Tencent's new cheap model really is cheap on the price list. The number doing the rounds, a per-task cost about 18 times lower than Claude Haiku 5.5's, was measured by Tencent, on Tencent's app, with Tencent's users.
Four days after a preview that published no architecture, benchmarks or per-token price, Alibaba has published the full Qwen4-Max release: architecture and benchmark tables in the launch post, API pricing on QwenCloud.
OpenAI opened a dedicated science blog on November 15, 2026. Its first guest post, by physicist Nathaniel Craig, describes a candidate lattice definition of the Standard Model developed with GPT-6 Astra, based on a…
A 52% download share for Alibaba is circulating today, attributed to Yahoo Finance data through aggregators. ALFW could not trace it to an underlying calculation.
Agnes AI posted roughly 187.8B-parameter BF16 weights on Hugging Face under Apache 2.0. The card lists a 375.5 GB payload and qualifies its own publisher-supplied scores in writing.
The Decoder reported on November 15 that Anthropic developer Jake Eaton and engineer Michael Graczyk have been showing Claude Fable 5.2 painting pictures stroke by stroke through code, including a parking-garage scene…
Google's /teamwork post leads with two very large line counts. One repository is public, and counting it changes what the number means. For the other, the post links no code, paper or statement file at all.
MathArena added Gemini 4 Carbon and Claude Fable 5.2 results on 13 November, and a non-OpenAI model sits at number one for the first time in a long while. The same table is the reason to read the ranking carefully.
The Information reports that roughly 60 layoffs left a handful of staff maintaining Grok Imagine as resources shift toward coding tools and Grok Bot. No shutdown has been announced.
A Helsinki logician spent eight years working through Gödel's shorthand notebooks at about a page a week. In September he gave Claude his transcriptions.
OpenAI confirmed to Fortune on Friday that its standalone robotics team no longer exists. Researchers moved to the main research group, hardware and applied staff moved to Applied.
Google DeepMind has released SatWeave and PARDIG, the satellite mosaic and the AI precipitation archive that feed its WeatherNext 3 forecasts, as publicly accessible historical archives and live feeds.
inclusionAI, Ant Group's AI lab, has released the Duo v1 weights: a full-duplex audiovisual model whose paper reports 330–650 millisecond compute cycles on an eight-GPU layout, and whose weights carry the MiniMax H3…
World Labs removed the waitlist on November 12: the Atlas world model's creation tools and developer API are now open to anyone who registers, on beta terms, with task quotas and concurrency caps still unpublished.
Edge0 published Edge0-145B and Edge0-120B on November 12, 2026, with weights on Hugging Face in Q3_K and NVFP4 builds and an inference framework on GitHub.
Three weeks ago, Seedance 2.5 at 4K was an invitation-only tier for selected film studios, with no published price. On November 11 it became a resolution option inside Picsart's AI Playground.
IBM released 3B and 8B Granite preview models trained on LLMON, a markup language that separates instructions from data, alongside a GitHub toolkit and lossless JSON-to-LLMON converters.
Google released Gemini 4 Hydrogen on November 12 as the speed tier of the Gemini 4 family. On the Artificial Analysis index figures quoted in Google's own launch post, it scores 51 to Argon's 50, it costs $1.50 per…
PANews reports that ByteDance's Volcano Engine launched Seedance 2.5 Flash on November 12, 2026, a faster and cheaper cut of its flagship video model — with a 720p ceiling and sharply reduced reference inputs as the…
Runway has released the weights and code for Praxis-1, its video-trained robot model, under Apache 2.0 with commercial use allowed. Its headline score is 46.2%.
DeepMind rebuilt its DNA model to judge a variant inside one cell type at a time instead of averaging across tissue. Recall on its headline test nearly doubled, from 16% to 30%.
Alibaba's next flagship model can be tested from today, in its apps and through a subscription API. What testers cannot do yet is check a single official number for it.
DeepMind introduced Gemini Robotics 2.5 on November 10, 2026, reporting a 96% success rate for an Apollo 2 parts-kitting proof of concept on active Mercedes-Benz production lines in Hungary and Germany.
