Skip to content
World clockEU--:--UK--:--USA--:--CN--:--PLDEFRIT中文EN

portal about AI and technologyevents · analysis · interviews · technical background

Search
LIVE
›

Google moves Gemini's free tier down a model as image and video tools split by price

On 3 October 2026, 9to5Google reported that Google's updated support document limits users without a Google AI subscription to the 3.5 Flash-Lite model from 9 October. That ends free access to 3.6 Flash and 3.1 Pro.

AI & modelsAnalysisGrace OkonkwoPublished: 4 October 20266 min readSources 9
Google moves Gemini's free tier down a model as image and video tools split by price
On 3 October 2026, 9to5Google reported that Google's updated support document limits users without a Google AI

The change lands in the Gemini app, not in the model releases themselves. According to 9to5Google, the support document updated that week says unpaid users will be cut to 3.5 Flash-Lite starting 9 October. AI Plus subscribers paying $4.99 a month keep Flash-Lite and Flash but lose Pro. Only AI Pro at $19.99 and AI Ultra at $99.99 or $199.99 keep all three models.

The Decoder read the same support page and reached the same table. Its report on 4 October adds the comparison that matters commercially. The arrangement mirrors OpenAI's freemium structure for ChatGPT, where the free product runs on a smaller model and the frontier tier sits behind a card.

Two details in the 9to5Google piece are easy to miss. AI Plus subscribers will get an email telling them when the change takes effect. AI Pro customers also pick up the Deep Think option for "maximum parallel reasoning," previously confined to the two Ultra plans. Google has said Gemini 4 Argon arrives first for AI Ultra subscribers.

What a cheaper image model buys

While Google tightens the top of its ladder, Ideogram went the other way on pricing. The company's own model page, published on 4 October, lists Ideogram 4.5 with four quality modes at native 2K resolution. The fixed price per image is $0.008 at very low quality, $0.03 at low, $0.06 at medium and $0.10 at high, with no subscription required. Its Precise Edit endpoint costs $0.22 at high quality, the same as a reference-based generation.

The pitch is multi-turn editing. Ideogram claims the model eliminates artifact buildup, so an edited crop can be stitched back into the original without pixel shifts or colour changes. The page also states that price does not change with prompt length, the number of reference images or output resolution. That is a direct shot at tools that bill by tokens or by iteration. The page does not address whether the artifact claim holds outside a vendor's own before-and-after gallery.

Ideogram shipped into a market where agents now do the most visible image and video work, not people opening an app. An open-source project called explainroo, posted to GitHub on 4 October, takes a script and renders a narrated MP4 locally. Kokoro reads the voice track in eight languages with 54 voices and no API key. Whisper times each word, Chrome draws the frames, and ffmpeg assembles the file. The repository's changelog dates version 0.4.1 to 20 October, when it kept arrow and chart labels whole while a shot holds.

The project is built for coding agents, not editors. The README says it works best with Claude Code and Opus 5.5, and that the agent checks its own output because it cannot watch the video. That is a small but telling detail about where multimedia production is heading. The person specifying the video may never see the frames until the MP4 exists.

Voice data, on request

Multimodal training data has its own price. BleepingComputer reported on 4 October that Anthropic has started prompting Claude users to share voice conversations for model training. The option is disabled by default, with a toggle under Settings, Privacy. The prompt reads: "Allow us to use your voice data to improve our AI models." Voice training is kept separate from the existing chat and Claude Code training controls, so a user can allow one and refuse the other.

Anthropic's move is voluntary and reversible, which puts it at the softer end of the spectrum. OpenAI's disclosure on 4 October sits at the harder end. The company said it disrupted a coordinated campaign to extract protected reasoning from its models. The earliest activity came in the first week of July, with high-volume spikes on 24 and 25 July of 16,000 requests from over 4,000 users. A related cluster of more than 15,000 users was fully disrupted by 28 July. OpenAI attributed a core cluster to individuals associated with Moonshot AI, the developer of Kimi, and said the operators did not break encryption or access stored conversations.

