Maker · 53 entries
Everything from Google
| Name | What to know | Price or memory | |
|---|---|---|---|
| mixed Strong benchmarks and best at hard reasoning, but slow, token-hungry and uneven hands-on. | $2 / $10 | ||
| promising Close to Fable on agent rankings at half the price, but only a day old. | $3 / $15 | ||
| mixed Strong at launch and low on hallucinations, but retired and overtaken by newer Gemini 4 models. | |||
| mixed Solid small open model for documents and images, but not the leader | ~6.7 GB | ||
| strong Tops Arena for text-to-video and ranks third for video editing, at a lowered price. | $1.5 / $9 | ||
| mixed Fast and cheap for everyday agent work, but behind rivals on several tests. | $1.5 / $7.5 | ||
| unproven Genie 4 is out for Ultra subscribers, but only Google's claims and a few complaints exist. | |||
| mixed Multimodal yes/no model with confidence scores, but pricier and less accurate than rivals | |||
| Gemini 3.8 Flash TTS adds text-prompt voice design and inline expressive tags; in API and AI Studio | |||
| unproven Google touts multi-agent features; no independent evidence they work well yet. | |||
| Suggested by a commenter as an alternative for letting agents control a browser | |||
| Predicts gene expression effects of variants from up to 1M-letter DNA and cell type; finds almost twice as many expression-changing variants; non-commercial API, commercial in private preview | |||
| unproven Reference tool for finding agent failures; results come only from Google's small demo | |||
| strong Gemini 3.8 Flash ranks near the top for image and video understanding | |||
| Continual recursive self-improvement lets harness optimizers reuse and revise prior edits across changing benchmarks or policy models by storing rationales, predicted and observed outcomes, and evaluation contexts in a s | |||
| unproven Impressive claimed security proofs for Signal, but only Google's word so far | |||
| Gemini 3.5 Flash Lite is about four times faster than 3.5 Flash for high-volume tasks; in API and AI Studio | |||
| Now supports self-hosted sandboxes | |||
| unproven Impressive claims from Google, but closed and untested by anyone outside. | |||
| Google's SLEEP alternates wake and sleep phases, consolidating memory and self-generated practice problems; near-perfect recall up to 10M tokens in tests | |||
| Proactive video agent that decides when to offer help during a live stream; code and models promised within 7 days | |||
| mixed Still the reference for natural-language image editing, but 2.1 trailed GPT Image 2.5 | |||
| Google demonstrates magic state cultivation on Willow, cutting errors about 40x to 99.99% fidelity; published in Science | |||
| Designs enzymes from scratch (structure and sequence), checked with AlphaFold 3; lab-tested enzymes for piperidine production and DEHP breakdown; not peer reviewed | |||
| Plant DNA model trained on 320 species; with a Gemma 4 agent ranked the true causal mutation first among about 2,500 candidates 90% of the time | |||
| Open multimodal embedding model: 270M text, +170M vision, +300M audio; runs on phones; about 1.5 GB, Apache 2.0 | ~1.5 GB | ||
| Cheaper high-volume Flash Lite TTS tier, 100+ languages and dialects | |||
| Google details space data center research; launch vibration and radiation tests passed, two-satellite test planned next year | |||
| Trillium TPUs survived launch vibration and withstood more radiation than expected in tests | |||
| Google report improves an agent's search strategy by replaying its past discovery trees; model weights stay fixed | |||
| Gemini 3.8 Live and Extended Thinking launch: 97 languages, speaks while calling tools, in Gemini API, Search Live and Gemini app | |||
| Took #1 on Arena text-to-video | |||
| Searchable ~1 PB database of predicted effects for 9+ billion human genetic variants, 30x AlphaFold DB | |||
| Complete male fruit fly brain and nerve cord map: over 166,000 neurons and about 125 million synapses, called the largest wiring map made | |||
| Defensive-security variant; beats GPT 5.6 Sol and Claude Mythos at vulnerability discovery and made 2.6x more correct Chrome patches than larger commercial models | |||
| Open-source zero-shot multivariate time-series foundation model, 330M parameters, pretrained on over a trillion time points | ~1.3 GB | ||
| Uses live satellite data for hourly global forecasts at resolution as fine as 5 km, versus 25 km every 6 hours for WeatherNext 2 | |||
| Extropic's Z1T training code is written in JAX | |||
| Productrise study: AI Mode shows the same products 21.6% more expensive than Google's traditional shopping search | |||
| Transcription model with verbatim and smart modes, speaker detection, word timestamps and automatic multilingual handling | |||
| Runs very fast in output tokens/sec on GPUs and is very usable as a local model | |||
| Agentic system that finds geographic data, trains models and predicts from natural-language questions on disease or food security | |||
| Took #1 on Arena vision | |||
| Sign language to text on phone camera, trained on 100,000+ hours across 50+ sign languages; launches with ASL on Gboard and Live Transcribe on Pixel 11 | |||
| Its SO400M variant was the starting point for North Micro Vision's custom-trained vision encoder | |||
| ER2 is the high-level reasoning model: plans tasks, corrects failures and coordinates multiple robots | |||
| Smaller version that runs fully offline on the robot; trusted tester program open | |||
| Hold the Fn key on macOS to dictate into any app; Gemini cleans filler words and self-corrections; optional Gemini reasoning mode | |||
| Google uses reinforcement learning on error-correction signals to retune qubit controls live, cutting logical error rate a further 20% | |||
| Turns a video model into one system predicting depth, normals, camera pose, segmentation and 3D keypoints; claimed to beat specialist models | |||
| Parametric 3D head model with 253 identity and 383 expression controls (eyes, teeth, tongue); outputs dense head mesh | |||
Plans
What each one gives you.Google AI Plus
Google AI Pro
Google AI Ultra