Skip to content

← X AI Daily

X AI Daily

Date19 November 2026 · ThursdayWindow2026-11-18 02:00 to 2026-11-19 07:51CoveragePartial

Translation

English translation of the Chinese original.

1. AI models, products and applications

Gemini 4 Neon goes live: $3/$15 per million tokens, 1M context, third on the AA Intelligence Index

At 05:33 Beijing time on November 19, Sundar Pichai announced that Gemini 4 Neon had begun rolling out; Google DeepMind followed at 05:34, Google AI Studio at 05:36 and the main Google account at 05:51 with pricing, benchmark placements and availability. Neon sits between Carbon and Hydrogen: a 1M-token context window, four thinking levels (low/medium/high/xhigh), $3 per million input tokens and $15 per million output tokens, with cached input at $0.15. It is available through the paid tier of the Gemini API (AI Studio, Android Studio), Gemini for business, Antigravity for paid Google AI Pro/Ultra subscribers, and the Gemini app for Pro/Ultra subscribers; a cyber-defense variant will reach eligible defenders in the Fairwind Program within days.

This was the highest-engagement event of the window. @sundarpichai's post drew about 366K views, 6,489 likes, 506 reposts, 360 replies, 587 bookmarks and 300 quotes; Google DeepMind about 239K views and 4,231 likes; Demis Hassabis about 104K views and 3,863 likes; the main Google account about 131K views and 1,267 likes; Google AI Studio about 60K views and 1,691 likes. Metrics were collected around 07:57 on 2026-11-19 (Shanghai time).

Verified: the official Google blog confirms the $3/$15 pricing, $0.15 cached input (a 95% discount), the 1M-token context window and four thinking levels, and states that Neon scores 61 at xhigh on the Artificial Analysis Intelligence Index, third overall, second on the Vals Index (70.9%) where it lands on the Pareto frontier, second on the new Agent Arena index overall (1336) and second on steerability, and Google calls it its lowest fabrication rate yet in internal Antigravity testing. The blog lists Google's own results: Terminal-Bench 4.1 at 65.2%, FrontierCode 2.0 at 46.2% at the default high setting, Terminal-Bench Science 70.5%, RiemannBench 84.0%, HLE-Diamond 71.3%, plus first place on Zapier's AutomationBench (56.3%) and Andon Labs' Vending-Bench 2 ($16,843). These are vendor figures.

Third-party measurement is more sober. Artificial Analysis's post (about 68K views, 1,031 likes) and its write-up confirm: Neon (xhigh) scores 61, one point above Carbon (high) and third overall, behind GPT-6.2 Astra (max) and Claude Fable 5.2 (max); cost per task is $6.44, about 17% below Carbon's non-discounted $7.76, and time per task is about 21 minutes, roughly 40% less than Carbon's ~36 minutes; output tokens per task are actually about 15% higher, so the cost advantage comes entirely from 25% lower token prices and the speed from roughly 110 versus 57 output tokens per second. In Antigravity CLI, AA measures a Coding Agent Index score of 59 (Carbon 55), in line with Claude Fable 5.2 (58) and Opus 5.5 (58) at far lower cost per task ($6.86 versus $12.61 and $15.47). AA also flags a pricing condition that is easy to miss: $3/$15 is 25% below Carbon's standard $4/$20, but Carbon currently sits at a discounted $2/$10 until November 30, so during the discount Neon is actually 50% more expensive.

Judgment: the pitch here is "Carbon-quality at Flash speed"; the price and throughput improvements are checkable while the capability gain is one point. Whether it is good value depends on how Carbon is priced after the November 30 discount ends.

Microsoft ships seven MAI models at Ignite, MAI Code 2 lands in GitHub Copilot

At Ignite 2026 in the early hours of November 18 Beijing time, Microsoft AI released seven models and said MAI Code 2 and MAI Code 2 Flash were available in GitHub Copilot. @github's announcement drew about 72K views, 525 likes and 38 reposts; @MicrosoftAI's post on Licensed Weights about 32K views and 403 likes.

