HX LLM Tech Briefing · 2026-11-04 (Issue 1)

HX Research | Beijing time 2026-11-04 (Wednesday) | Coverage: 2026-10-28 to 11-04, focusing on 11-02 to 11-04

Highlights

This week's biggest story was capability and finances laid out side by side. In its amended S-1, Anthropic disclosed revenue of $33.24 billion for the first three quarters and purchase commitments of about $895.4 billion, and stated in its risk factors that, by internal estimates, it may soon reach the threshold in its Responsible Scaling Policy (RSP) for significant acceleration of progress driven by AI. On evaluation, SemiAnalysis launched GDPx, a 206-task agent benchmark on which the best model, Claude Opus 5.5, scores only 12.0% — long-horizon professional work is far from solved. Chinese model releases concentrated on domestic compute and agentic office work: Meituan's LongCat-2.5 (1.6 trillion total parameters, trained entirely on domestic compute) and the official release of SenseTime's SenseNova 6.8 Flash; China's National Development and Reform Commission (NDRC) said it will accelerate foundational, comprehensive AI legislation. On the platform side, Anthropic and xAI each added default watermarks to API text output, and Claude introduced message threads that keep the conversation on the server — a sign that compliance capabilities and the cost of long sessions are becoming the main axes of API competition.

I. Model releases

1. Meituan releases LongCat-2.5: a 1.6-trillion-parameter natively multimodal model trained entirely on domestic compute (11/02)

Meituan released LongCat-2.5, a natively multimodal model aimed at long-horizon agent tasks⁠^[1]. According to the company, it has 1.6 trillion total parameters and was trained entirely on domestic (Chinese) compute; mid-training used sequences of up to 1M tokens for long-context training. Per-GPU throughput for long-context SFT was 86% higher than a DDP baseline. Inference uses W4A4 (both expert weights and activations quantized to 4 bits); compared with a W8A8 INT8 baseline, throughput rises by about 40%, the wait for the first token falls by about 16%, and the time to generate each subsequent token falls by about 20%⁠^[1]. The company says its scores on several benchmarks fall between Claude-Opus-4.8 and Claude-Opus-5; these are vendor-reported and have not been independently reproduced. Media reports put the active parameters at about 48B per token; that figure does not appear in the excerpts of the official page, so we treat it as a media report for now.

Why it matters: trillion-scale MoE has now completed full training runs on domestic compute for two consecutive generations. Engineering practices such as W4A4 and supernode communication optimization are directly relevant to other Chinese teams constrained by single-card memory.

2. SenseTime releases SenseNova 6.8 Flash (official release, 11/02)

SenseTime launched the official release of SenseNova 6.8 Flash⁠^[2], emphasizing long-term memory across tasks and sessions and proactive interaction (for example, alerting users to affected items when information changes), as well as generation of animated slide decks, animated posters, Blender 3D, SVG and web designs, with integration into Figma workflows. The company says it "surpasses some international frontier models" in agent collaboration, data analysis and presentation generation, but gave no specific scores. The model is available on the SenseNova Token Plan⁠^[2].

Why it matters: Chinese vendors' Flash-tier models are shifting their focus from answering questions to delivering editable work products; the contest is now about how completely they handle office and design work.

3. Flower Labs releases Endeavor 1.1: private deployment, plus inference inside the EU (11/03)

Flower Labs released Endeavor 1.1⁠^[3], which can be used as a hosted service, in a private cloud or on premises. It provides readable reasoning traces; users own the outputs and the traces and may use them to train their own models. New in this release are GPU inference inside the EU and image input. According to the company, GPQA rose from 92.0 to 94.4 and HumanEval from 98.2 to 99.9⁠^[3]. The company claims its overall intelligence score exceeds the mainstream products of OpenAI and Anthropic, but that score is an equal-weighted average of only four benchmarks — GPQA, AIME 2026, HumanEval and IFEval — so its scope is narrow and it should be read with caution.

Why it matters: data sovereignty, auditable reasoning and the right to train your own models on the outputs are becoming selling points in Europe against US closed-source APIs.

