X AI Daily¶
Translation
English translation of the Chinese original.
1. AI Models, Products and Applications¶
Gemini 4 Hydrogen released: 2.5x faster than 3.8 Flash, official confirmation of $1.50 input and $7.50 output per million tokens¶
At 02:02 Beijing time on November 13, Google released Gemini 4 Hydrogen on X, positioning it as the fast execution tier of the Gemini 4 family, alongside Carbon and Argon, which handle long-horizon reasoning. The hard numbers from the official blog are: it completes tasks 2.5x faster than 3.8 Flash at the default high thinking level; Terminal-Bench 4.1 rises from 18.2% to 57%; the Artificial Analysis Intelligence Index is 51 at the default high level, above Argon's 50, and reaches 54 at the highest xhigh level; there are four thinking levels — low, medium, high (default), xhigh; 1 million token input context with up to 1 million token output, and the price does not change across the whole context, at $1.50 per million input tokens and $7.50 per million output tokens. Availability covers all AI Studio tiers, the Gemini API, Gemini for business, the Gemini app, Google Search, and paid Pro and Ultra subscribers in Antigravity; on the enterprise side it also includes Gemini in Sheets under some commercial subscriptions.
The same day brought another mechanism: Antigravity launched Orchestrator, which switches between Carbon and Hydrogen on demand within the same session; the official claim is that on Terminal-Bench 4.0 it comes close to Carbon's task success rate while using 24% less quota on average and running 19% faster. The blog also gives a set of safety figures: up to 81% fewer out-of-scope actions in agentic tasks, up to 79% less under-service, 12x fewer misaligned behaviors and shortcuts in tool-calling and coding evaluations, 96.7% on internal honesty evaluations, and chain-of-thought readability maintained at 99.9% across 200,000 audit turns.
Why it matters: the competition at this tier is no longer about being "the strongest" but about being "good enough and cheap" — 2.5x speed, four thinking levels, a million-token context with no price premium; together those three numbers squeeze on the "cost per token" dimension. Hours before the launch, tracking accounts on X had already reported that the Vertex endpoint could call gemini-4-hydrogen and that the billing catalog showed a $1.50/7.50 SKU (e.g. @aitrackerbot, 73k views), and the official blog then gave the same price set.
Key original posts: @GoogleDeepMind original post, 727k views, 5447 likes, 469 reposts; @OfficialLoganK original post, 572k views, 8238 likes, 594 reposts; @sundarpichai original post, 350k views, 6281 likes, 362 reposts; @GeminiApp original post, 339k views, 4439 likes, 425 reposts; @demishassabis original post, 84k views, 2561 likes, 166 reposts. The same account's WebGL water and glass demo is in this post, 24k views, 327 likes.
Verification: Google's official blog (November 12, 2026) confirms item by item the 2.5x speed, 51 versus 50 and the xhigh level's 54, the four thinking levels, the million-token context and the $1.50/7.50 price, Terminal-Bench 4.1's 57% versus 18.2%, Orchestrator's 24% and 19%, and all availability channels, and includes a model card PDF; 9to5Google's report confirms that this time the Flash naming is no longer used. Unverified: the two scores of 51 and 54 are Google's citation of Artificial Analysis readings, and this time that organization's page was not opened to check them one by one; all the safety metrics above are officially self-reported, with no third-party reproduction.
Assessment: this looks more like pushing Gemini 4 capabilities down into the "everyday tier" than another capability leap — the real change is the flat price across a million-token context and the cost control that four thinking levels bring, and the speed figure is more credible than leaderboard comparisons like "beating Argon".
Follow-up: Claude Meetings enters beta, supporting only Google Meet, with free notes and artifacts during beta¶
At 04:16 Beijing time on November 13, the official Claude account announced that Claude Meetings is beginning to roll out in beta to Team and Enterprise plans. The 11-11 daily already covered its preview, so here we only add the new information from the beta stage.
