X AI Daily¶
Translation
English translation of the Chinese original.
1. AI Models, Products and Applications¶
OpenAI ships GPT-6.2 Astra and Sol: Astra's input and output prices fall 20%, second place on a third-party index¶
At 01:54 Beijing time on November 18, OpenAI's official account @OpenAI announced that GPT-6.2 Astra and GPT-6.2 Sol were available. Both models reached ChatGPT Work and Codex the same day for Plus, Pro, Enterprise and Edu users; Pro 500 subscribers can use Ultrafast with either model, and Pro 200 subscribers get Ultrafast for the first time, starting with GPT-6.2 Sol. Developers can call gpt-6.2-astra and gpt-6.2-sol through the API. The announcement post had about 4.597 million views, 28.6K likes, 2,991 reposts, 1,727 replies, 4,259 bookmarks and 2,792 quotes; data was collected around 08:00 on 2026-11-18 Shanghai time.
Pricing is the main story this time. Verified: the official blog post states that GPT-6.2 Astra is priced at $8 per million input tokens, $40 per million output tokens and $0.40 for cached input, 20% below GPT-6 Astra on input and output, with a 60% discount on cached input; GPT-6.2 Sol keeps GPT-6.1 Sol's API token prices. OpenAI's announcement thread in its developer community lists the same Astra prices and gives Sol at $2 and $10 per million input and output tokens; the rollout surfaces, API model IDs and Ultrafast access also match that announcement. Artificial Analysis's evaluation post additionally gives cached input for Sol at $0.10.
OpenAI's capability claims centre on professional output and scientific work: Astra scores 5.2 points above GPT-6 Astra on Agents' Last Exam; on the internal GDPval v2 it gains 6.2 percentage points while using 22% fewer output tokens; Sol trails Astra by 0.5 points on FrontierCode 2.0 at about one-fifth the cost. On safety, OpenAI says both models reach the "Critical" threshold for cybersecurity in its Preparedness Framework, and that it therefore tightened safeguards on high-risk cybersecurity requests while planning access for vetted defenders through Daybreak. Methodology and conditions are in the GPT-6.2 system card; these are vendor-run tests, not independent reproductions.
Third-party evaluation gives a finer picture. Artificial Analysis's post (about 336.6K views, 2,662 likes, 237 reposts, 114 replies, 403 bookmarks, 125 quotes) says Astra gains 5 points on the AA Intelligence Index and ranks second, behind Claude Fable 5.2 and ahead of Gemini 4 Carbon at 60 and Claude Opus 5.5 at 59, with the index scores given as 62 for Astra and 65 for Fable 5.2 according to The Decoder's citation of AA; Sol gains 2 points to 57, matching GPT-6 Astra at roughly 15% of its cost per task. The same post notes no progress in coding: Astra is flat with its predecessor on the Coding Agent Index and Sol is 3 points below GPT-6.1 Sol, driven by a 14-point drop on SWE-Atlas-QnA-AA. According to The Decoder's report, AA also measured hallucination rates of 59% for Astra and 55% for Sol on the hard AA-Omniscience question set, a measure that applies to unusually difficult questions rather than everyday use.
What makes this worth attention is that the generation shifts the pitch from "stronger" to "equal or slightly stronger capability at a clearly lower cost per task", while routing the sharpest capability into a controlled channel. In reach, it is the most engaged item of this window by a wide margin over other AI news that day. Judgement: gains concentrate in professional deliverables, science and 3D, with coding a clear weak spot; whether it changes developers' model choices depends on whether the API prices translate into real usage.
First Person Labs launches Second, making "the model of you" a portable local asset, with a $10M seed led by a16z¶
Between 02:02 and 02:18 Beijing time on November 18, First Person Labs' company account, co-founder @kenziyuliu and @a16z announced the product Second and a $10M seed round. Ken Liu's launch post had about 106.6K views, 639 likes, 93 reposts, 115 replies, 399 bookmarks and 73 quotes; a16z's investment post had about 49.2K views, 621 likes and 531 bookmarks; the company's post had about 6,999 views and 71 likes.
The product logic is simple enough: Second installs on a Mac, continuously aggregates email, messages and working context, distils them into a portable user model, then serves your priorities, to-dos, writing samples and preferences to outside agents such as Claude, Muse and Grok Bot through an API. Verified: the co-founder announced a seed of about $10M led by a16z; Dealroom's report likewise describes about $10 million led by a16z's infrastructure team. a16z's investment announcement confirms the investment in Second and states that it runs on the device by default, with users controlling which applications and agents can access it, but the announcement text does not state the round size.
