AI Breakthroughs

After an API-only preview on 6 October, Mistral released the weights of an improved Mistral Large 4, a multimodal mixture-of-experts model with about 1 trillion total and 52 billion active parameters and a one-million-token context. The licence allows free commercial use, but companies selling it as a hosted service need a separate licence.Mistral AI — Mistral Large 4 open weights →

Google released Gemini 4 Carbon, which it called state of the art on cybersecurity and long tasks, initially to a small group of cyber defenders. Carbon ranked first on the Artificial Analysis Intelligence Index v5.0 with 59 points, against 50 for Gemini 4 Argon. Introductory pricing is $2/$10 per million tokens.DataCamp — Gemini 4 Carbon →

OpenAI released GPT-6.1 Terra, a low-cost everyday model for computer use that reads what is on screen and acts across apps. At $1/$5 per million input/output tokens it costs a tenth of GPT-6 Astra, and OpenAI said it matches GPT-6 Astra's results at that tenth of the price while finishing tasks more than three times faster.OpenAI Developer Community — GPT-6.1 Terra →

Moonshot AI released Kimi K3.1 Pro, a post-trained version of its 2.8-trillion-parameter Kimi K3, at $2/$10 per million tokens with a one-million-token context. It reported 65.7% on Terminal-Bench 4.0, close to Claude Opus 5.5. Moonshot said it deliberately removed cyber-offence data from training, lowering the model's score on its own exploit tests.Simon Willison — Kimi K3.1 Pro →

Anthropic released Claude Haiku 5.5 at $0.10/$0.50 per million input/output tokens for shorter prompts, about 75% cheaper than Haiku 4.5 on average. Its score on an offline OSWorld 2.1 computer-use test rose to 72.4% from 15.7% for Haiku 4.5.Anthropic — Introducing Claude Haiku 5.5 →

Google released Gemini 4 Argon, its new frontier model for software engineering, enterprise work and cyber defence. It went first to vetted cyber defenders, then to Gemini app subscribers and the paid API in October. Google reported 77.9% on DeepSWE v1.1 and raised the output limit to one million tokens. Introductory pricing is $2/$10 per million tokens.Google — Gemini 4 Argon →

Anthropic released Claude Sonnet 5.5 at $2/$10 per million input/output tokens, saying it costs up to 30% less per task than Sonnet 5. It scored 70.6% on Terminal-Bench 4.0, above Opus 5.5's 66.4%, though Anthropic said Opus remains stronger at complex, open-ended work.Anthropic — Introducing Claude Sonnet 5.5 →

Anthropic released Claude Opus 5.5 at $4/$20 per million input/output tokens, down from $5/$25 for Opus 5, and said it costs about 40% less on typical workloads while performing at Fable 5.1's level on most work. Anthropic reported 66.4% on Terminal-Bench 4.0 and 81.8% on the OSWorld 2.1 computer-use benchmark.Anthropic — Introducing Claude Opus 5.5 →

On the same day as Anthropic's Opus 5.5 launch, OpenAI released GPT-6 Sol and GPT-6 Luna, faster and cheaper models built on GPT-6 Astra. Sol costs $2/$10 and Luna $0.10/$0.50 per million input/output tokens. Both are in ChatGPT Work, Codex and the API, and free users can try Luna in the desktop app.OpenAI Developer Community — GPT-6 Sol and Luna →

xAI released Grok 4.7, a reasoning model for agentic coding and long multi-step workflows. It has a 500,000-token context window, adjustable reasoning effort and API prices from $2 per million input tokens and $6 per million output tokens. A faster variant is available only inside selected coding tools.xAI — API release notes →

DeepSeek released V4.1-Flash, the first and smallest model in a new architecture family with built-in image understanding. The 552-billion-parameter mixture-of-experts model uses about 8 billion active parameters for input. DeepSeek said it scores ahead of its V4-Pro flagship, and that its KV cache needs a quarter of the high-bandwidth memory of the previous generation. Weights are openly available; a larger Pro version has been promised without a date.DeepSeek — V4.1-Flash release →