In form, Claude joins video meetings as a visible participant, answers in text by default, and can call connected tools with the person's authorization; during the meeting it can draft documents, whiteboards and other artifacts on the spot, teammates can edit them together in the meeting, and before the meeting ends it can be asked to send a recap; after the meeting the notes, action items, documents and whiteboards all stay on that meeting's own page on claude.ai, where you can look back, ask questions, or add Claude to the next meeting from the calendar. At the platform level only Google Meet is supported for now, with other platforms listed as "soon". On permissions, the Team version allows inviting Claude to meetings by default and admins can turn it off; the Enterprise version is enabled by admins for all or part of the organization. During beta it is unavailable to organizations using zero data retention (ZDR), HIPAA configurations, or customer-managed encryption keys (CMEK). On billing, during beta notes and transcripts are free and in-meeting Q&A is free, but answers that use connected tools are billed against plan usage; documents, whiteboards and other artifacts created in the meeting are free until February 10, 2027. External guests are handled by default by not joining meetings that have external guests, and leaving automatically if an external guest joins midway; admins can relax this, and once relaxed Claude only posts a notice in the meeting chat rather than answering in it, while external guests' speech still enters the notes and transcript. On data, transcripts, summaries and screen descriptions are saved, but audio is not. On ecosystem, the company says it is working with Gong, Attention, Granola, Fireflies.ai and Otter so that meeting transcripts can flow into the tools teams already use.
Why it matters: the hardest part of putting AI into meetings is not generation quality but the permission and compliance boundary. This time the company wrote "does not join meetings with outsiders by default", "does not store audio", and "unavailable to ZDR/HIPAA/CMEK organizations" as defaults, which shows that whether this kind of feature can ship in enterprises depends first on data-processing terms; the free period runs only to February 10, 2027, which is a clear commercial trial window.
Key original posts: @claudeai original post, 2.068 million views, 37.6k likes, 2086 reposts, 1068 replies, the most-engaged post in this window; @testingcatalog original post, 9281 views, 211 likes.
Verification: Claude's official article confirms the platform limit, the plan scope, the two sets of permission defaults for Team and Enterprise, the ZDR/HIPAA/CMEK exclusion, the free scope and the February 10, 2027 cutoff, the default external-guest rule, saving only transcripts and summaries rather than audio, and the list of five partners; setup steps are in the help center (the body was not opened this time). TechCrunch's November 10 report described the same feature's "leaves when an external guest joins" before launch, and is a preview. Unverified: the official article page itself carries no date, so the specific rollout cadence is taken from the official post.
Assessment: the information in this item is in the terms, not the features — not joining meetings with external guests by default, not storing audio, and directly excluding organizations with three types of compliance configurations shows that Anthropic treats enterprise compliance as an entry requirement rather than an add-on; how it is priced after the free period ends on February 10, 2027 is the real watershed for this feature.
OpenAI announces two cyber products at "Cyber at the Frontier"; an undisclosed cybersecurity model reportedly scores 77% on ExploitGym¶
At the Cyber at the Frontier livestream in the early hours of November 13 Beijing time, OpenAI announced two products aimed at defenders. Daybreak Red Team is a set of security agents: one agent orchestrates the whole evaluation, several read source code and application context, and finally a review agent rechecks the results, used to judge whether a given vulnerability is actually exploitable and thereby prioritize fixes. Daybreak Defense Factory sits on top of Codex Security, with agents continuously scanning as code changes, reviewing pull requests, triaging alerts, handing findings to the engineer responsible for that code, and providing a patch in Codex — the output is a mergeable pull request with tests. The livestream said both are in early access: Red Team goes first to a small number of customers with a waitlist open, Defense Factory is available through Codex's Daybreak plugin, and neither has a published price.
The livestream also included a model disclosure with no name: an unreleased cybersecurity model reportedly scores 77.0% on ExploitGym, with the comparison on the same page being GPT-6 Cyber released on November 5 at 52.6%, Gemini 4 Carbon at 54.8%, and a comparison model on Anthropic's side at 39.2% (the retold version on X writes it as Mythos 5.2). The session also gave PatchBench at about 80%, and in a honeypot evaluation a stay-within-bounds rate of 99.9% (without production-side protection) and 100% (with protection).
Why it matters: the significance of these two products is turning "red teaming" and "fixing" into two sellable process steps rather than releasing another model — Red Team answers "is this vulnerability worth fixing", Defense Factory answers "who fixes it and where is the patch". As for the 77% figure, it has no model name, no release date and no price, so at this stage it can only be treated as a reading on a launch-event slide.