Verified: Microsoft AI's Ignite page lists the seven models as MAI Cyber 2 Flash, MAI Code 2, MAI Code 2 Flash, MAI Image 2.7, MAI Realtime 1, Dragon Copilot Transcribe (built on MAI Transcribe 2) and Decision 1.1. The MAI Code 2 page states that averaged over DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1, MAI Code 2 reaches 98% of GPT-6.1 Sol's performance at 40% of the cost, and 96% of Claude Opus 5.5's performance at 5% of the cost; in blind frontend comparisons by 258 annotators, developers preferred its output 64% of the time. The footnotes give the conditions: three benchmarks were run on Microsoft's internal MSBench platform with the GitHub Copilot CLI and no subagents, FrontierCode scores and costs come from Artificial Analysis, and Code 2's cost is shown at its 50% promotional price. In the same launch, MAI Licensed Weights lets enterprises fine-tune frontier MAI models on their own data and run them inside their own Azure environment; the second-generation Maia 200 accelerator is in production and claimed to deliver 1.7x more tokens per watt; MAI Image 2.7 is claimed #1 on Artificial Analysis image editing at about one-fifth the price of GPT Image 2.6 Sunburst (max); MAI Realtime 1 starts at $0.01 per audio minute; Decision 1.1 is claimed 50x faster and 79x cheaper than GPT-6.1 Sol. Apart from the FrontierCode figure, all of this is Microsoft's own testing or its own claims.

Judgment: the axis is price and "customize it yourself", not a capability jump; the comparisons rest on internal benchmarks and a promotional price, so they should not be treated as settled before third-party replication.

Apple and Cornell release FastFrames: 36 ms per frame at 480p, video generation at interactive latency

At 02:59 Beijing time on November 18, the paper FastFrames: Video Generation at Interactive Latency was submitted as arXiv 2611.23226 v1 (authors from Apple and Cornell; project page and code on GitHub); the weights had gone up on Hugging Face a day earlier. Discussion in the window was set off by @_akhaliq's post, with about 39K views, 547 likes, 67 reposts and 600 bookmarks.

Verified: the arXiv abstract states that FastFrames is a family of fast autoregressive video diffusion models: the 1.3B variant runs at 36 ms per frame (28 RPS) on a single B200 while matching the quality of a state-of-the-art model that is nearly 8x slower, and the 14B variant runs at 45 ms per frame (22 RPS) on an 8x B200 node with better visual quality. The paper's key finding is that operations treated as implementation details, such as VAE decoding and KV cache updates, account for over half of end-to-end latency, and it replaces them with lightweight alternatives. The Hugging Face model card gives the output format as 832x480, 81 frames at 16 fps, under Apple's own apple-amlr license rather than a standard open-source license; the repository releases all checkpoints, including the causal teacher, consistency-distilled and DMD-distilled models. Chinese- and Japanese-language posts followed with hands-on tests, for example @pratikg (about 6,843 views, 117 likes) running it through MLX on a 64GB M5 Pro Mac mini — an individual port, not official support.

Judgment: the value of this work is turning "tens of milliseconds per frame" into an interactive medium (camera moves driven by keystrokes) rather than another higher-quality video model; but the 14B variant needs eight B200s, so consumer hardware remains out of reach.

OpenAI publishes an update to its research acceleration report: autonomous completion of 2–4 hour tasks rises from 23% to 46%

At 07:08 Beijing time on November 19, OpenAI's @daveholtz posted that the company was releasing new data on how AI is accelerating AI research, an update and expansion of its first research acceleration report from September (the post text is truncated in the API, with about 1,622 views and 36 likes — far below model-launch posts).

Verified: the OpenAI report page (November 18) states that from March to October the estimated success rate of agents completing tasks that would take a human two to four hours without human intervention rose from 23% to 46%, and for four-to-eight-hour tasks from 13% to 35%; valued at API list prices, the median researcher used about $2,400 a day and the 90th-percentile researcher about $23,000, up 62% and 105% from $1,500 and $11,000 in September; the official page states that internal deployment (internal Codex, ChatGPT and API usage combined) still accounts for only 2% of total research compute. From August 1 to October 18, 2026, estimated success rates without human intervention were higher for Communicate (69%) and Analyze (58%) than for Build (43%). The report also indicates that agents cannot yet reliably automate multi-week frontier R&D.

One second-hand retelling needs correcting: a summary post on X gives $7,000 as the baseline for the 90th percentile, but the official September baseline is $11,000; $7,000 is an earlier figure and is not used here.

Judgment: what matters is the framing — success rates are bucketed by how long the task would have taken a human, so 46% applies only to the two-to-four-hour band, and the spend is valued at API list prices rather than actual cash. Reading it as "R&D is nearly automated" goes beyond what the report concludes.

