Google's Gemini tiers, and what the rest of the AI release cycle is doing to speech models
On 9 October Google will cut free Gemini users down to Flash-Lite, 9to5Google reported on 3 October, closing off 3.6 Flash and 3.1 Pro. The same week brought a flurry of speech and translation model launches that almost none of those users will be able to reach.

Google's change is narrow and blunt. The support page, as summarised by 9to5Google on 3 October, says users without a paid Google AI plan lose access to 3.6 Flash and 3.1 Pro. AI Plus subscribers at $4.99 a month keep Flash-Lite and Flash, and lose Pro. Only AI Pro at $19.99 and AI Ultra at $99.99 or $199.99 keep all three.
Why it matters for anyone who uses Gemini for translation or dictation: your billing tier now sets the model doing the work, not your task. A free user who wants a harder language pair gets Flash-Lite. The same support document says AI Pro customers gain the Deep Think option, which until now sat on the two Ultra plans. It also says the app is adding low, medium and high effort levels for each model. Neither Google post explains what happens to voice features specifically. 9to5Google notes it is still unclear where Gemini 4 Argon lands, and whether it counts as a Pro model or a tier above.
A crowded month for speech, and not much of it from Google
Speech and translation have been the busiest corner of the AI release cycle this autumn. Recent coverage points to several launches. Tencent open-sourced a model for real-time Chinese-English interpretation. Block launched Dessa, an AI model trained entirely on human speech. Microsoft introduced new MAI models aimed at real-time and multilingual voice. Google added speech translation to Meet in 63 languages. Indian researchers released SraVaani, a speech model covering 65 Indian languages. Those launches are the context in which Google's tier change lands. The pattern is that speech keeps getting cheaper and more local. Meanwhile the assistant most people already have quietly gets a paywall around its better models.
The most concrete new entry in that run is Ideogram 4.5, published on the company's own model page on 4 October. It is an image editing model rather than a speech one, but its pricing is the clearest signal of where per-call AI costs are going. Ideogram lists a fixed price per image that does not change with prompt length, number of references or resolution. The price is $0.03 at low quality, $0.06 at medium, and $0.10 at high for prompt-only generation. It is $0.22 at high quality when references or a precise edit are involved. Very low quality is priced at $0.008 for the reference and edit endpoints and is not offered for prompt-only generation. No subscription is required.
That flat fee is worth holding next to Google's subscription ladder. Ideogram charges per image and lets anyone pay as they go. Google charges per month and gates the model. For a developer building a translation tool, the first model is a line item that scales with use. The second is a ceiling that arrives on the 9th of the month.
The eavesdropping problem nobody solved
Voice is also the input type with the worst privacy record, and the past few days produced a clean example. BleepingComputer reported on 4 October that Anthropic has started asking Claude users to volunteer their voice conversations for model training. The prompt reads: "Allow us to use your voice data to improve our AI models." The setting sits under Settings, Privacy, is off by default, and is separate from the existing switch for chat and Claude Code sessions. A user can therefore allow voice training while keeping text out. Anthropic says users can turn it off or delete the data later.
That is the reasonable version of the ask. It is still an ask, and it arrives in the same month that Google is making the better Gemini models a paid product. The two moves pull in opposite directions for a user who wants good speech recognition without paying twice: once in money, once in recordings.
Whisper-style local models are the obvious escape hatch, and the tooling around them is improving. Jerry Gamblin, who serves on the CVE Program's quality and automation working groups, published a run on 4 October. He used Cloudflare's Clef models through Ollama's new decision endpoint to classify all 27,489 CVEs published in the previous 60 days. Clef is a 27B model fine-tuned from Qwen 3.8, Clef Flash is a 9B model fine-tuned from Qwen 3.5, both Apache 2.0, weights on Hugging Face, both runnable locally. He reports Clef Flash answered two questions per CVE at a median under a second each on a laptop, with no API key and no bill. That is a classification task, not translation, but it shows what a 9B model on a laptop now does at speed. Gamblin's finding, that 4.2% of non-kernel CVE descriptions never state a security impact, says more about CVE writing than about models.
What the model makers are fighting over
OpenAI used its own security page on 4 October to describe a different fight. It said it disrupted a coordinated campaign to extract protected reasoning from its models, with the earliest activity in the first week of July. The operators did not break encryption or read stored conversations, according to the post. They manipulated interactions so hidden reasoning became visible to the requester. That included copying encrypted reasoning from one conversation and asking a model in another to decrypt and transcribe it. OpenAI says activity spiked on 24 and 25 July to 16,000 requests from over 4,000 users. It says related prompt patterns across more than 15,000 users were fully disrupted by 28 July. It attributes a core cluster to individuals associated with Moonshot AI, the developer of Kimi, and says it is unclear whether all operators came from a single actor.
None of that is about speech. It matters to anyone building on top of a hosted model, though, because it shows how much of a model's value sits in the parts the provider does not publish. A translation feature that depends on a vendor's reasoning layer inherits that vendor's security posture and its terms. The same applies to a voice feature that depends on a vendor's transcription stack.
On the research side, Lewis Mitchell's paper on arXiv, submitted on 1 October, offers a cheaper way to keep fine-tuned models from collapsing into repetition. The method filters training text by an entropy rate estimated from raw text alone, with no model log-probabilities. In a six-generation QLoRA experiment on Llama-3.1-8B, logprob-based filtering gave no significant text-diversity benefit on any metric (p > 0.23). The entropy filter produced 42% more unique trigrams, 30% more vocabulary and 19% less repetition, all at p < 0.001. The paper is 17 pages, eight figures, and listed for NeurIPS 2026. For anyone fine-tuning a speech or translation model on synthetic data, that is a concrete recipe rather than a warning.
The app store is moving into the chatbot
TechCrunch reported on 29 September that OpenAI's Dev Day announcements, taken together, turn ChatGPT into a place where apps are discovered and run. ChatGPT, which OpenAI says has 1.2 billion weekly users, will suggest apps mid-conversation when one would help. An expanded plug-in architecture now supports extensions that render interactive panels inside the chat. "Sign in with ChatGPT" carries a user's AI allowance into third-party apps, with 16 launch partners including Cognition's Devin, Notion, Vercel and Figma. If that works, the assistant becomes the distribution channel, and the model behind it becomes the thing everyone else has to route through.
Set that beside Google's tier change and the shape of the market is clearer. Google is charging for access to its better models. OpenAI is giving away reach and keeping the model. Anthropic is asking for voice recordings. Ideogram is selling images by the unit. Each one answers the same question differently: who pays for inference, and what do they get in return.
For a reader trying to pick a translation tool this month, the practical questions are narrower. Which model does the free tier actually run, and will that change without notice. Where do the recordings go, and is there a switch. And if the answer to either is unsatisfactory, is there a local option that is good enough. On the evidence of the past week, the third answer is getting better faster than the first two.
Sources
8- 01Gemini app limiting what models free and AI Plus users can accessEN
- 02Google's new Gemini tiers cut free users to its weakest model and lock $5/month subscribers out of ProEN
- 03Ideogram 4.5: The most precise edit modelEN
- 04Anthropic asks Claude users to share voice data for AI model trainingEN
- 05Decision Models vs. CVE DataEN
- 06Disrupting a coordinated model-distillation campaignEN
- 07No Model Required: Text Entropy Rate Filtering Mitigates Iterative Fine-Tuning CollapseEN
- 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.
Comments
0- No comments yet — be the first.