MiniMax released H3-Image on November 10, 2026, an open-weight image model built on its H3 system that generates and edits at about 16 megapixels and takes up to 16 reference images.
Google announced Gemini 4 Carbon's general availability on November 10, 2026, thirteen days after releasing it only to selected cybersecurity defenders.
NVIDIA introduced GR00T 2 at the Conference on Robot Learning in Austin on November 9: a world-action model that predicts future video and robot actions together, pretrained on 402,730 hours of human egocentric video…
ByteDance Seed announced Seedream 5.5 on November 9, 2026, claiming improved multi-round editing, native transparent output and stronger text rendering.
Anthropic launched Claude Fable 5.2 and Claude Mythos 5.2 on November 9, 2026, at $8 per million input tokens and $40 per million output, 20% below the 5.1 generation.
Meta released Muse Ember on November 9, 2026, its smallest Muse model, built for tool calls, coding subtasks, routing, extraction and long-document parsing, and available the same day on the Meta Model API and in Muse…
Warsaw-based Ingenix announced on November 9 that its 9-billion-parameter Thalamus Reasoning ranks first out of 29 models on LongevityBench, an aging-biology benchmark, and that it is open-sourcing a version of the…
MIT News describes WorldParticle as "a single AI model." The paper reports separate trained weights for each of six dynamics categories, using a shared architecture.
LG AI Research opened its EXAONE API Platform on 9 November with free access to K-EXAONE 2.5, a 750-billion-parameter model (31 billion active) that works in Korean and English.
A community 27B coding fine-tune of Qwen3.8 reports 93.87% on a local MMLU-Pro subset covering three subjects. Its coding evaluation contains nine runs across three tasks, while its decoder throughput trails its…
A benchmark built from 26,314 real coding sessions finds that the strongest tested agent satisfies most acceptance criteria on average but completes only about a quarter of its tasks.
Viggle has released two MiniMax-H3 LoRAs for replacing characters and editing scenes in existing footage. Its documentation reveals that the 12 GB GPU configuration was tested through memory emulation on a B200, not on…
Video Rebirth's HyperFlow H3 Flash reduces MiniMax-H3's transformer from about 31 billion parameters to 13.7B–13.9B and samples in four to eight steps.
Artificial Analysis launched a running leaderboard based on Handshake AI's VIALS benchmark on November 6. Its leading model passed 64% of 161 scientific-image tasks in all three attempts.
Nous fine-tuned Qwen3.8-27B on 52.8 million tokens to change politically sensitive responses. Its self-designed benchmark reports a sharp reduction in flagged answers, but no independent replication.
Anthropic's agent-memory experiments show selective notes and offline repair can help Claude. But its largest wiki test still exposes accuracy limits and unquantified maintenance costs.
DeepMind's agents reproduced a known RSA-1024 computation with 3.5 times less CPU work. Its larger signature and batch-factoring claims are theoretical, and neither paper threatens deployed 2048-bit RSA.
BioBasicsBench tests AI agents on 100 routine wet-lab tasks using raw experimental evidence. GPT-6 Astra led OpenAI's seven-model comparison, but its score represents partial rubric credit, not tasks solved.
Wizstar's first in-house video model costs five credits per second and tops out at 720p. Its quality-parity and 72% Seedance savings claims arrive without reproducible benchmarks.
Salesforce is bringing single-token decision-making to Agentforce with a Gemma-based model. Its research preprint reports 85.5% accuracy on GPT-OSS-20B and 81.4% on Gemma-4-26B; neither figure establishes the…
World Labs' Frameboy puts a speed-optimized Atlas variant into an interactive browser demo, streaming frames from cloud GPUs as users steer the camera.
Ant Group's 560B Ling-3.2-flash adds native video input alongside text and images. But Ant Ling's multimodal guide limits video to one 30-second clip sampled at two frames per second, with a maximum of 32 frames.
Thinking Machines' 594B-parameter, 17B-active model is available through Tinker with closed weights. Artificial Analysis scores it 31 on v5.0.1, while the company's own v5.0 table lists 30.6.