4. Pending: full weights of Kimi K3.1 Pro promised "before November 4"

Moonshot AI's technical blog of 10/23 describes K3.1 Pro as its "3T-class" open-weight model with a 1M context, and promises that the full model weights will be released before November 4, 2026⁠^[4]. According to the company, it scores 65.7% on Terminal-Bench 4.0, above GPT-6 Astra (58.2%) and slightly below Claude Opus 5.5 (66.4%). The company also disclosed that cybersecurity-related data was removed from the mid-training and RL stages for the open-weight release: on xdev-bench-hard (pass@3), a checkpoint without the removal scores 90.5% and the final version 81.0%⁠^[4]. As of publication (the morning of 11/04, Beijing time), the moonshotai organization on Hugging Face still had no public K3.1 repository; the latest was still Kimi-K3 (checked by our team through the HF API).

Why it matters: if released on time, this will be the largest open-weight release of the week. Its approach of deliberately weakening cybersecurity capability before opening the weights may also become a reference point for open-weight safety practice.

II. Research papers and evaluation

5. SemiAnalysis launches the GDPx agent benchmark: best score 12.0% (11/02)

GDPx draws its tasks from 14 professional RL-environment companies: 206 tasks across 11 domains — chip design, cybersecurity, biology, finance, law, quantitative trading, recursive self-improvement (RSI), hardware design, materials, alignment and software engineering. Each task comes with a real environment and a default time limit of 8 hours; the headline score is pass@1 averaged over 3 attempts⁠^[5]. The leaderboard shows Claude Opus 5.5 at 12.0% (about $51.41 per task), Gemini 4 Carbon at 10.9% (about $16.29 per task) and GPT-6 Astra at 9.6% (about $53.99 per task); the best open-weight model is DeepSeek 4.1 Flash 1017 at 5.4% (about $0.94 per task)⁠^[6]. SemiAnalysis notes that, under the binary scoring originally planned, every model scored 0 on the finance and law tasks, and that the totals of Opus 5.5 and GPT-6 Astra were pulled down by safety-classifier blocks⁠^[5].

Why it matters: evaluations of the GDPval kind, which grade a document produced in one go, are no longer enough; evaluations with real environments and long time limits are replacing them. On value for money, Gemini 4 Carbon comes close to the top at about a third of the per-task cost, which is worth watching.

6. CUA-Master: scaling computer-use agents through "open-world recursive improvement" (arXiv 2611.03126, 11/02)

The paper proposes a nested two-level flywheel: the outer loop uses open-world corpora and the model's current weaknesses to guide the next round of data construction; the inner loop mines real workflows from instructional videos as seed tasks, then synthesizes verifiable complex tasks, and adds a "snapshot, roll back and retry" mechanism for collecting trajectories⁠^[7]. According to the paper, CUA-Master-27B scores 80.2 on OSWorld Verified, 36.9 on OSWorld 2.0, 43.9 on WeaveBench and 25.0 on Agents' Last Exam⁠^[7].

Why it matters: the bottleneck for computer-use agents is moving from the model itself to the data flywheel. Mining real workflows from videos is a low-cost, scalable data source.

7. CoworkOS: an agent framework for long-horizon collaboration (Renmin University of China, arXiv 2611.03178, 11/02)

CoworkOS separates task strategy (left to the model) from execution management (left to the runtime) and uses a three-layer architecture that supports synchronous or asynchronous delegation to sub-agents. Humans and agents can edit the same document, slide deck, CAD model or Blender scene together, with layered workspaces and persistent semantic memory⁠^[8]. According to the paper, with GLM-5.3 as the backbone and five rounds of automated post-training of Qwen3.6-35B-A3B, the average score across 8 benchmarks rose from 33.19 to 42.95 (+9.76)⁠^[8].

Why it matters: agents doing their own post-training has now closed a small-scale loop in an open-source framework, which echoes this week's Anthropic disclosure about RSI.

8. uni-robot-agent: open-weight models can be robot-manipulation agents too (arXiv 2611.01560, 11/01)

The authors propose a unified framework for robot-manipulation agents. According to the paper, without privileged inputs, an off-the-shelf Qwen3.8-Flash-Next scores 31.39 on RoboDojo, above the published GPT-6 Astra-RoboProbe result (28.97), and performs competitively on 19 real-world tasks. The authors reduce effective robot manipulation to four general capabilities and recommend developing them together⁠^[9].

