed the supply bottleneck was US export licensing; the bottleneck is now import-side pacing, with the buyer deliberately slowing arrivals in exchange for control over how compute is allocated. This comes from a single wire-service report, with no second source confirming the licence volume or the price multiple, and no indication of how long it will last.

Two related supply signals arrived the same day. SemiAnalysis says Oracle, OpenAI, Anthropic and Meta have become Google TPU customers, and that Google is building out the surrounding software and developer ecosystem; the summary mentions cost advantages for TPUv7 on some model workloads without giving full test conditions or figures, so it should be read as workload-specific analysis. Separately, The Verge reports that Amazon, Microsoft, Alphabet and Meta have disclosed 2026 capital expenditure plans totalling at least $720 billion, mostly for AI infrastructure — that is total capex, not AI-specific spend, and the report notes that returns on the expansion have yet to materialise.

Sources:

High-value briefs

  • Nace AI open-sources the Drex decision-model family: four sizes at 8B, 9B, 35B (3B active) and 27B. Given a state and a fixed set of options, the model returns a probability per option without generating text. The 27B version lands within 1.6 points of Claude Opus 5.5 on Decision Bench v4 and answers in 183 milliseconds instead of 3.5 seconds; it scores 67.52 on the public Decision Index 0.4.1 suite against 60.70 for Jev 1.14.0. Training adds RLAF, which calibrates targets from a teacher’s probabilities and verified answers and contributes 2.0 index points; a consistency loss cuts the answer-flip rate when options are reordered from 6.3% to 4.2%. https://x.com/omarsar0/status/2118693892300439722
  • ScamWorld measures how often web agents get defrauded: it reconstructs multilingual scenarios including online shops, investment platforms and advance-fee scams. Agent baselines driven by six frontier models were defrauded 11–35% of the time; they were more likely to proceed on payment requests than on requests for credentials or wallet details, and in some trajectories the agent noticed something was wrong and still accepted added conditions from the site.
  • Science’s Last Exam exposes the independent-research gap: 34 institutions built 113 tasks across nine disciplines, testing hypothesis formation, experiment planning, scientific software operation and experiment-video comprehension. The strongest configuration on the initial leaderboard scored 37.5, and every model entered in the end-to-end research track scored zero. That reflects performance on this benchmark, not the absence of value in all research assistance.
  • ReTerm closes the loop between failure and training: researchers at Fudan, CUHK MMLab and Shanghai Jiao Tong keep the agent’s failure states, let the model try multiple recovery paths, then convert verified outcomes into reinforcement signals. ReTerm-9B improved by 14.6, 13.2 and 16.4 points on Terminal-Bench 2.0, SWE-bench Verified and OpenThoughts-TBLite respectively.
  • Iris C2Agent sidesteps key revocation: Tencent’s security analysis describes an 8.9MB Go sample that looks for AI coding tool processes and local model services. The report separates static key theft, quota hijacking and active-session borrowing, noting that revoking an API key does not cover the third path, and that endpoint checks should also look at anomalous process calls and local communication.
  • EMBERS raises fault-discovery throughput: it analyses 1,932 execution traces from three classes of system, distils nine defect categories, then mutates research questions and intermediate search inputs accordingly. Across 16 combinations of framework, model and dataset, throughput at equal budget reached 3.21 times the strongest baseline.
  • QSGS fine-tunes on a tiny data budget: it jointly considers sample relevance to the target task and redundancy between samples, then uses gradient-residual analysis to locate coverage gaps and synthesise additions. The authors report Mistral-7B matching or beating full-data performance at a 0.5% budget; the conclusion is limited to the models and tasks they tested.
  • uni-robot-agent standardises robot tool interfaces: after researchers from Tsinghua and Berkeley unified tool parameters, units and feedback formats, the same Qwen3.8-Flash-Next model scored 31.39 on RoboDojo versus 1.90 with RoboProbe. These are scores under specific test conditions rather than general success rates, and fine manipulation and long-horizon tasks remain weak points for open models.
  • OpenAI partners with Hacktron on vulnerability discovery: the two will test connections to third-party services and use frontier models to strengthen defences. The same day OpenAI and FMF and its member firms published a framework for judging when bypassing an AI system’s safeguards could enable serious cyber harm, and for prioritising stronger protections accordingly. https://openai.com/index/partnering-with-hacktron
  • Claude Code put user email addresses into web requests: some versions wrote the user’s email into requests, making it visible to sites. The official guidance is to upgrade to 2.1.317 or later, with a complete fix expected early next week. https://x.com/KanikaBK/status/2118732394014658822
  • Google DeepMind published two cryptography advances: the team says it found weaknesses in legacy 1024-bit RSA. Attack conditions and applicable scope require the original papers; a summary-level account is not enough to judge real-world impact. https://x.com/demishassabis/status/2118862690135277809
  • A human-attestation pull-request check: an app running on Cloudflare Workers attaches a check to each new pull request that starts red, and only turns green after a person logs in with GitHub, clears Turnstile and ticks a statement that they have read the code — a gate against agent-written code nobody has read. https://x.com/QingQ77/status/2118682057023037467
  • World Labs’ Frameboy demo generates explorable worlds in real time: built on the Atlas model, it streams generated scenes from cloud GPUs, with the optimised version reaching up to 15 frames per second and supporting precise camera control. This is an interactive demo rather than a guarantee of performance across scenes, and Fei-Fei Li shared it.
  • Hugging Face’s open robotics stack, by the numbers: close to 100,000 public robotics datasets (up 48%) and more than 30,000 models (up 41%); LeRobot passed one million downloads, and v0.7.0 adds new world-action models, interactive robot control and remote GPU inference. Microduck and Reachy mini are expected to pass 50,000 units sold this year. https://x.com/ClementDelangue/status/2118751330927780154
  • BioBasicsBench tests routine wet-lab analysis: OpenAI worked with enzyme-design company Ligo on 100 tasks drawn from real experiments, checking whether agents can read raw lab data, identify errors and judge whether the evidence supports a conclusion. It is positioned as a way to find where models must improve, not as a ranking.
  • Agents Anywhere connects local agents to any device: it runs no models itself, instead exposing already-authenticated Codex, Claude Code and DeepSeek Harness sessions through a connector so you can watch progress, reply, approve permissions and open a terminal from a phone, tablet or browser. It sits at the workbench layer and does not change model capability. https://x.com/Pluvio9yte/status/2118720959431774271