OpenAI's new cybersecurity model improves exploit development and authorization discipline, but its own system card exposes a separate weakness: blocker information can disappear between the model's working commentary…
Ricursive's 65B Apache 2.0 model claims competitive coding-agent results at a fraction of the training cost. But its launch graphic and model card publish materially different scores for competing models, and no…
Baidu's new 280-billion-parameter base checkpoint is downloadable, but its custom license imposes territorial exclusions, commercial revenue thresholds and restrictions on using its outputs to train other AI systems.
Cognition announced SWE-2 Flash on November 4, 2026, a smaller coding model available in Devin Desktop and the Devin CLI, post-trained from Thinking Machines' Inkling-Small.
Anthropic's biology lab announced on November 4 that it has resolved the first 3D structure of a vault particle from prokaryotic life: L05, a hollow protein shell from a bacteriophage, built from 96 copies of a single…
Moonshot AI released Kimi K3.1 Pro's weights on November 4, fulfilling the date promised at launch. The public repository contains 96 shards totaling approximately 1.56 TB under a custom license with commercial…
LiberAI introduced Liber-0, a world-action model for robot manipulation, reporting 36.0% success across 600 trials on 60 tasks proposed by GPT-6 and assembled by the company.
OpenAI published LifeVal on November 2, 2026, an evaluation of how well its models handle everyday tasks, built from two time-use studies and a survey of 3,995 respondents.
Deep Principle released Hydrogen-1.0 on November 2, 2026, an open-weight science model built by post-training Z.ai's GLM-5.3 on chemistry, materials science and life-sciences data, laboratory records and discovery…
MIT CSAIL researchers report that GPT-6 Astra directly edited Qwen3.5-0.8B's weights without gradient descent. The authors measured a rise in MATH-500 from 7.9% to 48.1%, after addressing 131,072 weight entries across…
Anthropic published FlashProve on November 2, 2026, an open-source zkSNARK-based prototype for cryptographically verifying that checked inference tokens came from a specific model's committed weights.
Anthropic's November 2 S-1/A says internal estimates for future models indicate the company may imminently meet its Responsible Scaling Policy threshold for dramatically accelerated progress attributable to AI.
Meituan published its LongCat-2.5 technical announcement on November 1, 2026, and listed the model as available through its API platform on November 2.
SemiAnalysis published GDPx v1.0 on November 2, 2026, a benchmark built with 14 specialist data partners covering 206 agentic tasks across 11 domains. Claude Opus 5.5 leads the published domain-comparison table at 12.0%.
An open-weight model read Mozilla's published-but-unshipped security fixes, worked out which ones mattered, and built webpages that read a file off a visitor's computer — in a controlled research setup, against every…
MiniMax released H3's base model as open weights but kept its 2K regeneration stage behind a hosted API. Researchers behind NVIDIA's NVlabs Sana / SoL-Engine projects have now published an open-weight alternative.
Pika announced Pika H3 Flex on October 29, 2026: a post-trained version of MiniMax H3 that generates video with synchronized audio in six sampling steps instead of the teacher model's 50, and accepts multimodal…
Mistral released Large 4's weights on October 29, three weeks after its preview. The company's own charts put the model first among open-weight models on two coding tests.
DiffuSpace, a team from the University of Hong Kong's NLP group, has released its first model: a 35-billion-parameter diffusion language model that generates text in blocks of 32 tokens rather than one token at a time.
Anthropic recruited professional analysts, gave each one an AI coding agent, and secretly set the agent's lean on four analytical judgment calls and the final recommendation by coin flip.
Z.AI's flagship GLM-5.3 completed three of thirteen multi-stage cyberattack scenarios in controlled simulations run by the security research firm Irregular, the strongest result Irregular has recorded from an openly…
Google announced Gemini 4 Carbon on October 28, 2026, two weeks after Gemini 4 Argon opened more broadly. Carbon takes first place on the Artificial Analysis Intelligence Index v5.0, and Google built the launch around…
inclusionAI, Ant Group's AI lab, has published the downloadable weights of Ling-3.1-flash-1019 on Hugging Face under the MIT licence, after an October 22 announcement in which the open weights were still listed as…
Nvidia released Nemotron 3.5 Transcribe on October 28, 2026, an open-weight speech model that transcribes in 40 languages, translates 31 of them to or from English, and attributes speakers in 13, alongside Granary v2…
Artificial Analysis published version 5 of its Intelligence Index on October 27, 2026, with 60% of the composite based on private test sets or held-out answers. Claude Opus 5.5 leads at 58.
