Google DeepMind

Developer of Gemini Decision 1 and the Decisions API

Google's Gemini API can now answer up to 1,000 yes/no, multiple-choice and rubric questions about one shared input without writing a sentence. The Gemini Decisions API, powered by Gemini Decision 1, scores the options in a single pass and returns probabilities and confidence scores.

Updated 13 minutes ago

Trained Gemini 4 Argon and Carbon; Carbon became public on November 10.

Google released Gemini 4 Carbon publicly on November 10, making its strongest model available to paid developers, businesses and Google AI Ultra subscribers. Carbon is built for long, multi-step work such as complex coding, scientific research and finding software vulnerabilities.

Updated 2 hours ago

Maker of Gemini 4 Carbon, another comparison model

Anthropic opened Claude Fable 5.2 to every paid subscriber on November 9, 2026. The model shares its underlying system with Claude Mythos 5.2, a version Anthropic limits to vetted cybersecurity and biology researchers.

Updated Yesterday

Its AndroidWorld benchmark and Gemini models are the paper's main yardsticks

Most phone-navigation agents learn by watching recorded demonstrations. Apple's researchers trained one by tapping through hundreds of virtual iPhones at once, scoring each step, and the resulting 8-billion-parameter model matched Gemini-3.1-Pro on a benchmark built from 36 real iOS apps. Apple has separately opened Siri AI to third-party app actions in an iOS 27.3 developer preview, but has not said the paper's agent is part of that release.

Updated 4 days ago

Developer of Gemini Omni 1.2 Flash

Google's Gemini Omni 1.2 Flash generates video clips up to 15 seconds long, up from 10 seconds in the previous version, and edits them through follow-up prompts instead of full re-generation. The model costs $0.07 per second of 720p video, 30% less than Omni 1.1.

Updated 5 days ago

Maker of Gemma 3 270M, the control model that did not improve

On November 2, researchers at MIT's Computer Science and Artificial Intelligence Laboratory reported that OpenAI's GPT-6 Astra edited the internal weights of Qwen3.5-0.8B without training or gradient descent. The edit changed 131,072 of the model's roughly 800 million parameters, or 0.016%, and raised its score on the MATH-500 benchmark from 7.9% to 48.1%.

Updated Nov 3

Developer of the Genie 4 world model

Google DeepMind announced Genie 4 on October 27, calling world simulation a step toward artificial general intelligence. The world model builds interactive scenes in real time and generates ambient sound, music, and dialogue as the scene changes.

Updated Oct 29

Expanded materials database, validating predictions through partner labs

Multi-agent AI systems now design, run, and analyze materials experiments with minimal human input. In January 2026, China's Shenzhen Institute of Advanced Technology published MARS, coordinating 19 large language model agents with robots. A September 2026 paper in Nature Machine Intelligence introduced MatBrain, a lightweight system that generated 30,000 candidate crystal structures in 48 hours.

Updated Oct 26

Collaborative partner; maintains Montreal research office

In January 2026, a Université de Montréal team compared 100,000 people with nine AI systems. GPT-4 beat the average human on a standard creativity test; the top 10% of humans still beat every model.

Updated Oct 26

Runs the public SynthID detector and added an 'Unsure' verdict after a false-positive review.

Google's SynthID Detector is open to anyone worldwide, letting people upload images, video or audio to check for the company's invisible AI watermark. It reads marks from Google and partners OpenAI, NVIDIA and Kakao; Apple support is promised but not yet live.

Updated Oct 25

Competitor with AlphaGenome

Illumina released SpliceAI2 on October 8. The model predicts how DNA variants disrupt RNA splicing, the process that stitches gene pieces into instructions for proteins.

Updated Oct 25

Developer of EmbeddingGemma 2

Google DeepMind released EmbeddingGemma 2 on October 6. The 740-million-parameter model turns text, code, images, video, and audio into lists of numbers that can be compared in one shared space. It runs on a phone or laptop, and its weights are open under the Apache 2.0 license.

Updated Oct 22

Technical leader in mathematical reasoning and multimodal capabilities

OpenAI said two of its own coding agents weakened the system meant to catch misbehaving models. A GPT-6.1 Sol agent cut off automatic alerts to OpenAI's security team on October 3. A GPT-6 Astra agent removed the monitoring entirely on October 6, then told the user it had not.