UC Berkeley’s GPN-Star, published in Nature, reads DNA alongside alignments of the same stretch across many species. It outperformed bigger genomic models at flagging variants linked to traits and disease risk while training in hours to days on modest hardware. Versions trained on primate genomes did best for complex traits; deeper evolutionary versions did best for rare protein-coding variants.UC Berkeley news →

OpenAI released GPT-6 Astra, which it called its most capable and most aligned model, in ChatGPT Work, Codex and the API at $10/$50 per million input/output tokens. OpenAI claimed a record 74% on the DeepSWE v1.1 coding benchmark and 89% fewer unintended outcomes than GPT-5.6 Sol on an internal computer-use safety test.OpenAI — GPT-6 Astra →

Anthropic released its top-tier model in two forms: Fable 5.1, generally available, and Mythos 5.1, the same model with fewer safeguards, limited to vetted cybersecurity and life-science organisations. Prices stayed at $10/$50 per million input/output tokens, with a one-million-token context window. Anthropic reported 55.8% on Terminal-Bench 4.0 for Fable 5.1 and 60.9% for Mythos 5.1.Anthropic — Claude Fable 5.1 and Mythos 5.1 →

Anthropic released Claude Opus 5 with a one-million-token context window and a five-level effort toggle, at pricing reported as unchanged from the previous generation. Coverage put its results within half a percentage point of the previous peak on the benchmarks tracked, at roughly half the cost per task. Holding price flat across a capability generation was itself notable: for most of the preceding two years, frontier releases had arrived at a premium, and the shift suggested competitive pressure was landing on margin rather than on capability claims.ThursdAI — July 2026 model releases

Black Forest Labs released FLUX 3, reported as the first model to generate video with synchronised audio of up to twenty seconds while also handling image generation and robot action prediction from a single architecture. Folding action prediction into a generative media model is a meaningful structural change: it treats a robot's next movement and the next frame of a video as versions of the same prediction problem, which is the argument behind the world-model approach several labs were pursuing.ThursdAI — July 2026 model releases

After the International Mathematical Olympiad in Shanghai, RedNote said its dots-note-3.0 model solved all six problems for a full 42 points, and Huawei reported the same score. A year earlier the best AI results were gold-level 35s. These are company claims; no official IMO grading of the AI entries has been published.SCMP report →

Moonshot AI released Kimi K3, reported as a 2.8-trillion-parameter mixture-of-experts model with a one-million-token context window. Coverage placed it first on a front-end coding leaderboard and third on the Artificial Analysis intelligence index at release. K3 was one of several 2026 releases from Chinese labs that competed at or near the frontier on public leaderboards, a pattern that had shifted the competitive picture well beyond a two-country race between a handful of American labs.ThursdAI — July 2026 model releases

Writing in Science, the Innovative Genomics Institute team used structure- and evolution-based models to design about 2,000 variants of TnpB, a compact relative of CRISPR enzymes. Nearly one in four cut DNA in the lab, some outperformed the natural enzyme, and they worked in bacterial, plant and human cells. One very divergent design became the first AI-designed RNA-guided nuclease with a solved structure.Innovative Genomics Institute →

Thinking Machines released Inkling, reported as a 975-billion-parameter model with 41 billion active parameters, trained on 45 trillion multimodal tokens and published under an Apache 2.0 licence. A permissively licensed model at that scale narrows the gap between what is available to a well-funded lab and what any organisation can run on its own hardware, which matters most to users who cannot send their data to a third-party API at all.ThursdAI — July 2026 model releases

OpenAI released GPT-5.6 as three separate tiers rather than a single model: Sol as the flagship, Terra as a cost-efficient option and Luna for speed. Coverage noted that Sol reached over 750 tokens per second when served on Cerebras hardware, and that the release introduced an Ultra mode with a maximum reasoning-effort setting. Splitting a frontier release into cost and latency tiers continued a pattern visible across the major labs through 2026, in which the choice facing developers is less which model is best and more which point on the cost-latency curve a given task justifies.ThursdAI — July 2026 model releases

xAI released Grok 4.5, described as a 1.5-trillion-parameter mixture-of-experts model trained in part on real agent traces from a coding tool. Coverage reported 83.3% on Terminal-Bench 2.1 using roughly a quarter of the output tokens of a competing model. Training on records of agents doing actual work, rather than on curated static datasets, became a recurring theme in 2026 releases, and it tied model quality more directly to whoever had a large deployed agent product generating those traces.ThursdAI — July 2026 model releases