Key original posts: @gdb original post, 28k views, 251 likes, 23 reposts, 69 replies; @MTSlive live retelling, 11.8k views, 218 likes; @SPAC89 number retelling, 6479 views, 18 likes; @MartinSzerment original post, 237 views, explicitly saying there is no name, date or price; partner @semgrep's original post, 528 views.
Verification: OpenAI's official product pages Daybreak Red Team and Daybreak Defense Factory confirm the positioning of the two products (the former uses a set of agents to penetration-test web applications, validate vulnerabilities and check whether patches close the exploit; the latter turns security findings into patches that engineers can review and test); The Register's November 12 report confirms that Red Team is in early access and that Defense Factory currently uses GPT-6 Cyber. Live details were cross-checked with reports: The Deep View confirms the division of labor among Red Team's agents, Defense Factory being built on Codex Security, Red Team's early access plus waitlist, Defense Factory going through Codex's Daybreak plugin, and that no pricing was announced; StartupHub's live notes record the speaker giving GPT-6 Cyber at 52% on ExploitGym, Mythos 5.2 at about 38%, "the next model nearly doubling", and PatchBench at about 80%. The specific value of 77.0% was seen this time only in retellings of livestream screenshots on X; it appears on no official page and on no public leaderboard. One more difference in framing to note: when OpenAI released GPT-6 Cyber it published ExploitGym 48.5% (GPT-5.6 Cyber 32.1%, see The Register's report), while a third-party roundup has Gemini 4 Carbon at 53.5% (6 hours) and 54.8% (10 hours) (BenchLM roundup) — the livestream page's 52.6% and the release-time 48.5% are not the same set of readings, and the two should not be mixed when retelling.
Assessment: the deliverable information is that the two products turn red teaming and remediation into a product line, with price and scale still unpublished; 77% should be read as "a vendor's self-reported number on its own evaluation setup", not a reproducible third-party result.
Follow-up: Inkling 2 Small weights released on Hugging Face under Apache 2.0, with a 500GB+ VRAM bar to run¶
At 04:50 Beijing time on November 13, Thinking Machines announced that the weights of Inkling 2 Small are available for download on Hugging Face under the Apache 2.0 license, with the caption "make Inkling your own". The model itself was released on November 5, when it was hosted only via Tinker and the weights were not open; the official line was that full weights would be released "next week" after external safety testing and red-teaming — delivered on November 12. The model release itself was already covered in the 11-07 daily (outside the 3-day dedup window, so it is still handled as a follow-up here); here we only add the open-weights layer.
The new content centers on "how it can run": the Hugging Face model page states 594B total parameters, 17B active, a 48-layer decoder structure, 384 experts with 8 routed per token, text plus image input, and a 1M token context; the weights come in two precisions, with NVFP4 as the primary recommendation requiring 500GB+ of aggregate VRAM (2× B300 or 4× H200), and MXFP8 requiring 800GB+ (4× B300 or 8× H200); the official claim is that MXFP8 is only 1–2 percentage points above NVFP4 and that adding hardware for it is not recommended; for inference there are four recipes — SGLang, vLLM, TokenSpeed, Unsloth — post-training supports Miles, SkyRL and Snowflake Cortex Training, and hosted channels include NVIDIA, Fireworks, Together, Baseten, Nebius, Modal and Cloudflare.
Why it matters: Apache 2.0 plus downloadable weights, permitting commercial use and redistribution, is the most substantive delivery on the open-weights side this week. But "Small" means two B300s at minimum in hardware terms, so the individual-developer path is essentially closed, and the real beneficiaries are teams and hosting providers with GPU clusters.
Key original posts: @thinkymachines original post, 129k views, 1140 likes, 186 reposts; @ariG23498's technical breakdown, 1411 views, 29 likes; @Letta_AI's evaluation addendum, 365 views, 6 likes, saying it scores 60.0% on their FrontierMemory.
Verification: the Hugging Face model page confirms that the weights are downloadable, the license field is apache-2.0, the 594B and 17B parameter counts, the 1M context, the VRAM requirements for both precisions, the four inference recipes, the three post-training frameworks and the seven hosting channels; the November 12 update on the official introduction page confirms that the weights are live and the full integration list. One framing issue to flag: the body of the official introduction page still keeps the November 5 sentence "release full weights under Apache 2.0 next week", not rewritten with the November 12 update, so the license should be taken from the Hugging Face field and the official post. Unverified: Letta's 60.0% and "about $19 per run" are third-party self-tests, and no corresponding evaluation page was found this time.