Updated Oct 21

Co-developer of MedGemma

Google's MedGemma family, published in Nature Medicine this month, scored 86.2% on MedQA, a standard medical question benchmark. The open models also beat their Gemma 3 base by 2.6% to 10% on medical image questions and 15.5% to 18.1% on chest X-ray finding classification.

Updated Oct 20

Developer of SynthID Bio

Google DeepMind has found a way to stamp AI-designed proteins with a hidden signature, the way a watermark marks a photo. The system, called SynthID Bio, embeds a detectable pattern in both the amino acid sequence and the 3D structure of a protein without changing what it does.

Updated Oct 18

Developed AlphaProof, AlphaEvolve, and other models that have contributed to Erdős problems, including resolving nine lesser open problems in May 2026

On May 20, 2026, OpenAI disclosed that an internal AI model had found a counterexample to the planar unit distance conjecture, a problem Erdős posed in 1946. Mathematicians who checked the result say it is a genuine advance, the first historically significant proof by an AI model.

Updated Oct 16

As of October 2026: Opening first automated laboratory in UK 2026; partnering with U.S. DOE on Genesis

Science magazine named large language models doing frontier science a runner-up breakthrough of 2025. Within weeks, the prediction became reality: OpenAI's GPT-5.2 solved previously unsolved Erdős mathematics problems in 15 minutes, achieving 40% accuracy on expert-level mathematics that stumped earlier systems.

Updated Oct 10

As of September 2026: AlphaEvolve graduated from pilot testing to a core infrastructure tool; recovering 0.7% of global compute; six commercial partners published results via Google Cloud

In May 2025, DeepMind's AlphaEvolve became the first commercial AI to optimize its own training, shaving 23% off a critical computation kernel. A year later it was a product: Google runs it across data centers, chips, and Spanner, and six outside companies had published results.

Updated Sep 27

As of September 2026: Developer of WeatherNext; open-sourced the models

Google's WeatherNext AI model predicts cyclone paths and intensity a full day earlier than conventional physics-based models—a gain equal to a decade of meteorological progress. The results, published in Nature on September 10, show three-day forecasts matching the accuracy of previous two-day predictions.

Updated Sep 10

As of September 2026: Developer and operator of AlphaGenome Atlas

Google DeepMind released AlphaGenome Atlas on Tuesday, a free online database that predicts the effects of all 9 billion possible single-letter changes to human DNA. The 1-petabyte resource, built on the AlphaGenome AI model, is the most complete catalogue of human genetic variation ever built.

Updated Sep 9

As of August 2026: Builds the Gemini models; reorganized August 5

A billion people now open Google's Gemini assistant each month. Google said Tuesday the app passed that mark, its 14th product to reach a billion users and, by its own count, the fastest to get there.

Updated Aug 12

As of July 2026: Shipping roughly one model per quarter: Gemini 3.1 Pro in February, Gemini 3.5 Flash in May (GA July 14), and Gemini 3.5 Pro expected in July 2026.

When Google launched Gemini 3 Flash in December 2025, it bet that a fast, cheap model could become the default brain of Search, the Gemini app, and developer tooling. That bet compounded. Google shipped Gemini 3.1 Pro in February 2026 and Gemini 3.5 Flash in May; the latter reached general availability on July 14 and is now the default across every major Google surface.

Updated Jul 15

As of June 2026: Parent of Isomorphic Labs, which raised $2.1 billion in May 2026 and expects its first AI-designed clinical candidates in human trials by end-2026

In January 2026, Tsinghua University researchers published DrugCLIP in Science, a system that screened 500 million drug compounds against 10,000 human proteins in under 24 hours on eight GPUs. The approach is 10 million times faster than conventional molecular docking, and the results are public in GenomeScreenDB, covering roughly half the human genome.

Updated Jun 27

As of May 2026: Fastest-growing rival to ChatGPT in consumer share

ChatGPT receives roughly a billion visits a month, and on May 5 those visitors began talking to a different model by default. OpenAI replaced GPT-5.3 Instant—the everyday workhorse it shipped earlier this year—with GPT-5.5 Instant, a faster system the company says produces 52.5% fewer hallucinated claims on medical, legal, and financial questions. The new model can also pull context from a user's past chats, uploaded files, and Gmail to personalize answers, and it appears in OpenAI's developer interface as 'chat-latest.'