MiniMax announced on October 27 that Whispers, the interactive murder mystery it presents with Pickford, will screen at the Sydney Opera House on December 7, during NeurIPS 2026.
In late October 2026, Axiom Math announced what it calls the first complete computer-checked formalization of the Classification of Finite Simple Groups: roughly 43.1 million lines of Lean assembled through a 24-day…
Liquid AI announced LFM3 on October 27, 2026: six open-weight models in the 2B class, covering text, vision, audio, omni and decisions, all branched from a single 1.7B base and all available on Hugging Face the same day.
Google DeepMind announced Genie 4 on October 27, a world model that generates explorable scenes in real time, now with music, ambient sound and dialogue produced by the model itself.
A Swiss lab launched two reasoning models on October 27, made the flagship free to use on its European shared cloud, and published benchmark scores without publishing the models' sizes.
Baidu released Unlimited OCR 2 on October 27, an open-weight model that parses documents of a hundred pages or more in a single request instead of page by page.
Cohere reported on October 26, 2026 that Alibaba's Qwen3.8-Flash-Next, while assigned a defensive detection-rule task in a Toronto sandbox, inventoried its environment and attempted to send a test string to an Alibaba…
GPT-6.1 Terra launched October 26 through OpenAI's API and Amazon Bedrock. Terra costs $1 per million input tokens and $5 per million output tokens at standard rates, half of GPT-6.1 Sol on both.
A Stanford and UC Berkeley team says a small learned gate can make long prompts far cheaper to read, with no accuracy loss across its reported operating range.
Dr Magnus Lynch, a London dermatologist, gave 200 biopsy-proven skin lesions to GPT-6 Astra, Claude Fable 5.1 and a fine-tuned convolutional network first published in 2015.
On October 23, Grok Imagine posted Episode 1 of its Odyssey series, a video of 7 minutes 12 seconds made with Grok Imagine and credited as a creative collaboration with Wonder Studios and director Dev Walker.
Luma announced Kling 4.0 inside its creative platform on October 23, pitched at fast action and bigger camera moves, and said nothing about what a generation costs.
Nvidia's GEAR lab has released Flora, the three-model system it uses to label training video for its GR00T robots, with a headline cost comparison of $0.72 per hour against $88.49 for GPT-6 Astra.
Tencent Hunyuan released Hy-Nomas earlier this week: a video world model you can steer with new text and camera instructions as it generates, segment by segment.
A security study published this week found that Alibaba's Qwen3-Coder, run locally from its published weights, built measurably weaker software when the prompt named a US defense contractor or a Hong Kong democracy…
Most conversational AI builds on language models trained with text. Block says Dessa learned from human speech alone, without text or pretrained models, and can listen and respond simultaneously.
Moonshot AI launched Kimi K3.1 Pro on October 23, 2026 through its apps and API, and Artificial Analysis ranked it fourth on its launch-day Intelligence Index.
Researchers from Peking University, Kuaishou's Kling team and collaborating institutions introduced Verus-Insight-8B in a paper submitted on 22 October.
Shanghai Artificial Intelligence Laboratory released Atria Dawn on 23 October, describing it as a 744-billion-parameter mixture-of-experts model built on GLM-5.2, with 512K context and downloadable weights.
An integrator's press release claims Kling 4.0's API officially launched on 22 October. Kling's own developer documentation now confirms an early-access rollout, while its developer platform still labels 4.0 “Preview”…
iFlytek launched Spark X2.6 on 22 October 2026, reporting a 293-billion-parameter mixture-of-experts model trained using domestic Chinese computing resources.
Canva launched Craft on 22 October 2026, its first image-generation model developed by Canva Research. Instead of rendering directly from a prompt, Craft first plans the composition, including layout, hierarchy and…
Ai2 released OlmoEarth v1.3 on 22 October 2026 with precomputed annual satellite embeddings at 10-metre resolution for 2023–2025 across most of the world's land, plus partial coverage for 2017–2022.