Why it matters: this reduces robot-agent research's dependence on closed frontier models, so that follow-on training research can be done on open models.

9. SWE-Trail: guiding data collection for coding agents with past trajectories (Ant Group, arXiv 2611.02074, 11/01)

SWE-Trail does not give past experience to the agent as a prompt (which would make training-time and test-time inputs differ); it uses the history only to decide which states deserve more sampling⁠^[10]. According to the paper, a 124B student model initialized from Ling-3.0-flash-base achieves average resolve rates on SWE-bench Multilingual and SWE-bench Pro that are 2.33 and 1.60 percentage points higher, respectively, than with independent sampling⁠^[10].

Why it matters: competition in coding-agent post-training now reaches down to how trajectories are sampled; how the sampling budget is spent directly affects data efficiency.

10. IdeaTaste: the "research taste" of LLMs is about 18%–26% of the human community's (CMU, arXiv 2611.03424, 11/02)

The benchmark has models write research proposals from a snapshot of the literature up to 2025 and compares them with the work actually published in 2026 — 6 models, 4 research fields and more than 90,000 proposals⁠^[11]. According to the paper, agents reach about 26% of human level on citation matching and only about 18% on idea matching; neither richer domain context nor a larger population noticeably narrowed the gap⁠^[11].

Why it matters: the weakness of "AI scientists" lies not only in execution but in choosing what to work on. This gives a quantitative reference point for claims made about such products.

11. Order-Only On-Policy Distillation (O³PD) (arXiv 2611.02112, 11/01)

O³PD has the student model reproduce only the ranking of the teacher's top-K tokens rather than fit the exact probabilities, so it is insensitive to the teacher's temperature and needs no temperature tuning⁠^[12]. According to the paper, across 6 teacher–student pairs, O³PD beats standard OPD on the HMMT25/AIME24/AIME25 average in every case, by up to 2.9 points, and by up to 5.6 points on AIME25⁠^[12].

Why it matters: this is a simple distillation improvement that adds almost no cost and is of direct practical value for post-training the reasoning of small models.

12. Follow the Gradient: tracing safety degradation back to the training data (DatologyAI, arXiv 2610.37481)

The paper proposes SimpleLoRIF: by backpropagating through only the last 4 Transformer blocks, it can compute gradient embeddings for millions of examples and find fine-tuning samples that look harmless but whose gradients align with unsafe behaviour⁠^[13]. According to the paper, removing the most relevant 20% of samples from Tulu 3 yields SafeTulu, which lowers the GCG attack success rate on Qwen, Llama and Olmo by up to 22%, outperforming filtering by an LLM judge⁠^[13]. (The paper was released on 10/27 and entered the HF trending list this week.)

Why it matters: the finding that benign data can also weaken safety now comes with a scalable data-filtering remedy, which is especially useful for teams fine-tuning open-weight models.

III. Products and pricing

13. Claude API introduces message threads: the server keeps the conversation, and later requests send only new messages (11/02, beta)

According to the Claude Platform release notes, long conversations and agent loops on the Messages API can use message threads (beta header message-threads-2026-08-12): the server stores the conversation, so each later request only needs to send the newly added messages, which lowers the time to first token. A thread expires 5 minutes after the last request and is deleted at most 1 hour after its last use. It is available on the Claude API, AWS, Google Cloud (Anthropic-hosted) and Microsoft Foundry⁠^[14]. (Media also reported that the time to first token in long sessions falls by about 35%; that figure does not appear in the release notes.)

Why it matters: the cost and latency bottleneck of long-horizon agents increasingly lies in resending the whole context every turn; keeping state on the server is becoming a standard API capability.

14. Claude text watermarking extended to Mythos 5, Opus 4.7 and Sonnet 4.6 (11/03)

According to the release notes, from 11/03 text generated by Claude Mythos 5, Claude Opus 4.7 and Claude Sonnet 4.6 carries Anthropic's text watermark. On Sonnet 4.6, if a request sets temperature/top_p/top_k to non-default values, the output may not be watermarked⁠^[14]. (Fable 5.1 / Mythos 5.1, released on 9/1, are already watermarked by default.)

