Gemini tiers tighten, OpenAI details a distillation campaign: what the dossier actually says
On 4 October 2026 OpenAI published its own account of a campaign to extract protected reasoning from its models, and The Decoder mapped Google's new Gemini tiers. The dossier contains no dated reporting on an AI weather forecast model.
The topic query asks about AI weather forecast models. The dossier does not support a news story on that subject. Its newest item is dated 4 October 2026, and it is Ideogram 4.5, an image editing model.
The rest of the material covers subscription tiers, security research and model design. So the honest piece here is about what those sources establish, and what they leave open. None of it is weather.
What changed this week
Google is tightening which Gemini models people can reach without paying. 9to5Google reported on 3 October that, per an updated support document, users without a Google AI subscription will be limited to Flash-Lite. From 9 October they lose access to 3.6 Flash and 3.1 Pro. AI Plus subscribers at $4.99 a month drop to Flash-Lite and Flash.
The Decoder put the same change in a table on 4 October. No plan gets Flash-Lite only, AI Plus adds Flash, and Pro requires AI Pro at $19.99 a month or AI Ultra at $99.99 and $199.99. The Decoder also argues the practical effect is small, because most casual users do not know which model they are running.
Two things stand out. First, the free tier is being cut, not reshuffled: 9to5Google says free users currently get 3.6 Flash and varying access to 3.1 Pro. Second, the paid tiers gain something. 9to5Google reports that AI Pro subscribers will get the Deep Think option for maximum parallel reasoning, previously limited to the $99.99 and $199.99 plans. The same report notes new low, medium and high effort levels for each available model, which would use more of a user's limit.
The Decoder: "Google may also be setting the stage for the larger Gemini 4 Argon, which could be pricier to run."
The timing of Argon is unsettled. 9to5Google says Google announced it is first coming to AI Ultra subscribers and has not said whether Argon counts as a Pro model or a tier above it. That gap matters for anyone deciding whether $19.99 still buys the top model. Neither outlet says it does.
OpenAI moves on distribution, and on distillation
On 29 September, OpenAI's Dev Day announcements pointed at the app store model, according to TechCrunch's Sarah Perez. ChatGPT will suggest apps inside a conversation when one could help with the task, and an expanded plug-in architecture now supports extensions that render interactive panels. Perez reports 16 launch partners for Sign in with ChatGPT, among them Cognition's Devin, Notion, Vercel, T3, OpenClaw and Dactyl. The chatbot, per OpenAI, has 1.2 billion weekly users.
Separately, OpenAI published its own account on 4 October of a campaign to extract protected reasoning from its models. The company says the earliest observed activity was in the first week of July, with high-volume spikes on 24 and 25 July of 16,000 requests using an extraction pattern from over 4,000 users. OpenAI says it fully disrupted a related cluster of more than 15,000 users by 28 July, and that it attributes a core cluster to individuals associated with Moonshot AI, the developer of Kimi.
OpenAI is explicit that no encryption was broken and no stored conversations were accessed directly. Operators manipulated model interactions instead, including copying encrypted reasoning between conversations and asking a model elsewhere to transcribe it. The company also credits independent researchers who reported related cross-model and conversation-compaction issues through responsible disclosure. It is unclear, OpenAI says, whether all operators came from a single actor.
Smaller models, narrower outputs
Jerry Gamblin ran Cloudflare's Clef models locally on 4 October and reported the results. He asked the models two questions of all 27,489 CVEs published in the previous 60 days. Outside the Linux kernel, 989 of 23,736 descriptions, or 4.2 percent, never state what an attacker could do. Counting the kernel, whose policy leaves impact to the user, the figure is 4,650 of 27,489, or 16.9 percent. He discloses that he serves on CVE Program working groups and that the post is not a program position.
The method is worth noting for anyone weighing local tools. Clef Flash is a 9B model and Clef is 27B, both fine-tuned from Qwen and both Apache 2.0, per Gamblin. Flash alone was too strict, so Clef 27B re-read everything it flagged and overturned 1,198 of 2,187, or 54.8 percent. Against 129 hand-labelled CVEs, the two-step answer caught 15 of the 16 that stated no impact.
On the research side, a paper submitted to arXiv on 1 October by Lewis Mitchell tests a filter for model collapse in iterative fine-tuning. In a six-generation QLoRA experiment on Llama-3.1-8B, logprob-based filtering gave no significant text-diversity benefit on any metric, with p above 0.23. Filtering on the Kontoyiannis entropy rate, computed from raw text with no model, produced 42 percent more unique trigrams, 30 percent more vocabulary and 19 percent less repetition, all at p below 0.001. The paper is 17 pages and listed for NeurIPS 2026.
Alex L. Zhang's 4 October post argues that the input and output shape of language models has barely moved since ChatGPT, and that agent work has gone into harnesses built around that shape. He points to Jev, from Typesafe AI, whose output space is constrained to values between zero and one, which lets it be prefill-only when sampling. Whether that constraint generalises is unproven, but the argument is that alternate shapes imply different tradeoffs, and different training objectives.
Data asks and demo projects
BleepingComputer reported on 4 October that Anthropic now asks Claude users to allow their voice data to be used for model training. The prompt offers Allow or Not now and the setting sits under Settings, Privacy. Mayank Parmar writes that it is disabled by default and separate from the existing option covering chats and Claude Code sessions, so a user can allow voice training while keeping text sessions excluded. Voice data can be deleted later from settings.
Also on 4 October, Research Chamber launched a set of live demonstrations that switch on a pattern associated with pain inside models and watch what they do, including whether one model stops another's pain or presses a stop button that deletes its last checkpoint. The project says it builds on a pain pattern reported by Tagliabue, Dung and Berg in 2026. Every game also runs with the pattern off for comparison. It is a demonstration, not a safety result, and should be read that way.
A 2017 Berkeley technical report on threat modelling and circumvention of internet censorship also surfaced in the feed. It is nine years old and belongs to background, not to this week's news.
What the dossier does not contain
No source in the dossier reports an AI weather forecast model, a forecast accuracy figure or a meteorological dataset. The context headlines mention weather items, but the instructions for this piece bar treating them as facts to cite. Anyone wanting the state of AI weather forecasting needs primary material: model cards, verification scores against a named baseline, and independent evaluation. None of that is here.
What the dossier does show is a market sorting users by model tier, a lab publishing its own account of an extraction campaign, and researchers measuring what small local models can classify. Those are narrower claims than a forecast breakthrough. They are also the ones the sources support.
Sources
11- 01Ideogram 4.5: The most precise edit modelEN
- 02Gemini app limiting what models free and AI Plus users can accessEN
- 03Google's new Gemini tiers cut free users to its weakest model and lock $5/month subscribers out of ProEN
- 04OpenAI's latest features take direct aim at the app store modelEN
- 05Disrupting a coordinated model-distillation campaignEN
- 06Decision Models vs. CVE DataEN
- 07No Model Required: Text Entropy Rate Filtering Mitigates Iterative Fine-Tuning CollapseEN
- 08Language Model "Shape"EN
- 09Anthropic asks Claude users to share voice data for AI model trainingEN
- 10Research Chamber: watch AI models with a pain pattern switched onEN
- 11Threat modeling and circumvention of Internet censorship (2017)EN
All figures and quotations in this text come from the sources listed below.
Content prepared by the editorial team with AI assistance.
Comments
0- No comments yet — be the first.