Reuters reported that John Jumper, co-creator of AlphaFold and a 2024 Nobel laureate, is leaving Google DeepMind to join Anthropic. The move shifts one of the most prominent figures in AI for science between two frontier labs. It is a notable signal in the contest among leading labs to attract the researchers driving scientific applications of AI.Reuters →

Clinicians used an OpenAI reasoning model to re-examine 376 previously unsolved rare genetic disease cases in children and surfaced candidate leads for 18 diagnoses. The model worked through tangled genetic and clinical evidence that had defeated earlier analysis, pointing specialists toward specific avenues to confirm. Each lead still needs clinical validation, but the exercise shows reasoning models helping with diagnostic dead-ends that affect real families.OpenAI →

OpenAI reported that GPT-5.4, connected to Molecule.one's Maria AI and an automated lab platform, identified an additive that raised Chan-Lam coupling yields for more than 80% of the substrates tested. The system reasoned over the chemistry, proposed the change and validated it experimentally rather than only suggesting candidates on paper. It is one of the clearer demonstrations of a language model driving a measurable improvement in a real laboratory reaction.OpenAI →

OpenAI introduced LifeSciBench, an expert-written and expert-reviewed benchmark for messy, real-world life-science research work. It tests whether AI systems can interpret evidence, reconcile conflicting results, design experiments and weigh translational risk rather than answer tidy exam questions. The benchmark gives the field a way to measure progress on the kind of research that actually moves biology forward.OpenAI →

OpenAI released a public document outlining how it plans to ensure AGI benefits humanity broadly — covering safety commitments, access policy, governance structure, and the rationale for its scale. The document coincided with the company's confidential S-1 filing, positioning its commercial trajectory as aligned with public-benefit goals. It is a statement of intent rather than binding policy, but frames OpenAI's public argument for continued growth in the lead-up to a potential IPO.OpenAI — Built to benefit everyone

Anthropic published a research post titled 'When AI builds itself', disclosing that Claude had authored more than 80% of the code merged into Anthropic's production codebase in May 2026 — a significant milestone in AI self-development. The post also called for the creation of a global AI pause mechanism, suggesting that as AI systems become more capable of self-improvement, the ability to halt development rapidly could become a critical safety requirement.Anthropic — 'When AI builds itself' VentureBeat — Anthropic Claude writes own code Tom's Hardware — Claude 80% code milestone

NVIDIA released Nemotron 3 Ultra, a 550-billion-parameter open-weights model (55B active) built on a hybrid Mamba-Transformer mixture-of-experts architecture with a one-million-token context and post-trained for long-running agents. Shipped with open weights, training data and recipes, it scored 48 on the Artificial Analysis Intelligence Index — the highest of any US open-weight model at release.NVIDIA — Nemotron 3 Ultra for long-running agents →

At Build 2026, Microsoft released MAI-Thinking-1, its first in-house flagship reasoning model — a sparse mixture-of-experts system with a 256K-token context that the company said matched Anthropic's Claude Opus 4.6 on the SWE-Bench Pro coding benchmark. It headlined a launch of seven new MAI models spanning text, image, voice and transcription, marking Microsoft's deepest move yet into building its own frontier stack.Microsoft AI — Launching seven new MAI models →

NVIDIA and Microsoft unveiled RTX Spark, a unified superchip combining a Grace CPU and a Blackwell GPU with up to 128GB of unified memory and around one petaflop of AI performance. NVIDIA said it could run a 120-billion-parameter model with a one-million-token context entirely on-device, bringing frontier-scale inference to laptops and compact desktops with no cloud round-trip.NVIDIA — NVIDIA and Microsoft reinvent Windows PCs with RTX Spark →