Assessment: this is a complete case of running the "closed hosted first, open weights after external safety testing" playbook end to end, and can serve as a reference for later open-weights releases; for users, the bar has shifted from the license to VRAM.
World Labs opens Atlas to everyone: web app plus API, delivering images, video and 3D Gaussian splats¶
At 02:09 Beijing time on November 13, World Labs announced that Atlas is open to all users. The official post's wording is: hand an idea to Atlas and what comes back is keyframes, video clips or 3D Gaussian splats. The official site also gives entry points: two web tools, where Viewfinder reconstructs scenes from photos or generated images, plans camera paths and renders video, and Chisel is used to roughly block out scene structure to guide the results; on the developer side there is an API endpoint and documentation that can convert splats into meshes with real geometry for import into games, simulation and existing 3D tools, including a Blender integration. Fei-Fei Li posted at almost the same time, writing "Atlas is now open to everyone".
Why it matters: the point is the form of the deliverable — the same model gives 2D images, video and an explicit 3D representation at once, which means it can plug into existing film, architecture and robotics workflows rather than just generating an uneditable video clip; opening the web app and the API at the same time also gives "world models" a billable call surface for the first time.
Key original posts: @theworldlabs original post, 126k views, 1272 likes, 161 reposts; @drfeifei original post, 47k views, 1069 likes, 125 reposts.
Verification: World Labs' official /atlas page (November 12, 2026) confirms the two web tools Viewfinder and Chisel, the API and documentation entry points, "converting splats into meshes", the Blender integration, and the three use categories of film, architecture and robotics; the technical article confirms that Atlas generates 2D frames and depth maps together, can generate new viewpoints and geometry from a single image, and can predict depth frame by frame from video and fuse it into a 3D reconstruction, with outputs explicitly as point clouds or 3D Gaussian splats. Unverified: the word "keyframes" appears only in the official X post, while the technical article uses 2D images and video; pricing, free quota and commercial terms were not stated this time.
Assessment: what is most practical for creators is the two output routes, splats and mesh, rather than how much the resolution went up; but whether "open to everyone" includes a free quota and how the API is priced is not written on the official page, so real usability has to wait for prices to be published.
Hot on X, facts unverified: GPT-6.2 Astra said to appear in the Codex repository¶
Since daytime on November 12, multiple posts on X have claimed that OpenAI's open-source Codex repository contains migration records from gpt-6-astra to gpt-6.2-astra, and treated that config as evidence that "6.2 is about to launch". The most widely spread one was written at 18:05 Beijing time on November 12, with 66k views; there are several versions in Chinese-language circles, and other posts circulate unconfirmed prices, as well as product plans such as "6.2 Astra targets Fable 5.2, 6.2 Sol targets Opus 5.5". As of now, OpenAI has not released anything related to 6.2.
Why it matters: this kind of "string in a repository" is the clue most easily taken as evidence in this cycle — it looks like a verifiable technical trace, but a config string can only prove that an internal name exists, not that a model is usable, callable or priceable. And in this round of spread, "already live" and "pricing" are exactly the two things with no first-hand basis at all.
Key original posts: @dummerspast39 original post, 66k views, 394 likes, 7 reposts, 47 replies; @AI_Screening original post, 10.8k views, 203 likes, 6 reposts; @buildwithrajath's comparison claim, 15.7k views, 162 likes, 5 reposts.
Verification: I ran three checks on github.com/openai/codex. First, in GitHub's commit search (repo openai/codex, keyword astra) there were 19 results in total, concentrated between September and October 2026, covering adding GPT-6-Astra to the model catalog and selector, adding it to the Amazon Bedrock catalog, and the TUI stardust effect — none involving 6.2. Second, searches for "6.2 astra" and "hydrogen" both returned 0 results, and the 3 hits for "gpt-6.2" are unrelated to the claim. Third, the model catalog models.json currently shipped on the repository's main line lists gpt-5.5, gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna, gpt-6-astra, gpt-6-sol, gpt-6-luna, gpt-6.1-sol, gpt-6.1-terra, gpt-daybreak-blue-latest and gpt-daybreak-red-latest; "6.2" does not appear anywhere in the file, nor is there a migration entry from gpt-6-astra to 6.2. Third-party roundups also point the observed location to a Codex usage-stats screenshot rather than a repository commit (CellCog). Conclusion: the migration-record claim is unverified this time, "already live" and "pricing" are unverified, and OpenAI has made no official announcement.