Meta's Muse passes 10 million downloads as Reuters documents overreach and cross-account leakage

In the early hours of November 19 Beijing time (afternoon of November 18 US Eastern), Reuters published an investigation: Meta's consumer agent Muse has more than 10 million downloads since its September launch, and about a quarter of people who set up accounts return daily; but more than two dozen internal posts reviewed by Reuters show Muse reaching into end-to-end encrypted Signal messages, inventing purchases that never happened, and sending users chats that appeared to belong to someone else, with employees worried about cross-account mixing and privacy exposure. Meta spokesperson Daniel Roberts called the examples "isolated incidents", saying many were hallucinations that have been addressed; Reuters, working with Signal and the AI Now Institute, replicated the access and concluded the issue was ongoing (on Android via notification permissions, on macOS via accessibility and screenshot tools, blocked on iPhone), and Apple's OS update this week closed that macOS pathway.

Discussion on X was far smaller than the story itself: a post pointing to the Reuters report drew about 418 views and 5 likes, and another summary about 860 views and 10 likes. Facts are taken from the Reuters story.

Judgment: this is a clear case of consumer agents being tested systematically on the cost of "permissions in exchange for capability": once an agent is asked to operate a computer on someone's behalf, OS permissions and third-party app boundaries become the security problem itself. Meta's "isolated incidents" framing and Reuters's replication still diverge, and both should be kept on the record.

Ming-3.0-audio-flash takes first place on the Artificial Analysis speech reasoning leaderboard

At 16:37 on November 18, the official account of Ant Group's inclusionAI, @AntLingAGI, posted that Ming-3.0-audio-flash ranks first globally on Artificial Analysis speech reasoning (Big Bench Audio) with 99.8% accuracy, ahead of Grok Voice 3.0, Gemini 3.8 Live Extended Thinking (High) and GPT-Realtime-2.1 High. The post drew about 62K views, 243 likes, 24 reposts and 122 bookmarks.

Verified: Artificial Analysis's speech-to-speech leaderboard currently lists Ming-3.0-audio-flash first on Big Bench Audio speech reasoning (99.8%, with 42 models evaluated on the board); the model itself is Ant's inclusionAI 126B-total, 6.5B-active audio-language model, previously open-sourced under the MIT license, as shown on its Hugging Face model page. This is therefore an existing open model's placement on a new leaderboard, not a new model release.

Judgment: first place on a single dataset and a single leaderboard shows that real-time speech reasoning is crowded and the leaders are close together, but it is not proof of a better end-to-end voice product.

Australia's Agentaus 1 regains attention in the window over a "DeepSeek wrapper" dispute

Trellis Data's 220B model Agentaus 1 actually launched on November 10 — the company newsroom states "Canberra, Australia, 10 November 2026", with 220B parameters, roughly 50% larger than its previous model, a global ranking of 43 for foundation models, and free, subscription, enterprise and API tiers. What is new in this window is that it was picked up again by Chinese- and English-language circles on November 17–18 and became embroiled in a "wrapper around a Chinese open model" dispute. Co-founder Michael Gately restated on X at 03:54 on November 18 Beijing time that "Australia's own model is open to the world today"; that first-party post drew about 1.64 million views, 2,331 likes and 198 reposts, the highest engagement on this topic in the window, and also states 84.4% on MMLU-Pro and a world ranking of 43. @DocumentingAGI's summary drew about 34K views, 272 likes and 21 reposts, and @cgtwts's write-up about 23K views and 160 likes. Metrics were collected around 08:05 on 2026-11-19 (Shanghai time).

Verified: beyond the company release above, Switzer's report and PS News attribute "84.4% on MMLU-Pro, 43rd globally at the time of testing and first in Australia" to the founder, with PS News noting Trellis plans to publish those results later; the company newsroom itself says only "a world ranking of 43" and gives neither 84.4% nor a benchmark table.

The controversy also needs recording: HuggingNews reported that a researcher ran three fingerprint tests against 72 known models and alleged a 95% match with deepseek-v4-flash-0731; Trellis responded that the model is proprietary and not a fine-tune of a Chinese model. The allegation has no complete materials available for independent review and does not change the vendor's published scores, so it is recorded only as an unresolved dispute.

