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

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

Search
LIVE
›

Google puts Flash-Lite behind the free tier as Gemini 4 Argon waits in the wings

Google will stop giving free Gemini app users access to Flash and Pro on 9 October, leaving them only Flash-Lite, according to an updated support document reported by 9to5Google on 3 October.

AI & modelsExplainerRachel NwosuPublished: 4 October 20268 min readSources 8
Google puts Flash-Lite behind the free tier as Gemini 4 Argon waits in the wings

Google will stop giving free Gemini app users access to Flash and Pro on 9 October. They will keep only Flash-Lite, according to an updated support document reported by 9to5Google on 3 October. The same document says AI Plus subscribers, who pay $4.99 a month, will lose Pro and keep Flash-Lite and Flash.

This is a downgrade, not a reorganisation. Until now the free tier included 3.6 Flash and varying access to 3.1 Pro, according to The Decoder's write-up on 4 October. After 9 October, all three models sit behind AI Pro at $19.99 a month or AI Ultra at $99.99 or $199.99 a month. The three models are Flash-Lite, Flash and Pro. What changes for the reader is simple. The model picker in the Gemini app stops being a menu of everything and becomes a price list. Anyone who relied on Pro for free, or paid $4.99 expecting it, gets a weaker model or a bigger bill. What is still unknown is how Google will treat its next flagship, Gemini 4 Argon, when it lands.

Two outlets, one support page, slightly different framing

The two write-ups diverge on timing. 9to5Google dates the change to 9 October. It says AI Plus subscribers will get an email explaining when their changes take effect. The Decoder frames the shift as starting in October 2026 without naming a day. Both trace it to the same Google support page, so the gap looks like framing rather than a contradiction. Neither outlet reports a statement from Google beyond the document.

9to5Google also reports that the limits arrive on top of a compute-based usage change Google made in May. AI Pro and AI Ultra users, meanwhile, are not losing anything. The $19.99 tier gains the Deep Think option for "maximum parallel reasoning," which until now was reserved for the $99.99 and $199.99 plans. The app also looks set to add "low," "medium" and "high" effort levels for each available model. That mirrors what Google AI Studio and Antigravity already offer.

The model nobody can place

Where Gemini 4 Argon fits is the open question. 9to5Google reports that Google said earlier this week that Argon is first coming to AI Ultra subscribers. The company has yet to specify whether Argon counts as a Pro model or sits in a new tier above it. That ambiguity matters more than the Flash-Lite cut. If Argon lands above Pro, $19.99 buys access to a 3.1-era model while the frontier sits one or two tiers up.

The Decoder raises a related possibility: that the tiering is groundwork for Argon, which could be more expensive to run. It also argues the real-world impact is likely small. Most casual Google AI users have no idea which model they are running, and the power users who do notice are probably already paying. That is a reasonable read, and it is also the read Google would prefer you to take.

The freemium shape is borrowed. The Decoder notes the setup mirrors OpenAI's model with ChatGPT, where free users get a capped experience and paying customers get the frontier. Google's version is blunter. It is not metering how much of a good model you use, it is removing the good model from the cheap plans entirely. There is a pricing table to go with it. Google's support page, as reproduced by The Decoder, gives Flash-Lite to everyone, Flash to AI Plus and above, and Pro only to AI Pro and AI Ultra. Four tiers, three models, one clear squeeze point at $4.99.

Elsewhere in models this week, the releases are not benchmarks

The rest of this week's model news is less about leaderboards than about what gets shipped and who can run it. Ideogram published the details of Ideogram 4.5 on 4 October. It pitches the model as a precise edit model at native 2K resolution, live in its app, its MCP integration and its API. The company says the model eliminates artifact buildup across edits. That is the thing that makes multi-turn editing possible rather than a one-shot trick.

Its pricing is per image and flat: $0.008 in very low quality, $0.03 low, $0.06 medium and $0.10 to $0.22 high. The higher figure covers generation with references and precise edits. Ideogram says the price does not change with prompt length, number of reference images or output resolution, and that no subscription is required. For anyone comparing image models on cost per usable output rather than on a demo reel, that is the number that matters.

On the open-weights side, the most interesting release this week is not a frontier model at all. Jerry Gamblin ran Cloudflare's Clef and Clef Flash over all 27,489 CVEs published in the 60 days to 3 October. He asked each description two questions: does it establish a security impact, and is it clear to a defender? Clef Flash, a 9B model fine-tuned from Qwen 3.5, answered both at a median of under a second per CVE on a laptop, with no API key.