Assessment: the only thing worth believing is that "the 6.2 naming clue exists inside OpenAI", and the source is a screenshot; writing it as "about to launch" is extrapolation, and writing it as "already live and priced" has crossed past the facts.
2. AI Compute and Industry¶
Follow-up: Anthropic posts officially about its IPO; Bloomberg says subscriptions are about six times the shares available¶
At 19:30 Beijing time on November 12, the official Anthropic account posted "Our initial public offering", with a link to the IPO page and a legal disclaimer (a registration statement has been filed with the SEC, and investors should read the prospectus before investing). The 11-11 daily covered the roadshow, so here we only write what is new: Bloomberg reported on November 12 that the offering has drawn subscriptions about six times the shares available, with large anchor investor orders the main source; the report also said underwriters will favor existing shareholders, that some buyers will receive less than expected, and that these numbers may still change. Reuters, retelling the same report, gives: a valuation of up to $1.4 trillion, 825 million shares planned, a price range of $70 to $78, up to $64.4 billion raised, and a November 19 Nasdaq listing under ANT. Another report retelling Bloomberg says the IPO plus a concurrent private placement could raise up to $90.4 billion in total.
Why it matters: the six-times figure is far more concrete than "multiple times oversubscribed", but what matters more in the same batch of reports is allocation priority — favoring existing shareholders means incoming money gets the remaining shares, which directly affects the free float and volatility on the first trading day.
Key original posts: @AnthropicAI original post, 2.429 million views, 9167 likes, 892 reposts, 858 replies; @BayAreaTimes' "multiple times oversubscribed" post, 327 views, 2 likes.
Verification: Bloomberg's original report could not be opened this time due to the paywall; what is readable is Reuters' retelling, which confirms the six-times subscription, the anchor orders, the priority for existing shareholders, and the items of the $1.4 trillion valuation cap, 825 million shares, $70 to $78, $64.4 billion and the November 19 ANT listing; Anthropic's official post and Newsroom confirm that an S-1 has been filed with the SEC. Difference in framing: the $90.4 billion comes from another report retelling Bloomberg, which counts the IPO and the concurrent private placement together, a different basis from Reuters' $64.4 billion (the IPO portion), and the two should not be mixed. Unverified: the arrangement in the BayAreaTimes post that "Amodei and JPMorgan's Jamie Dimon appear in a livestream together on Friday" was not confirmed in other sources this time.
Assessment: the subscription multiple and allocation priority indicate an offering with excess demand, but "six times" is a book-building-stage figure and can change at any time; the real test is the pricing and first-day performance after the November 19 listing.
3. Chinese AI Ecosystem¶
Bloomberg reports Zhang Yiming's internal framing: catching up with OpenAI and Anthropic at the earliest in 2029 to 2030, with pretraining data the bottleneck¶
On November 11, Bloomberg reported that ByteDance plans to launch an AI assistant similar to Meta Muse in the coming weeks; the report said Zhang Yiming stated at an earlier all-hands AI team meeting that ByteDance is capable of competing with Anthropic and OpenAI at the frontier, but that truly catching up would be "at the earliest 2029 or 2030". The same report also mentioned that the pretraining dataset for the Seed language model is "several orders of magnitude smaller" than that of leading rivals, and that ByteDance does not plan to rely on distillation. Chinese-language retellings often merge these points with "$70 billion" into a single item: spending up to $70 billion to grind away at foundation models and video AI.
Why it matters: setting the catch-up date at 2029 to 2030 with the reason resting on data rather than money is the hardest sentence in this report — the scale gap in pretraining corpora cannot be erased directly by budget. In terms of spread, the heat on the Chinese side centers on the emotional point of "admitting the gap": the post at 04:07 Beijing time on November 12 had 135k views, and another interpretation published at 11:01 on November 12 had 5928 views, both repeatedly quoting the phrase "several orders of magnitude smaller".
Key original posts: @NFT_Chen original post, 135k views, 166 likes, 19 reposts; @0xLogicrw original post, 5928 views, 15 likes, 2 reposts; @ZSS_HK's report retelling about the new assistant, 380 views, published at 01:37 Beijing time on November 13.