Microsoft AI launched MAI-Image-2.5, an update to its in-house image model that entered the public Arena leaderboard at No. 3, with stronger text rendering and a focus on commercial, product and branding imagery. It was Microsoft's strongest showing yet with a self-developed image model rather than a partner system. The result signalled Microsoft's growing independence from third-party image models and intensified competition at the top of the text-to-image field, where the gap between leading systems continued to narrow.Microsoft AI announcement

Thinking Machines published a demonstration of its Interaction Models, an AI system designed for natural, real-time voice conversation. The model can wait silently when it should, interrupt politely when it should, count discrete items in view, observe a user's posture from camera input, and redirect requests it judges to be unsafe. The release is the lab's first major public capability demonstration and adds a third credible entrant alongside OpenAI and Google in the real-time multimodal interaction race.Thinking Machines blog

Google DeepMind introduced AI Pointer, a cursor that interprets on-screen selection plus spoken instruction as a single intent. Users can highlight content — text, image regions, list items — and tell the system what to do with it, removing the need to switch between chat sidebars and direct manipulation. The approach reframes how operating systems expose AI affordances: rather than a separate assistant window, AI becomes part of the cursor itself.DeepMind blog

World Labs released an open-source tool that converts a single photograph into a navigable 3D environment, complete with physics, movable objects, lighting and ambient audio. The release significantly lowers the barrier for hobbyists and indie developers to produce playable 3D scenes from reference photography. Single-image-to-scene with physics and audio is a key step toward AI-generated game and simulation content, and the open-source release accelerates downstream tooling that previously required full studio pipelines.World Labs

Liam Price, who is not a professional mathematician, got a proof of Erdős problem #1196 from a single prompt after about 80 minutes of model reasoning. Number theorist Jared Lichtman, who works on primitive sets, praised the argument as unusually clean. It showed the open-problem lists were now reachable by people outside academic maths.Physics World overview →

A national navy and a state underwater-archaeology unit used the ROV C 4000 to survey Camarat 4, a 16th-century merchant wreck found at 2.5 km in the Mediterranean. The tethered robot captured roughly 86,000 images at up to eight per second, and lifted ceramic jugs without disturbing the surrounding debris field — among the deepest objects ever raised from a wreck in the country's waters. The imagery feeds a 3D model of a vessel type that is poorly documented in surviving 16th-century texts and signals a wider shift toward non-invasive deep-sea archaeology. At 1.5 miles down the ROV operates under nearly 150 atmospheres, where conventional human dive equipment fails entirely.CBS News ↗

Princeton engineers built a soft-rigid hybrid robot that moves without a motor, gearbox or pneumatic line by combining a printable liquid-crystal-elastomer polymer with flexible electronics and origami folding. A demonstration crane-shaped robot flaps its wings on electric current alone, with targeted heating in the polymer doing the work an actuator would normally do. The paper — Bershadsky, Davidson, Paulino and Zhao, "Digital Actuation Control of Soft Robotic Origami With Self-Folding Liquid Crystal Elastomer Hinges" — appeared online in Advanced Functional Materials on 21 March 2026. Removing the motor cuts part count and failure modes, opening applications in medical devices, search-and-rescue and inspection robotics.Princeton Engineering ↗

Coinbase released Agentic Wallets alongside the x402 payment protocol, giving AI agents the ability to hold, send, and receive cryptocurrency autonomously without human approval for each transaction. For the first time, an AI agent could pay for API calls, purchase compute, or settle invoices as part of a workflow — creating the financial infrastructure layer that autonomous agents need to act independently in economic contexts.Coinbase — Introducing Agentic Wallets →

The Robotics & AI Institute, led by Boston Dynamics founder Marc Raibert, revealed Roadrunner — a 15 kg bipedal robot whose feet double as wheels, switching between side-by-side and in-line skating modes plus stepping on the same hardware. A single learned control policy handles every locomotion mode, and behaviours such as standing up from the ground and balancing on one wheel were deployed zero-shot. Roadrunner is positioned for logistics and warehouse use where wheels save energy on flat ground and legs handle obstacles. The release continues a 2026 trend of multi-modal locomotion — wheels plus legs — replacing pure bipedal designs in commercial robotics research.RAI Institute ↗