Why it matters: watermarking is spreading from new models to existing ones, and content provenance is becoming the default configuration.

15. xAI: Grok API text output watermarked by default from 11/07

xAI's developer documentation says that from 2026-11-07, text output from 7 models — grok-4.7, 4.6, 4.5 and 4.3, and three variants of grok-4.20 — will carry an invisible watermark by default, including structured output and tool-call arguments. The response header x-watermark: true indicates that a watermark has been applied. Team administrators can turn it off per model, but must confirm that an applicable legal exemption exists and take on the labelling obligation themselves; the detection API is available on application only. The company says the move is to comply with the EU AI Act⁠^[15].

Why it matters: read together with item 14, this shows the labelling obligations of the EU AI Act pushing the major vendors to watermark by default at the API layer. Enterprises that rely on downstream rewriting or on text detection need to assess the impact.

16. OpenAI sharply raises security bug bounties (11/02)

According to the company, the median of the reward ranges for medium, high and critical vulnerabilities has on average risen to more than four times its previous level; critical vulnerabilities in core products such as ChatGPT, Codex and the API platform can earn up to $30,000, a special top critical tier up to $100,000, and up to about $172,000 once all reward multipliers are included. It has also added a report-quality multiplier, five loyalty-bonus levels and a bonus pool for focused campaigns, and will pay the difference retroactively on bounties awarded in the 30 days before the announcement⁠^[16].

Why it matters: as models themselves get better at finding vulnerabilities, frontier labs are also raising incentives for outside white-hat researchers — both attack and defence are speeding up.

17. OpenAI: using models to accelerate engineering for the physical world (11/03)

OpenAI described internal uses of its models in CAD, simulation, PCBs, manufacturing drawings and BIM coordination. According to the company, its hardware team's model usage (measured in tokens) has grown 23-fold since February, and two employees used Astra in their spare time to design a piece of test equipment in one week. The company also acknowledged that Astra does not consistently account for manufacturers' capacity constraints, tolerances and assembly details⁠^[17].

Why it matters: after programming, mechanical and electronic engineering may be the next fields where agents are deployed at scale — consistent with GDPx's choice of hardware and chip design as key domains.

18. Media report: Grok Bot cuts usage limits as Musk calls serving costs "unsustainably high" (11/01–11/02)

According to TechCrunch, over the Halloween weekend some Grok Bot users found their weekly usage limits used up early. Musk replied on X that Grok Bot's serving costs are unsustainably high relative to revenue. The new limits have not been announced⁠^[18]. (This comes from a media report and an executive's social-media post; xAI has not issued a formal announcement.)

Why it matters: agent products consume far more inference per use than chat products, and the mismatch between subscription pricing and actual cost is starting to show.

IV. Open source and tools

19. DeepSeek Harness launches Creator Mode

The official DeepSeek Harness page lists four run modes: standard, PTC, minimal and creator. Creator mode can inspect the current runtime and try out Cordis plug-ins in memory, and is used to create custom agent presets. The Harness developer preview is open source and can be run with npx @deepseek-ai/dsh web⁠^[19]. According to media reports, the mode launched around 11/02, and the company also published the prompts for 7 demo plug-ins under the MIT licence⁠^[20] (the launch date comes from media reports; the official page gives no date).

Why it matters: agents writing their own plug-ins through conversation turns the harness from a fixed product into a platform that can extend itself, continuing the Super-agent idea DeepSeek put forward last week.

20. NVIDIA releases Cosmos Framework and Cosmos NIM (11/02)

An NVIDIA technical blog introduced the open, modular Cosmos Framework, which covers world-model training, RL (DPO, GRPO, PPO, OPD, DiffusionNFT), distillation and quantization, and inference; it supports models including Cosmos 3, Alpamayo, GR00T, Qwen and pi0.5, and provides Cosmos 3 Nano simulation recipes for autonomous driving, robotics and surgery. Cosmos NIM provides accelerated inference; according to the company, speculative decoding is on average about 2–3× faster on H200, and Cosmos3-Nano generation needs 58 GiB of GPU memory (BF16) / 44 GiB (FP8)⁠^[21].