🕐 Selected hourly signals

PT time Signal Why it is worth remembering
01:00 Someone reports Opus 5.5 wrote a WASM 2.0 interpreter from the spec in 40 minutes for $8 and passed the full test suite A single personal experiment without the test set or a reproduction, but the cost and time are worth recording
09:00 A curator of Chinese AI discussion argues the “3–4 months behind” gap was never about capability but about distillation lag One author’s framing without checkable data; filed as an unverified lead
11:00 Claude Code cloud sessions can infer dependencies, network access and required keys, then save the environment with keys invisible to the model Moves from configuring an environment each time to reusing one, a prerequisite for cloud-side agents
12:30 Inkling 2 Small reaches 81.7% on ARC-AGI-2 at $0.27 per task, and 96.5% on ARC-AGI-1 at $0.08 per task Reasoning benchmarks are starting to report per-task cost alongside capability
14:20 A secondhand claim that Meituan released a 1.6-trillion-parameter model at about $0.30 per million input tokens A single relayed post with no official release or independent reproduction; a lead only
16:00 Cursor publishes a report on stabilising the combination of RL and self-distillation in production Product teams are starting to publish training detail that can be checked against academic work
16:30 GitHub’s five fastest-growing repositories of the day are all agent infrastructure: semantic ontology, voice, an official skills library, a collaboration workspace and a design system They fill the same gap — agents must understand business semantics, follow speech and fit into existing tools
19:00 Anthropic’s long post on memory engineering drew 194,000 views and reached the top of the day’s reading list Memory has moved from a product feature to a system component that needs separate design

Editorial conclusion

What matters most today is not any single model but the direction of three curves: capability is climbing steeply, unit price is falling steeply, and the revenue curve underpinning high valuations has not kept pace with the second. Security shows the same thing from the other side — when finding a vulnerability costs four figures, the balance of attack and defence gets recalculated. As for METR’s acceleration curve, it is the one that most needs independent review and is most easily over-read.

Sources and method

Reviewed 20 hourly captures and 9 named sources in the target date directory, 278,274 bytes in total, with a rich signal pool. The Chrome Developers blog, Google Research, Cline and XiaoHu.AI published nothing new that day; AI Valley and HubToday fell back to their most recent articles, and HubToday’s body covers events from November 5 and earlier, so it served only as cross-day reference. Vendor claims and single-post figures are labelled as such in the text.

WeChat QR code for 智简 Smart&Concise

FOLLOW ON WECHAT

智简 Smart&Concise

Search in WeChat for independent development and AI updates.