Blog · AI Agent News · Safety & policy VIEN
AI News Nov 18: X Flags "llm_slop", GPT-6.2 Cuts API 20%
From Gemini Tasks, Apple FastFrames at 35.4 ms, GPT-6.2, and AI Studio to X's "llm_slop" code rule: five Nov 17–18, 2026 AI updates and what they cost you.

In this article
- 1. Gemini Tasks goes free in six countries with calls and payments
- 2. Apple FastFrames 14B & 1.3B hits 35.4 ms, yet demands 44.8 GiB VRAM
- 3. GPT-6.2 Astra & Sol cut Astra API prices 20%, stall on coding
- 4. Google AI Studio adds custom domains and Android builds—with a catch
- 5. X's "llm_slop" code rule: 7-day throttle or 30-day shadowban?
- My take: Read the fine print before chasing AI speed
Across November 17–18, 2026, AI labs pushed harder on autonomous agents, fast video generation, and lower API pricing, while quietly tightening hardware requirements, billing traps, and spam filters. Below are four major product and model releases before we unpack X's open-source llm_slop enforcement rule at the end.

1. Gemini Tasks goes free in six countries with calls and payments
On November 17, 2026, Google Labs VP Josh Woodward and Gemini App on X rolled out Tasks free for eligible adults in the United States, Japan, South Korea, Canada, Australia, and India. Switching the home-screen toggle from Chat to Task runs multi-step background workflows powered by Gemini 4 Hydrogen.
In the US, users can also forward tasks to [email protected] or text Gemini over Google Messages and iMessage, according to Google Blog updates. Tasks signs into portals via Google Password Manager, checks out with Google Pay under user spending caps through AP2, and calls local businesses (except in five US states) inside hardware-isolated VMs per account. Users outside those six launch nations, including Vietnam, still have to wait for free access.
2. Apple FastFrames 14B & 1.3B hits 35.4 ms, yet demands 44.8 GiB VRAM
Apple researchers led by Gene Chou published open weights for FastFrames-L (14B) and FastFrames-S (1.3B) on November 17–18, 2026 via the FastFrames project page and shared by @_akhaliq. On a single NVIDIA B200 GPU, interactive video synthesis runs at 35.4 ms per step.
On that same B200 GPU, FastFrames runs 8–20x faster than Causal Forcing++ (272 ms) and Self-Forcing (717 ms), driven by a 0.9 ms pixel decoder and 2.2 ms KV-cache refinement. Before trying to run it locally, keep the hardware floor in mind: the smaller 1.3B demo requires Linux x86-64, CUDA 12.8, and peaks at 44.8 GiB VRAM, well beyond typical desktop GPUs as noted in our FastFrames deep dive.
| Nov 17–18, 2026 Story | Headline Number | Practical Catch for Users |
|---|---|---|
| Gemini Tasks (Google) | Free in 6 launch nations | Calling, email/SMS channels, and checkout stay US-only at launch |
| Apple FastFrames 14B & 1.3B | 35.4 ms/step on 1 B200 | Even the 1.3B demo peaks at 44.8 GiB VRAM on Linux + CUDA 12.8 |
| GPT-6.2 Astra & Sol | Astra API -20% ($8/$40) | Coding scores stalled; hard-question guess rate hits 55–59% |
| Google AI Studio | Domains from $12/year | Irreversible Cloud Billing upgrade; Play publishing US-only |
X llm_slop_user rule | 7-day SpamHighRecall | Scoring classifier stays closed-source; triggers remain opaque |
Also readMeta Muse Signal & Bank Balance Report: Fact vs. Fiction3. GPT-6.2 Astra & Sol cut Astra API prices 20%, stall on coding
OpenAI released GPT-6.2 Astra for specialized work and GPT-6.2 Sol for everyday tasks on November 17, 2026 across ChatGPT Work, Codex, and the API, according to The Decoder. Astra's API pricing drops 20% to $8 per million input tokens, $0.40 cached, and $40 output, while Sol stays at $2 input and $10 output.
Sam Altman highlighted on @sama that $200/month Pro subscribers now get Ultrafast mode on GPT-6.2 Sol—up to 4.5x faster generation at 6x quota burn—while acknowledging both models still need polish to feel "nice to work with." Independent Artificial Analysis tests show coding scores flatlined (Astra unchanged; Sol down 3 points vs GPT-6.1 Sol), and both hallucinated on 55–59% of tough AA-Omniscience questions at max effort.