The results are a benchmark of prose, not of reasoning. Outside the Linux kernel, 989 of 23,736 descriptions, or 4.2%, never say what an attacker could do. The kernel is its own case: its policy leaves impact to the user. So 3,661 of its 3,753 records state none either, which pushes the all-CVE figure to 4,650 of 27,489, or 16.9%. Gamblin serves on CVE Program working groups and says nothing here is a program position. He found Flash alone too strict: the 27B Clef model overturned 1,198 of 2,187 flags, 54.8%. Against 129 CVEs he labelled by hand, the two-step answer caught 15 of the 16 that state no impact.

That is the kind of benchmark release that rarely makes a launch post and changes more day to day work than a leaderboard point. A small model that runs locally and classifies text cheaply is a different product category from a frontier chatbot. The Clef weights are Apache 2.0 on Hugging Face, runnable through Ollama.

Two papers, one warning about training on your own output

Two research items this week both circle the same problem: models trained on model output degrade. Lewis Mitchell's paper, submitted to arXiv on 1 October, tackles model collapse in iterative fine-tuning. The experiment used six generations of QLoRA on Llama-3.1-8B. Logprob-based filtering, the established baseline that needs access to model probabilities, gave no significant text-diversity benefit on any metric (p > 0.23). Filtering by a non-parametric entropy rate estimator computed from raw text alone produced 42% more unique trigrams, 30% more vocabulary and 19% less repetition, all at p < 0.001.

The finding is narrow and the setting is fully synthetic, single-lineage fine-tuning. But it points at something practical: you can filter training data for diversity without a model in the loop. The paper lists 17 pages, 8 figures and NeurIPS 2026 in its comments.

Alex L. Zhang's blog post on the "shape" of language models, published on 4 October, argues from a different angle. He notes that the input and output shape of language models has stayed roughly static since ChatGPT. All agent work is built as a harness around an autoregressive model rather than the other way round. He cites Jev from Typesafe AI, a model whose output space is constrained to the interval from 0 to 1, as an example of an alternate tradeoff space. It is a design argument, not a benchmark result, and it should be read as one.

A security story that is not a model release

The week's most consequential AI security disclosure has no benchmark attached. OpenAI said on 4 October that it disrupted a coordinated campaign to extract protected reasoning from its models. The earliest activity came in the first week of July. The operators did not break encryption or reach stored conversations, the company says. They manipulated model interactions so hidden reasoning could be reproduced at scale.

OpenAI's account gives dates and volumes. Activity began on 1 July. High-volume spikes on 24 and 25 July consisted of 16,000 requests from over 4,000 users. Related prompt-pattern activity across a cluster of more than 15,000 users was fully disrupted by 28 July. OpenAI attributes a core cluster of the activity to individuals associated with Moonshot AI, the developer of Kimi. It says it is unclear whether all operators came from a single actor.

Read alongside the Gemini tier change, the two stories describe the same market from opposite ends. Google is tightening what its cheapest customers can reach. OpenAI is dealing with people who went around the front door to get at reasoning the model does not show. Neither company has said how much the respective changes cost them. Neither has published a number for what the extracted reasoning was worth.

What the Gemini change does not tell us is whether Argon will justify the ladder. Google has confirmed Argon is coming to AI Ultra first. It has not said where Argon sits relative to Pro. Until that is settled, the 9 October cut is a pricing decision with a product question still hanging over it.

Comments 0

Sources

8
  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. 04Decision Models vs. CVE DataEN
  5. 05No Model Required: Text Entropy Rate Filtering Mitigates Iterative Fine-Tuning CollapseEN
  6. 06Language Model "Shape"EN
  7. 07Disrupting a coordinated model-distillation campaignEN
  8. 08OpenAI's latest features take direct aim at the app store modelEN

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

Content prepared by the editorial team with AI assistance.

Rachel Nwosu

Rachel Nwosu

AI, models and technology

Rachel Nwosu covers AI, models and technology for FLASH24, working from public model documentation, benchmark releases and repository histories rather than press summaries, and she skips announcements that arrive without reproducible numbers. She checks training-data claims against dataset cards and reruns reported metrics where code is available. She spends much of her week interviewing researchers and engineers, tracking model launch calendars, and comparing vendor benchmarks with independent evaluations. Outside the desk she runs 3D printers, restores old computers, and tests how models learn from internet junk. She does not publish benchmark figures she cannot trace to a source.

Newsroom →

Comments

0
  1. No comments yet — be the first.

Write a comment

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