Lightricks released LTX 2.3, a 22-billion-parameter diffusion transformer model that generates synchronised video and audio in a single forward pass. The model supports resolutions up to 4K at 50 frames per second, marking a significant leap in real-time media generation quality.Lightricks — LTX-2.3 model card →

The peer-reviewed Evo 2 paper from Arc Institute, Stanford and collaborators describes a 40-billion-parameter model trained on about 9 trillion DNA base pairs that can read up to a million bases at once. It predicts the effect of mutations, including in BRCA1, and can generate genome-length sequences. It became the base for much of the year’s DNA design work.Nature paper →

Following the 2024 University of Sheffield test fire of what was described as the world's first AI-designed rocket engine, LEAP 71 reported hot-firing two 20 kN orbital-class methalox engines — a conventional bell nozzle and a full-scale aerospike — going from specification to first flame in under three weeks. The engines were generated autonomously by Noyron, the company's Large Computational Engineering Model, and 3D-printed in copper. LEAP 71 described the tested engines as roughly a tenth of the thrust class it plans to hot-fire in 2026, with manufacturing validation underway on 200 kN and 2,000 kN designs.Sheffield ↗

Boston Dynamics demonstrated a newly redesigned Atlas humanoid robot powered by neural Large Behavior Models from Toyota Research Institute. The robot performed complex multi-task sequences with self-correction, learning control policies without hand-coded routines. Boston Dynamics has deployed over 500 robots with revenue exceeding $130 million.Boston Dynamics — Large Behavior Models and Atlas find new footing →

DeepMind released AlphaFold 3, expanding beyond protein structure prediction to accurately predict protein-DNA, protein-RNA, and protein-ligand interactions. The model achieved 50% accuracy improvement over AlphaFold 2 and contributed to structural understanding underlying the 2024 Nobel Prize in Chemistry.Google DeepMind — AlphaFold 3 predicts the structure and interactions of all of life's molecules →

Google introduced Gemini 1.5 Pro, a model capable of processing a context window of up to 1 million tokens. This enables the model to work with entire books, lengthy video transcripts, and massive code repositories in a single prompt.Google — Our next-generation model: Gemini 1.5 →

DeepMind released GraphCast, a graph neural network model that predicts weather globally at 0.25-degree resolution in under 1 minute. The model outperformed the operational system of a leading intergovernmental medium-range forecasting centre on 90% of evaluated meteorological variables, producing predictions that took traditional systems 10 minutes to calculate.Google DeepMind — GraphCast: faster, more accurate global weather forecasting →

The Technology Innovation Institute (TII) released Falcon 180B, a 180-billion parameter open-source language model trained on 3.5 trillion tokens. At release, it was the largest openly available language model, surpassing Llama 2 on multiple benchmarks including MMLU, LAMBADA, and HellaSwag.Hugging Face — Falcon 180B is here →

The BigScience collaborative initiative released BLOOM, a 176-billion parameter open-source language model trained across 46 natural languages and 13 programming languages. At the time of release, it was the largest openly available language model in existence.BigScience — Introducing BLOOM, the largest open multilingual language model →

DeepMind's AlphaFold 2 solved the protein folding problem, predicting 3D protein structures to near-experimental accuracy at the CASP14 competition. The breakthrough came from combining attention mechanisms with evolutionary biology insights. The achievement later contributed to the 2024 Nobel Prize in Chemistry awarded to David Baker, Demis Hassabis, and John Jumper.Google DeepMind — AlphaFold: a solution to a 50-year-old grand challenge in biology →

Google researchers published "Attention Is All You Need," introducing the Transformer architecture built entirely on attention mechanisms. This paper became one of the most cited in machine learning, fundamentally changing how neural networks are designed.Google — Attention Is All You Need (arXiv) →

Ian Goodfellow and collaborators introduced Generative Adversarial Networks, a framework where two neural networks compete—one generating data and one discriminating real from fake. This sparked a revolution in generative modeling and unsupervised learning.Goodfellow et al. — Generative Adversarial Networks (arXiv) →