Why it matters: world-model toolchains are converging on the mature practices of LLM post-training, which will lower the bar for embodied-AI teams to customize world models.

V. Policy and industry

21. NDRC: accelerating foundational, comprehensive AI legislation (11/03)

According to Xinhua (reposted on the Chinese government's website), Wang Ruomeng, deputy director-general of the NDRC's Department of Innovation and High-Tech Development, said at a press conference on 11/03 that during the 15th Five-Year Plan period China will speed up foundational, comprehensive AI legislation; establish safety standards and ethical guidelines covering the whole AI life cycle; support AI regulatory sandboxes for different fields and promote mutual recognition of regulatory results; and work to include new occupations such as model trainers and data annotators in the national occupational classification. He also said that 38 countries have joined the World AI Cooperation Organization as founding members⁠^[22].

Why it matters: this is the first time officials have explicitly named comprehensive legislation and life-cycle safety standards as work priorities within the 15th Five-Year Plan framework. The compliance boundaries for Chinese model vendors may move from scattered regulations toward unified legislation.

22. Anthropic's amended S-1: $33.24 billion in revenue for the first three quarters, and a risk factor on nearing the RSI threshold (11/02)

According to CNBC's report on the amended S-1, revenue for the nine months to 9/30 was $33.24 billion, up about 1,200% year on year; Q3 revenue was about $17 billion, up about 47% quarter on quarter (Q2 and Q1 grew 144% and 132% respectively); the nine-month net loss was $104 billion, mainly from fair-value changes in convertible notes and derivative liabilities; and purchase commitments and lease agreements total about $895.4 billion, almost all of it related to compute and infrastructure⁠^[23]. According to the risk-factor text quoted by TechCrunch, Anthropic says that, based on internal estimates for future models, it may soon meet its RSP's definition of significant acceleration of progress driven by AI; it has not determined that any current model has crossed that threshold, and it is considering revising the RSP to ensure that humans stay in control of frontier-model development⁠^[24].

Why it matters: this is the first time a leading lab has listed AI accelerating AI R&D as a material risk in a statutory filing. Compute commitments of nearly $900 billion also show that the pace of compute supply has become a core variable in valuation.

23. OpenAI launches the Grid Protectors Program: data centres cut load when the grid is under strain (11/02)

OpenAI announced a nationwide Grid Protectors Program under which it agrees with utilities and regulators that its future compute campuses will reduce their power draw when the grid is under strain, and commits to paying its own costs for electricity and supporting infrastructure. Its campus in Effingham, Georgia is the pilot: the agreement with Georgia Power includes up to 1,000 MW of flexible demand response, which the company says is equivalent to the average demand of about 700,000 households⁠^[25].

Why it matters: power is replacing chips as the main constraint on compute expansion. Flexible load may become a precondition for approval of large AI campuses.

HX view (our team's judgements, not news reporting)

1. Evaluation and pricing of long-horizon agents will be rewritten across the board within 6 months. GDPx's single-digit scores, Claude's server-side threads and the tightening of Grok Bot's limits show that the industry's understanding of both capability and cost is still framed around single-turn chat. We expect quotes based on per-task pricing and real-environment evaluation to become the main reference for enterprise procurement.
2. Watermarking and provenance will go from a compliance option to an API default. Anthropic and xAI extended default watermarking to existing or flagship models in the same week, driven mainly by the labelling obligations of the EU AI Act. Chinese vendors' APIs aimed at overseas markets will very likely follow.
3. AI accelerating AI R&D is turning from a research topic into a governance issue. Anthropic's wording in its S-1, together with open-source closed-loop automated post-training experiments such as CoworkOS, means that RSI-related monitoring indicators may enter regulatory review and investment due diligence. Watch for Anthropic's subsequent revisions to the RSP.
4. The combination of domestic compute and open weights is still accelerating. If Kimi K3.1 Pro is open-sourced as promised, then together with LongCat-2.5's engineering experience on domestic compute, the gap between Chinese open models and closed frontier models may narrow further. But as GDPx shows, the gap on long-horizon real-world tasks remains clear.