What stands out is the framing: not another stronger assistant, but a fight over who owns the context about you. Judgement: the hard part here is not the model but persuading users to hand their most sensitive data to a local daemon, and whether the outside agent ecosystem will adopt this interface.
Alljoined releases AIR-O, generating an image of what someone is attending to in real time from non-invasive brain signals¶
At 01:00 Beijing time on November 18, brain-computer interface company Alljoined's official account @alljoined released AIR-O. The original post says the model uses only non-invasive brain data to generate an image from scratch in real time depicting where a person is focusing attention, and calls it the first of a new generation of algorithms that read the shape of thoughts directly from brain activity. The post had about 99.4K views, 1,347 likes, 163 reposts, 104 replies, 825 bookmarks and 91 quotes.
Discussion focused on the limits of the claim. Replies from neuroscience-adjacent accounts point to the spatial resolution and signal-to-noise limits of EEG and read the demo as coarse reconstruction of attended visual content rather than decoding arbitrary thought; others credit the company for an unusually honest non-invasive approach. These are opinion samples from replies, not independent verification.
To be verified: as of this issue, Alljoined's site offers no technical blog, paper or benchmark page for AIR-O, and no third-party reproduction appeared on X, so the capability can only be recorded as a company claim. Judgement: one of the most eye-catching research-flavoured launches of the window, but the mind-reading framing is running ahead of the evidence.
SID.ai releases SID-2, replacing single-model retrieval with compute-scaling swarms¶
At 02:27 Beijing time on November 18, SID.ai co-founder @maxrumpf released SID-2. According to the original post (about 82.5K views, 1,090 likes, 109 reposts, 122 replies, 867 bookmarks, 48 quotes), SID-2 is built on compute-scaling swarms, finds better results than Opus and Astra while being 8x faster and 20x cheaper, and is available through the API the same day.
Verified: the company site repeats the same marketing claims and states that a single agent reasons and searches in about 300ms, scales to as many as 24 coordinating agents, has a 256K context per agent, and is available through the API. The 8x, 20x and "better than Opus and Astra" figures are vendor claims; this issue could not check finer NDCG or per-task cost tables against a fully verifiable technical-report text, so those numbers are not presented as verified facts.
Judgement: "8x faster and 20x cheaper" is the headline on the site and launch post, and actual results will depend on the task and the number of agents; what can be confirmed is the product thesis — using parallel retrieval agents — not an independent finding that it beats frontier models across the board.
SenseTime open-sources SenseNova-Loop-9B: 9B parameters, 56 layer executions per pass, claims parity with Qwen3.8-27B¶
At 22:45 Beijing time on November 17, SenseTime's official account @SenseTime_AI announced the official open-source release of SenseNova-Loop-9B. The original post says the model has 9B parameters and is competitive with Qwen3.8-27B across multiple benchmarks, and lists gains from preview to release: NL2Repo 19.91 to 41.36, AutomationBench 14.33 to 36.67, GAIA 37.86 to 69.90. The post had about 59.3K views, 876 likes, 106 reposts, 35 replies, 705 bookmarks and 48 quotes.
Verified: the Hugging Face model card and the GitHub repository confirm an Apache 2.0 licence, about 9B total parameters with roughly 14.6B inference-equivalent parameters, 32 physical decoder layers of which 24 middle layers are executed twice with shared weights so one forward pass amounts to 56 layer executions, interleaved linear and full attention with two-stream mHC, and a context window of up to 512K tokens. The model card describes a low-memory deployment path using 4-bit weights and an 8-bit KV cache. The model card's own comparison is more granular: AutomationBench ties Qwen3.8-27B at 36.67, AIDABench is 65.31 versus 65.01, and NL2Repo is 41.36 versus 44.60, so the lead is not across the board.
What matters is the route chosen: instead of scaling parameters, it trades layer looping for computational depth and sells single-GPU deployability. Chinese-language discussion quickly turned to quantization and local deployment, including claims about the size of the 4-bit build and running long contexts on lower-memory GPUs, but most of that is individual testing; the model card only describes a low-memory path with 4-bit weights and an 8-bit KV cache, and a specific 15GiB figure was not verified there. Judgement: a 9B model approaching part of a 27B model's ability is a real option for small teams running locally, but "parity" holds only on some benchmarks.
2. AI Compute and Industry¶
NVIDIA fiscal Q3: $115.8B revenue, up 103% year on year, $107.6B data center, $130B guidance¶
After the US close on November 17, early morning of November 18 Beijing time, NVIDIA reported results for the third quarter of fiscal 2027, ended October 25. Verified: the official press release states revenue of $115.8 billion, up 103% year on year and 20% quarter on quarter; data center revenue of $107.6 billion, up 110% year on year and 21% quarter on quarter; GAAP gross margin of 74.8%; GAAP and non-GAAP diluted earnings per share of $2.70 and $2.68; and fourth-quarter guidance of $130.0 billion plus or minus 3%, assuming no data center compute revenue from China. The release also says Vera Rubin NVL72 has begun shipping, described as the fastest product ramp in company history, with racks at CoreWeave, Google Cloud, Microsoft Azure, Oracle Cloud Infrastructure and Nebius.