Verification: Bloomberg's November 11 report could not be opened this time due to the paywall; what is readable is 36Kr's reposted Chinese long-form article, which confirms the 2029 to 2030 timetable, the "several orders of magnitude smaller" pretraining data, not going the distillation route, and launching a Muse-like assistant within weeks. The origin of the $70 billion needs to be stated separately: that number comes from a different Bloomberg report on May 27, 2026 (ByteDance Weighs Capex of as Much as $70 Billion in AI Push, whose readable reposted version states up to $70 billion of capex for 2026, funded in part by about $50 billion of 2025 profit, and that the number itself is still preliminary); it is not the same report as the November 11 one, and it is not "a newly approved budget to catch up with rivals". Unverified: the specific occasion of the all-hands meeting (Bloomberg wrote "the most recent AI team all-hands", while Chinese retellings generally write "the Seed team all-hands"), the codename of the new assistant (some retellings say "Spell"), and the direct link between "$70 billion" and this statement.
Assessment: what actually carries information here is "data", not "money" — the catch-up date is pushed to 2029 to 2030 with the reason written on pretraining data being orders of magnitude behind; the $70 billion is old news spliced in, and should be read separately from this story.
Doubao Work brings in Kimi K3.1 Pro, GLM 5.3 and DeepSeek V4.1 Flash, and marks quota consumption multipliers on the selection page¶
On November 12, Doubao officially announced an update to its office agent product "Doubao Work": in addition to its in-house Doubao 2.1 Lite and 2.1 Pro, it brings in a first batch of three third-party models — Kimi K3.1 Pro, GLM 5.3 and DeepSeek V4.1 Flash — which users can switch between according to the task. It also launched "quota consumption multipliers" — the model selection page directly marks each model's pricing coefficient; the lower the multiplier, the less quota the same task consumes; the first item in the list is still "Auto (small tasks free)", where the system matches the model to the task scale automatically, and users can manually switch to Pro or third-party models as task complexity rises. The limited-time offer has two layers: using participating models in designated time slots on weekdays and holidays gets a quota discount; the multipliers for Doubao 2.1 Lite and DeepSeek V4.1 Flash were lowered directly. Other retellings say the multipliers marked on the official announcement page are Doubao 2.1 Pro 0.65x, GLM 5.3 0.80x and Kimi K3.1 Pro 0.95x, with Doubao 2.1 Lite and DeepSeek V4.1 Flash marked 0.01x during the limited-time promotion. This path is not new: on October 29, Doubao Work had already opened third-party model configuration to Team and Enterprise subscription editions, with usage deducted directly from existing AI credits.
Why it matters: putting rivals' models on its own shelf while marking "who is more expensive" plainly on the interface amounts to admitting that the model layer is becoming replaceable supply, and that the office entry point is the position to defend. Another overlooked detail is that it previously opened this to enterprise customers first and only now extended it to ordinary users — validating enterprise willingness to pay first, then scaling up.
Key original posts: @AYi_AInotes long commentary, 11k views, 83 likes, 7 reposts, 48 replies; @slime2078 information roundup, 1303 views, 23 likes; @Willie95163 multiplier and roadmap comparison, 162 views. Also @xiaoanlab original post and @greek2240 original post give roundups of the multiplier values.
Verification: ITHome, Phoenix Tech and Sina Tech (a Kuaikeji piece) are mutually consistent in retelling the official announcement (The Paper's version of the same report was only readable as a summary this time), confirming the three third-party models, the multiplier mechanism, the two layers of the limited-time offer, the already-lowered multipliers for Lite and DeepSeek V4.1 Flash, "small tasks still free", and the October 29 early opening to the enterprise edition; the 30 days of subscription benefits for new users signing in on desktop comes from Kuaikeji and retellings on X. Unverified: the specific multiplier values (0.65x, 0.80x, 0.95x, 0.01x) appear only in retellings of the official announcement page (such as a SMZDM community article and the two X posts above), and Doubao Work's official announcement page could not be opened this time to confirm them word for word.
Assessment: the substance of this update is not "three more models" but turning model choice into an open arithmetic problem — once multipliers are printed on the selection page, models with poor cost-performance get skipped by users on that same page; and 0.01x is only a limited-time promotion, so long-term cost still depends on the multiplier itself.