References

1. [1]Meituan: Meituan releases the LongCat-2.5 model, a new breakthrough for Chinese models on Chinese chips (in Chinese) (2026-11-02) https://www.meituan.com/news/NN261102280001343
2. [2]SenseTime: SenseNova 6.8 Flash official release (in Chinese) (2026-11-02) https://www.sensetime.com/cn/news/sensenova-6-8-flash-release-20261103-1421
3. [3]Flower Labs: Introducing Endeavor 1.1 (2026-11-03) https://flower.ai/blog/2026-11-03-introducing-endeavor-1.1
4. [4]Moonshot AI: Introducing Kimi K3.1 Pro (technical blog) https://www.kimi.ai/blog/kimi-k3-1-pro
5. [5]SemiAnalysis: GDPx: Measuring Frontier Model Capabilities Across the Most Valuable Domains (2026-11-02) https://newsletter.semianalysis.com/p/gdpx-measuring-frontier-model-capabilities
6. [6]GDPx official leaderboard https://gdpx.semianalysis.com/
7. [7]CUA-Master (arXiv 2611.03126) https://huggingface.co/papers/2611.03126
8. [8]CoworkOS (arXiv 2611.03178) https://huggingface.co/papers/2611.03178
9. [9]Democratizing Robot-Use Agents (arXiv 2611.01560) https://huggingface.co/papers/2611.01560
10. [10]SWE-Trail (arXiv 2611.02074) https://huggingface.co/papers/2611.02074
11. [11]IdeaTaste (arXiv 2611.03424) https://huggingface.co/papers/2611.03424
12. [12]Order-Only On-Policy Distillation (arXiv 2611.02112) https://huggingface.co/papers/2611.02112
13. [13]Follow the Gradient (arXiv 2610.37481) https://huggingface.co/papers/2610.37481
14. [14]Claude Platform Release Notes (entries for 11/02 and 11/03) https://platform.claude.com/docs/en/release-notes/overview
15. [15]xAI Docs: Output Watermarking https://docs.x.ai/developers/advanced-api-usage/watermarking
16. [16]OpenAI: Rewarding defenders with bigger bug bounties (2026-11-02) https://openai.com/index/rewarding-defenders-with-bigger-bug-bounties/
17. [17]OpenAI: Accelerating engineering for the physical world (2026-11-03) https://openai.com/index/accelerating-engineering-for-the-physical-world
18. [18]TechCrunch: Elon Musk throttles Grok Bot limits due to 'unsustainably high' costs (2026-11-02) https://techcrunch.com/2026/11/02/elon-musk-throttles-grok-bot-limits-due-to-unsustainably-high-costs/
19. [19]DeepSeek Harness developer preview (official page, in Chinese) https://deepseek.com/harness/zh/
20. [20]TAO Media: DeepSeek Adds Chat-Based Plugin Creation to Harness With Creator Mode (2026-11-02) https://www.tao.media/deepseek-adds-chat-based-plugin-creation-to-harness-with-creator-mode/
21. [21]NVIDIA Technical Blog: Build World Models for Interactive Simulation with NVIDIA Cosmos (2026-11-02) https://developer.nvidia.com/blog/build-world-models-for-interactive-simulation-with-nvidia-cosmos/
22. [22]Chinese government website (Xinhua): Deepening the "AI+" initiative and promoting high-quality development of the marine economy (in Chinese) (2026-11-03) https://www.gov.cn/zhengce/202611/content_7084907.htm
23. [23]CNBC: Anthropic reports $17 billion in Q3 revenue… (2026-11-02) https://www.cnbc.com/2026/11/02/anthropic-ant-s-1-quarter-financials.html
24. [24]TechCrunch: Anthropic warns that it's very close to RSI capability (2026-11-02) https://techcrunch.com/2026/11/02/anthropic-warns-that-its-very-close-to-rsi-capability-aka-self-building-ai/
25. [25]OpenAI: Introducing the Grid Protectors Program (2026-11-02) https://openai.com/index/grid-protectors-program