This earnings release had no official entry point on X. Within the window the @nvidia account only posted items such as an operations executive's award, with no earnings post, and the bot re-ran two zero-result checks, so spread came mainly from traders and summary accounts, for example SmallCapSnipa's summary post (about 5,697 views, 47 likes), engagement far below model-launch news. All figures follow the investor relations release; second-hand arithmetic in posts was not used.
The structure is what stands out: growth is still pulled by a single data center engine, and the outlook explicitly assumes no data center compute revenue from China. Judgement: beyond the year-on-year increase, the more useful information is in the guidance assumptions and the share of data center revenue.
Isomorphic Labs closes a $5 billion Series C co-led by Greenoaks and Sequoia¶
At 20:00 Beijing time on November 17, Isomorphic Labs' official account @IsomorphicLabs announced a $5 billion round. The company's post had about 106.4K views, 873 likes, 76 reposts, 38 replies, 130 bookmarks and 68 quotes; Demis Hassabis's post had about 531.2K views, 11.1K likes, 1,139 reposts, 498 replies, 1,351 bookmarks and 314 quotes, the widest-reaching funding item of the window.
Verified: the company announcement states a $5 billion Series C co-led by Greenoaks and Sequoia Capital, with Andreessen Horowitz and other new investors participating, continued significant investment from Alphabet, and participation from all existing investors: Thrive Capital, MGX, GV, CapitalG, Temasek and the UK Sovereign AI Fund. The money is to scale its proprietary drug design engine IsoDDE and its AI-designed candidate pipeline.
Where the capital went matters: this is the largest AI-related round in the window, and the target is drug design rather than a general-purpose model, with every existing investor following on. Judgement: as model capability gaps narrow, capital is moving toward AI that directly produces patentable assets.
3. Chinese AI Ecosystem¶
Kling releases Image 4.0: native 8K, 2,000 characters per image, up to 10 cohesive images¶
On November 17, Kuaishou's Kling released the image model Kling Image 4.0. Its press release says the model supports native 8K output, up to 2,000 characters of text in a single image, up to 10 style-consistent images from one prompt, more than 70 preset fonts, and API access for enterprises and developers. The official account @Kling_ai posted a demo at 00:39 Beijing time on November 18 with about 3,873 views, 50 likes, 4 reposts, 16 replies, 10 bookmarks and 3 quotes.
Engagement is low, a typical official launch that has not yet built high interaction. Chinese-language discussion centred on feature explainers: @GeekCatX noted in an earlier teaser post (about 745 views) that the official poster offered only three adjectives and no launch date, and post-launch posts were mostly feature introductions and template ideas rather than scaled-up testing. The press release cites the Artificial Analysis image leaderboard to say the preview led on text-to-image and image editing and ranked third in UIUX design scenarios for text-to-image; that is a vendor citing a third-party leaderboard, not an independent evaluation.
Judgement: native 8K and long-text layout target specific pain points in commercial design workflows, and explain the direction of this update better than raw specs; engagement and third-party validation have not caught up.
A same-prompt comparison of Doubao Work and WorkBuddy: same model, same prompt, bills differing by up to 37x¶
At 13:36 Beijing time on November 17, Chinese-language account @AYi_AInotes posted a comparison of two domestic office agents. The original post, with about 126.5K views, 150 likes, 8 reposts, 39 replies, 65 bookmarks and 17 quotes, says that on tasks such as analysing order tables and building web pages, running the same model with the same prompt in Doubao Work and WorkBuddy produced a maximum bill difference of 37x across 34 runs on both sides, with the widest gap in the Kimi K3.1 group.
This is an individual test, not an official pricing conclusion: billing involves each platform's plan multipliers, caching and rate limits, so a bill difference cannot be converted directly into a model price difference, and the post does not publish the full per-group data. Reposts and discussion simplified it into "Doubao is cheaper" without auditable detail.
Judgement: the value of such comparisons is not the ranking but the reminder that, for domestic office agents, platform billing rules rather than model capability often decide the choice.
Coverage gaps for this issue: the November 16 daily report is missing, so history de-duplication covers one day less; NVIDIA's earnings had no official X post in the window, so figures follow the investor relations release; Alljoined's AIR-O has no independent technical material and remains a company claim.