✓“Today, we’re launching EmbeddingGemma 2, expanding beyond text to unify code, images, video, and audio in a shared embedding space.”
✓“Built on the Gemma 4 architecture and released under a commercially permissive Apache 2.0 license, EmbeddingGemma 2 has 740 million parameters, making it optimal for on-device inference.”
✓“Using Matryoshka Representation Learning (MRL), developers can dynamically truncate output vectors from 768 dimensions down to 512, 256, or 128 dimensions.”
Day 7Users report · 14
›A developer who runs a website on embedding models said it is a drop-in replacement for both the first EmbeddingGemma and an image embedding model, with better quality, less memory and faster embeddings.
›A user added image, audio and video input to their own local embedding tool and measured it on an M3 Pro laptop: about 78 short texts, 4 images or 6 thirty-second audio chunks a second.
›A user's first tests found it no better or faster than the first version for text-only retrieval or clustering, and preferred a specialist model for music audio.
~“Stronger benchmark performance, faster task completion, and shorter outputs than the previous version”
~“DeepSeek-V4.1-Flash-1017 uses reinforcement learning with self-feedback to improve end-to-end project development”
Day 7Users report · 12
›A tester who benchmarked the new weights against the old ones found fewer output tokens at every reasoning level, with the lowest level now usable, and the same intelligence and tool-calling pattern.
›A user ran three hard questions with it and two Claude models as worker agents in Claude Code; it came out cheaper and more correct than the mid-tier Claude model.
›A user who ran it over a weekend for simulation code found it good, but got a better result from a Qwen model on the same task, and called it too early to judge.
~“Gemini 4 Argon is now globally available today for developers and enterprises via the API in AI Studio and Gemini platform, for Google AI Ultra subscribers in Google Antigravity, and Google AI Ultra and Pro subscribers in the Gemini app and AI Mode in Search.”
~“Today, we’re rolling out Gemini 4 Argon directly inside Google Antigravity for Google AI Pro and Ultra subscribers — available in the Antigravity desktop app and the Antigravity CLI.”
✗“The model launches today at an introductory price of $2 per million input tokens and $10 per million output tokens (with cached input tokens at 95% off).”
Day 7Users report · 10
›AI Pro subscribers objected that Argon was Ultra-only in Antigravity while Google One pages advertised it for Pro. One reply noted that a few Argon questions filled a five-hour usage window.
›Pro users asked for Argon in Antigravity, and several said older Gemini models had also vanished from their model picker that week.
›Google AI Ultra for Business subscribers say Argon shows in the list but is not accessible in the Antigravity app or CLI; one was told it was due to the rollout.
~“It performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5.”
?“Input and output tokens are $4 and $20 per million, 20% less than Opus 5.”
~“In addition to the price drop, we’re increasing five-hour usage limits on Pro, Max, Team, and seat-based Enterprise plans.”
Day 7Users report · 7
›A Claude Code user compared local transcripts week over week and found the weekly limit draining about 3.6x faster after the 25 Sep reset with the same workflow and model mix.
›Another user reports weekly and session limits going much faster than the previous week while using a single conversation.
›A user with an unchanged setup measured about 2x thinking tokens and 1.6x output per request from 1 Oct, with worse judgment, and says the same shift showed up in claude.ai chat.
~“GPT-6.1 Sol is available starting today to all Plus, Pro, Business, Enterprise, and Edu users in ChatGPT Work and Codex.”
~“In the coming days, we’ll also offer GPT-6.1 Sol Ultrafast, with up to 8x faster token generation compared to its standard speed in Codex.”
~“GPT-6.1 Sol is more transparent about its limitations and more reliable at respecting user intent and safety constraints.”
Day 7Users report · 5
›A Pro user could not find 6.1 Sol in the Codex desktop app although it showed in Work. A maintainer later posted the minimum desktop, CLI and VS Code versions.
›Users outside the US said they still could not see the model in the ChatGPT desktop or web app two days after launch, while one UK user said it was already the Codex CLI default. Staff suggested updating the app.
›OpenAI announced Ultrafast for the API, Codex and ChatGPT Work, limited to Pro 500 and eligible Enterprise/Edu plans. A user asked for a trial, since most people can't tell if it's worth the upgrade.
~“Muse Spark Open is the first model in its class to run on your desktop using two Apple M5 Ultra Mac Studios with the latest MLX framework.”
?“Muse Spark Open is available today through API platforms at Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI.”
?“We’re releasing the weights under the Muse Model License with limited commercial terms, which allow many developers to use the model for free.”
Day 7Users report · 6
›A commenter quoted Meta's suggested Mac Studio setups and said they could find no information on inference speed.
›A commenter noted the licence requires large inference providers to get a separate licence from Meta, which they expect to reduce price competition between hosts.
›On the model's Hugging Face page, a user objected that downloading requires handing over personal data and said it put them off trying it.
?“a new tri-hybrid architecture and training recipe allow LFM3-1.7B to outperform our previous LFM2.5-2.6B while using 35% fewer parameters.”
?“the text model fits in about 1 GB at 4-bit precision.”
?“256K context window for the text model.”
Day 7Users report · 2
›A user tried the omni and text models as MLX builds in several Mac runners and none would load, with a safetensors read error.
›A user said the vision-language model gave better text-only answers than the 1.7B text model in the playground, which answered a simple animal-sound question wrongly.
?“At max effort, Terra is over three times as token efficient as Claude Opus 5.5, and one-fourth its price per token, resulting in a ~9x lower cost per task.”
?“Terra is also much quicker, completing tasks over three times faster than both GPT-6 Astra and GPT-6.1 Sol.”
?“At maximum reasoning effort, GPT‑6.1 Terra solves 50.3% of tasks, more than double GPT‑5.6 Terra’s published score of 21.5%.”
Day 7Users report · 1
›A user running their own benchmark said Terra at max effort scored below Sol and cost about twice as much, while medium effort looked like the best use and cheaper than Sol.