Judgment: the real selling point of sovereign AI is data residency and procurement compliance, not leaderboard rank; but both "43rd" and "homegrown" are currently vendor claims, 84.4% does not even appear in the company's press release, and until the fingerprint dispute is resolved, "homegrown" should not be treated as verified.

MyClaw 3.0: multiple agent harnesses inside a cloud computer

At 21:58 on November 18, @MyClaw_Official launched MyClaw 3.0, pitched as "raised AI teammates with tools built in": any harness including Claude Code, Codex, Hermes, Pi and OpenClaw with shared context, more than 5,000 live data sources, a cloud computer with up to 64 vCPU, 128GB RAM and 1.28TB storage, support for browser and Slack/Teams/Telegram/WhatsApp channels, and a setup of about 30 seconds. The post drew about 128K views, 518 likes, 221 reposts, 291 replies, 334 bookmarks and 283 quotes — the second-highest-engagement product post of the window.

Verified: the website confirms the product structure — six runtimes (OpenClaw, Hermes Agent, Pi, Claude Code, Codex, DeepSeek Harness) on one cloud computer, 50+ selectable models, one memory shared across agents, a built-in Data Router that needs no API keys, and task assignment from Slack, Telegram and WhatsApp. The site does not state the "5,000+ data sources" or "64 vCPU" figures; those appear only in the launch post and are recorded as vendor claims.

Judgment: this is not a new model but existing agent runtimes packaged as "cloud teammates" with a data layer added. Reposts and quotes are close to half of likes, suggesting this kind of agent-infrastructure story travels further right now than model benchmarks.

OpenAI's blog publishes a Dean Ball essay, "The Silent Consensus"

On November 18, OpenAI's blog published The Silent Consensus, a long essay by Dean Ball; a note at the top says the piece is adapted from a Harvard lecture and represents the author's personal views, which may not reflect those of OpenAI or his colleagues. The author's X post (02:12 on November 19) drew about 125K views, 1,027 likes, 127 reposts, 62 replies, 637 bookmarks and 72 quotes — the highest-engagement OpenAI-related post in the window apart from Gemini 4 Neon.

Verification status: the essay's existence, publication date, author and the personal-views disclaimer can all be checked on the linked page; the arguments in the post are the author's claims, which this report neither restates as fact nor evaluates.

Judgment: its significance is that an OpenAI executive is joining the public debate about "AI and human primacy" in a personal-views format, putting the industry's internal default assumptions on the table; this is a discourse event and changes no model or product facts.

2. AI compute and industry

Architect Labs uses AI to design a new CPU architecture: Frostbite claims 4.59x performance per watt over a traditional CPU

At 06:58 Beijing time on November 19, Architect Labs team member @ekinthesumbul posted about Frostbite, a general-purpose CPU designed from scratch around interaction nets, with the ISA, microarchitecture, RTL, verification and final GDS all produced by the company's AI systems. The post drew about 143 views, 8 likes and 1 repost — a new release that has not yet drawn high engagement.

Verified: the company blog (November 18) states that across 25 workloads (10 used for search and tuning, 15 held out for final evaluation, including FFT, sorting, graph search, JSON parsing, Huffman coding and k-d trees), one Frostbite quad achieves a geometric mean of 4.59x performance per watt versus "a modern, silicon-proven superscalar CPU" (not named for legal reasons; 0.895mm² at the same process, dual-issue in-order), with 6.02x on the search set and 3.83x on the held-out set, and 8.77x performance per area on average; the chip reaches 800MHz at synthesis on 7nm and 500MHz in the final GDS, with a quad plus four L3 nodes at 0.792mm²; the architecture removes most of the cost of deep cache hierarchies and coherence protocols, keeps data in small tile-local memories, and exhibits work-stealing behaviour no human specified. Planned next steps include releasing the RTL and ISA, an FPGA software playground and a preliminary operating system.

Judgment: this is a rare checkable disclosure for "AI-designed hardware" (with a held-out set plus area and frequency conditions), but it is still self-reported with no third-party replication; the 4.59x figure can only be independently tested once the RTL is actually published.

Xlue starts a "health world model" with a $35 million seed round

At 01:37 on November 19, @XlueAI announced that it is live: the company is building a foundation model for human health that understands health states, predicts how they evolve and simulates how treatment changes their course. The post drew about 22K views and 62 likes, with 16 reposts.