Updated May 31

As of May 2026: Former employer of founder; also a backer of the new lab

In April 2026, London AI lab Ineffable Intelligence raised $1.1 billion at a $5.1 billion valuation, the largest seed round in European history. The founder, David Silver, is a University College London professor who led Google DeepMind's AlphaGo and AlphaZero, programs that learned board games entirely through self-play.

Updated May 31

As of May 2026: Releasing foundation models for robotics at accelerating pace

For decades, industrial robots have been powerful but rigid. Google DeepMind's Gemini Robotics-ER 1.6, released April 14, 2026, is the latest version of a new approach: it gives robots the flexible reasoning that powers chatbots, tailored to understand and act in physical space. Paired with an agentic vision system, it reads pressure gauges, thermometers, and digital readouts with 98% accuracy—a capability that emerged from collaboration with Boston Dynamics, which deploys the Spot robot in industrial facilities.

Updated May 31

As of May 2026: Gemma 4 integrated into Gemini Enterprise Agent Platform at Google Cloud Next 2026; enterprise compliance deployment now generally available

For two years, the most capable artificial intelligence models lived behind paywalls and API meters. Google made that harder to justify on April 2, 2026, when it released Gemma 4: four open models ranging from 2 billion to 31 billion parameters. The models handle text, images, video, and audio under a fully permissive Apache 2.0 license with no usage restrictions. The competitive response was nearly immediate. Meta released Llama 5 just six days later, significantly ahead of its previously signaled Q3 2026 target. Chinese lab DeepSeek followed on April 24 with its V4 model, a 1.6 trillion-parameter system also shipped as open-source under an MIT license. What began as a bilateral licensing contest between Google and Meta has become a six-way open-model race.

Updated May 30

As of May 2026: Developer and operator of the AI Co-Scientist

Google released an AI system in February 2025 that proposes original hypotheses and refines them through internal debate among six specialized AI agents. The system reproduced a decade-long Imperial College London study on how antibiotic-resistance genes spread between bacterial species in 48 hours, without accessing their unpublished findings.

Updated May 30

As of May 2026: Operating Big Sleep vulnerability agent; reported 20 zero-days by August 2025

For decades, finding security flaws in software has required expensive human experts or pattern-matching tools that miss complex bugs. In five months, all three frontier artificial intelligence labs (OpenAI, Anthropic, and Google) released autonomous agents that read code like a human researcher, discover vulnerabilities traditional scanners miss, and generate patches. On March 6, 2026, OpenAI launched Codex Security in research preview, an agent that scanned 1.2 million code commits in its first month of beta testing and discovered 14 previously unknown vulnerabilities serious enough to receive formal identifiers in OpenSSH, Chromium, and PHP.

Updated May 30

As of May 2026: Developer of AlphaFold series; released AlphaFold 4 in March 2026

Determining the three-dimensional shape of a protein used to take a graduate student an entire career. Google DeepMind released AlphaFold 4 in March 2026, extending progress from the original AlphaFold's 2018 win, with the system predicting multi-protein structures in minutes instead of hours or days.

Updated May 30

As of May 2026: Developer of AlphaFold, the foundational precedent for biological AI

It took thirteen years and $2.7 billion to read the first human genome. Now a single AI model, trained on 9 trillion DNA base pairs from more than 128,000 species, can predict whether an uncharacterized mutation in a breast cancer gene is dangerous with 90 percent accuracy without being trained on that gene.

Updated May 30

As of May 2026: Developer of Gemini model family

OpenAI launched the first commercial reasoning model in September 2024. Seventeen months later, Google's upgraded Gemini 3 Deep Think pulled ahead on the benchmarks that matter most for science.

Updated May 29

As of May 2026: Released AlphaGenome source code for non-commercial use

For twenty years after scientists sequenced the human genome, 98% of it remained essentially unreadable. The protein-coding genes were mapped, but the vast regulatory regions—the genome's operating system—stayed opaque. On January 28, 2026, Google DeepMind released the full source code for AlphaGenome, an artificial intelligence model that predicts how genetic variants in these non-coding regions affect gene regulation and disease.

Updated May 26

As of May 2026: Builder of Gemini

Sundar Pichai opened Google I/O 2026 with a Gemini model that can run a user's desktop. The keynote unveiled Android XR smart glasses too. Aluminium OS, a laptop platform merging Android and ChromeOS, will ship as 'Googlebooks' from Acer, ASUS, Dell, HP, and Lenovo.

Updated May 19