4. Google AI Studio adds custom domains and Android builds—with a catch
Google AI product lead Logan Kilpatrick announced on November 18, 2026 via @OfficialLoganK that AI Studio now supports buying custom domains inside the Publish panel and building full-stack Android apps with Firebase. The Starter Tier remains free with *.ai.studio subdomains and a browser-based 1200p Android emulator.
Look closely at AI Studio Build mode docs before clicking Publish with your own domain, though. Attaching a custom domain requires upgrading the Google Cloud project to paid billing—an irreversible one-way switch off the free Starter Tier—plus non-refundable domain fees from $12/year and Cloud DNS from $0.20/month, while personal Google Play publishing is US-only at launch.
5. X's "llm_slop" code rule: 7-day throttle or 30-day shadowban?
On November 17–18, 2026, warnings from @obchakevich_ sparked fears that AI-written replies trigger hidden llm_slop_user penalties on X. Inspecting xai-org/x-algorithm (lib.rs) confirms the automated penalty exists, though the mirrored duration is shorter than many threads claim.
In enforcement_user.yaml, rule act_add_llm_slop_label checks for "llm_slop_user" in score.labels and attaches SpamHighRecall with ttl_msec: 604800000 (7 days, lowered in October 2026 from an August snapshot of 30 days). While active, existing followers still see your posts in Following, but OON_USER_LABEL_DROPS drops them from out-of-network discovery feeds like For You without extending the timer on repeat hits.
My take: Read the fine print before chasing AI speed
Across all five Nov 17–18, 2026 stories, the real story is not raw benchmark speed—it is the fine print attached to every upgrade. Gemini Tasks and AI Studio offer generous free entry points that quickly hit regional gates or one-way billing switches; GPT-6.2 Astra cuts API fees by 20% while still guessing on tough facts; and FastFrames hits 35.4 ms only if you have 44.8 GiB of server VRAM. For creators and indie builders, X's llm_slop rule is the clearest warning of all: saving two minutes by automating replies is a bad trade if it quietly locks you out of For You reach for a week.
Sources
- Gemini App on X — Tasks rollout X @GeminiApp
- Google Blog — Gemini App updates Google Blog
- X @_akhaliq — Apple FastFrames-L 14B & FastFrames-S 1.3B X @_akhaliq
- Gene Chou (Apple AIML) — FastFrames genechou.com
- The Decoder — OpenAI launches GPT-6.2 Astra and Sol The Decoder
- X @sama — Sam Altman on GPT-6.2 Sol Ultrafast X @sama
- X @OfficialLoganK — Google AI Studio custom domains & Android X @OfficialLoganK
- Google AI Studio Docs — Build mode & Custom domains Google AI Studio Docs
- xai-org/x-algorithm — lib.rs GitHub xai-org/x-algorithm
- xai-org/x-algorithm — enforcement_user.yaml GitHub xai-org/x-algorithm
- X @obchakevich_ — cảnh báo nhãn llm_slop_user X @obchakevich_
Safety & policyGoogle GeminiOpenAI & ChatGPT
Safety & policy · Nov 21, 2026Nvidia Without Jensen Huang: Where Do CUDA and GPUs Go?Analyzing Nvidia after Jensen Huang: the 60-report leadership gap, Claude Mythos and AMD pressure on CUDA, and what GPU buyers face next.
Safety & policy · Nov 21, 2026Apple Music Eats 8–11 GB With Zero Downloads: How to Fix ItStreaming only, yet Apple Music takes up 8–11.5 GB on your iPhone? See why Lossless caching bloats Documents & Data and 4…
Coding agents · Nov 21, 2026Apple's Computer-Use AI Agent Listing and What It Means for MacAnalysis of Apple's proactive AI agent job listing, how it compares with Gemini on Mac, and macOS 27.3 permission splits for Mac…