Verified: Forbes reported on November 18 that the Carnegie Mellon spinout raised a $35 million seed round led by Sarah Guo's Conviction; RuntimeWire's report adds that founder Chenyan Xiong is a CMU associate professor who spent five years on large language models at Microsoft Research, that the company was spun out in 2025, and that it focuses on predicting cancer risk from longitudinal electronic health records (CATCH-FM). The company site uses the same "Health World Model" positioning.

Judgment: the "world model for medicine" pitch is still a vision, and public materials contain no independently checkable efficacy data; the size of a $35 million seed round shows capital will pay a premium for clinical data plus foundation models.

Voyage Labs raises a $12 million seed round as its AI travel agent Atlas enters private beta

Announced on November 17 and spreading on X from the early hours of November 18 Beijing time: Manik Gupta, formerly head of product at Uber with a Google Maps and Microsoft Teams background, founded Voyage Labs, whose first product Atlas is a conversational travel agent covering planning, price comparison, booking and in-trip change monitoring from inspiration to the trip's end, with a private-beta waitlist opening that day. Spread was mostly through aggregator accounts: @WhaleInsider's post drew about 46K views and 164 likes, and @lauraveyne's write-up about 5,510 views and 59 likes.

Verified: the PR Newswire release states the $12 million seed round was led by Kleiner Perkins with SignalFire, SV Angel and others participating; TechCrunch's report confirms Atlas is in private beta, positioned as an assistant for the whole journey "from inspiration to home".

Judgment: vertical agents differentiate on whether they can plug into inventory and booking systems and handle real transactions like rebooking, not on the model; "private beta plus waitlist" means there is no completion rate to evaluate yet.

3. Chinese-language AI ecosystem

Tencent Hunyuan open-sources Hy4 flash weights (follow-up)

At 12:39 on November 18, @TencentHunyuan announced that Hy4 flash weights were open, releasing Hy4 flash, Hy4 flash-FP8 and Hy4 flash-Base in one go, stating 224B total parameters, 12B active, a 1M context window and native multimodality. The official post drew about 157K views, 1,692 likes and 198 reposts. Metrics were collected around 08:05 on 2026-11-19 (Shanghai time).

Chinese-language commentary is supplementary here: @NFT_Chen (about 6,815 views, 22 likes) framed it as Tencent formally open-sourcing; @4395Aron (about 298 views) noted the official wording was two variants; @taiyanghere (about 176 views) pointed out explicitly that "what was opened is the weights, not a new model launch — the model itself shipped two days ago". Engagement on these posts is low, but the information is specific.

Verified: Pandaily's report and neurohive's summary confirm the three variants were published to Hugging Face, ModelScope and GitCode, with FP8 being 8-bit quantized and Base not instruction-tuned; the specs match the November 16 launch.

Deduplication: the November 17 daily covered the Hy4 flash launch and pricing, so this section only covers the new step of open weights.

Judgment: open weights turn a "cheap everyday work model" from an API into something self-hosted, and the real value depends on what hardware can run a 224B model — the window contained no credible local throughput test.

Doubao Work: 8 million weekday active users and RMB 2 billion enterprise ARR three months after launch

On November 18, Doubao officially announced results for Doubao Work three months after launch: weekday active users (DAU) passed 8 million, and enterprise annual recurring revenue passed RMB 2 billion; the product had added a "workbench" feature on November 16. Discussion on X lagged well behind the official cadence: @Stargridcompass's post drew about 1,329 views, 8 likes and 13 replies — a new item that has not yet formed high engagement.

Verified: PEdaily's relay states that "on November 18, Doubao officially announced that three months after launch, Doubao Work's weekday active users (DAU) passed 8 million and enterprise customer annual recurring revenue (ARR) passed RMB 2 billion", with Sina and NetEase carrying identical wording and Huxiu's news flash matching; both figures are company statements, with no third-party audit or definition of scope (for example, whether free-quota users are included).

Deduplication: the November 18 daily compared billing for the same tasks between Doubao Work and WorkBuddy but did not report these figures, so only the new operating data is written here.

Judgment: if the definitions hold, this is a rare public disclosure of both activity and revenue for a Chinese office agent; but "weekday active" and "enterprise ARR" are not standard industry definitions, and cross-product comparisons need the definitions aligned first.

Coverage gap for this issue: the November 16 daily does not exist, so historical deduplication is missing one day.