One sentence matters for anyone shipping multimodal products, and it is OpenAI's own. This manipulation "is not a vulnerability unique to OpenAI's models." The company shared information through the Frontier Model Forum. Extracted reasoning, its post argues, can train another model without carrying over the safeguards applied to the original's user-facing outputs.

Decisions instead of paragraphs

Not every new model is trying to write. Jerry Gamblin serves on the CVE Program's quality and automation working groups and states that nothing in his post is a CVE Program position. He ran Cloudflare's Clef and Clef Flash against all 27,489 CVEs published between 4 August and 3 October 2026. Clef is a 27B model fine-tuned from Qwen 3.8 and Clef Flash a 9B model from Qwen 3.5. Both are Apache 2.0 with weights on Hugging Face, and both are reachable through the decision endpoint Ollama added in 0.35.

The models return choices and probabilities rather than text. Outside the Linux kernel, 989 of 23,736 descriptions, or 4.2%, never state a security impact. The kernel's policy leaves impact to the user. Including the kernel, the figure rises to 4,650 of 27,489, or 16.9%. Flash alone was too strict, so the 27B model re-read its flags and overturned 1,198 of 2,187, or 54.8%. Against 129 CVEs he labelled by hand, the two-step answer caught 15 of the 16 he judged to state no impact.

That is a classifier doing a job most people assumed needed a language model's prose. Alex L. Zhang makes the broader argument in a blog post on 4 October. The input and output shape of language models has barely changed since ChatGPT, he writes, and almost all agent work has gone into building a harness around that shape rather than changing it. He points to Jev, a model from Typesafe AI whose output is constrained to a number between zero and one, as an example of a different tradeoff.

A paper submitted to arXiv on 1 October by Lewis Mitchell offers a related result from the training side. Filtering synthetic text with a non-parametric entropy rate estimator, with no model involved, gave +42% unique trigrams and -19% repetition across six QLoRA generations on Llama-3.1-8B. Logprob-based filtering showed no significant benefit.

None of this settles what Google's tier change means for the average user. The Decoder's assessment is that the real-world impact is likely small. Most casual users do not know which model they are running, and power users already pay.

The harder question is whether image and video features follow the same wall. Neither the 9to5Google report nor the support document summarised by The Decoder says how the new effort levels, low, medium and high, will be metered across plans. 9to5Google notes it is not clear where Argon will sit. Until Google publishes that, the pricing table is the only reliable guide.

Comments 0

Sources

9
  1. 01Gemini app limiting what models free and AI Plus users can accessEN
  2. 02Google's new Gemini tiers cut free users to its weakest model and lock $5/month subscribers out of ProEN
  3. 03Ideogram 4.5: The most precise edit modelEN
  4. 04Explainer videos and product demos made by your AI agent. Free and open sourceEN
  5. 05Anthropic asks Claude users to share voice data for AI model trainingEN
  6. 06Disrupting a coordinated model-distillation campaignEN
  7. 07Local models on 27,489 CVEs: 4.2% omit security impactEN
  8. 08Language Model "Shape" - Alex L. ZhangEN
  9. 09No Model Required: Text Entropy Rate Filtering Mitigates Iterative Fine-Tuning CollapseEN

All figures and quotations in this text come from the sources listed below.

Content prepared by the editorial team with AI assistance.

Grace Okonkwo

Grace Okonkwo

AI, models and technology

Grace Okonkwo covers AI, models and technology for FLASH24, working from primary sources such as model cards, API documentation and benchmark papers rather than vendor summaries. She checks training data provenance, evaluation conditions and reported scores against the underlying datasets before any figure reaches print. She interviews researchers and engineers directly, tracks release calendars from major labs, and compares successive model versions on the same tests. Her own self-hosting, home-network and documentation-reading habits feed straight into that desk, since she tests tools on her own hardware first. She does not publish benchmark claims without a reproducible method.

Newsroom →

Comments

0
  1. No comments yet — be the first.

Write a comment

Comments are public. We do not publish abuse, spam or advertising.