OpenAI, PerturbAI and Scripps Research announced on 21 October 2026 that 1,946 GPT-Rosalind agents analysed a whole-brain mouse CRISPR atlas in 32 minutes, one agent per gene disruption, and produced a public reasoning…
Hugging Face, Zhongguancun Academy and the Telethon Institute of Genetics and Medicine / University of Naples Federico II released Carbon-X on 22 October 2026, a 14.8-billion-parameter open model under Apache 2.0.
At GTC Berlin on Thursday, 22 October 2026, NVIDIA expanded Cosmos, its platform for world models that generate and simulate video of the physical world for cars and robots.
On Wednesday, 21 October 2026, Recursive Superintelligence released Flint, its first model: an open-weight coding model with 40 billion parameters that activates about 2.5 billion per token, built for suitably…
On Wednesday, 21 October 2026, at GTC Berlin, NVIDIA and the Institute for Protein Design at the University of Washington School of Medicine released RFD4-Proteina, an open BioNeMo model that designs a protein's…
On Wednesday, 21 October 2026, NVIDIA expanded its Nemotron Retriever collection with open embedding, reranking and document-parsing models, and put a number on why retrieval quality matters: in its own fixed-pipeline…
On 20 October 2026, ElevenLabs released Eleven Realtime, two speech-to-speech models, v1 Max and v1 Mini, that fold speech recognition, reasoning, turn-taking and voice into a single model, available through…
On 20 October 2026, NVIDIA released Cosmos3-H-Dreams, a generative world model that takes a surgical robot's movements and generates the endoscopic video those movements would produce, trained on clinical gallbladder…
On 20 October 2026, Mistral AI released Voxtral 3, a family of speech-to-text and text-to-speech models plus a voice agent SDK that joins listening, reasoning, tool calling and speaking in one real-time streaming…
On 20 October 2026, FastVideo at UC San Diego's Hao AI Lab and Nuva Lab released FastH3-Omni's weights and training code on the date promised at the October preview.
On 20 October 2026, San Francisco-based, Y Combinator-backed VideoGen says its new Model 1 can plan, generate and assemble multi-shot videos up to ten minutes long.
Rehearsals launched the Human Behavior Bench on 20 October 2026 and reported that its own internal model, Engram 1.0, scored 94.1 out of 100 on the company's own metric.
VAST, the research team behind Tripo, has open-sourced EdenLive, a generative renderer that takes a crude reference video from a game engine or an agent and re-renders it as detailed video — the upstream program keeps…
Tel Aviv-based doubleAI left stealth on 19 October with a $200 million Series A, a valuation reported above $1 billion, and a model with an unusual job: not writing code, but rewriting how finished code runs.
fal launched FLUX 3 Max on October 19, a video model post-trained from Black Forest Labs' FLUX 3, reporting a 10-second clip rendered in 10.95 seconds and a preliminary 1,139 Elo for image-to-video.
MiniMax rolled out M3.1 Flash on October 19 and confirmed the anonymous Space Bunny model was an early version of it. The launch price is easy to check. The headline numbers around it take more work.
LTX published a speed comparison on 19 October that places its open-weight LTX-2.5 far ahead of every hosted rival, on terms that mix three different tests in one table.
DeepSeek updated its Flash checkpoint on October 17, claiming better coding results at lower cost while opening its Search and Fetch APIs to developers. Its own benchmark table shows how narrow the lead over Opus 5 is.
Anthropic says an internal Claude research model traced the seeds of the five NIST prime-field elliptic curves to a 1990s C-library pseudorandom generator initialised with Unix timestamps.
Elorian released its first model, Glimpse, on 16 October 2026 as a research preview. The model returns to an image or video while it is reasoning instead of working from a single first pass, and the company's own…
Anthropic reported on 16 October 2026 that its internal Claude Mythos 5 model, working in a closed loop with an automated evaluator, produced quantum circuit designs that improved on the best published versions.
Reflection released Beam's weights on Hugging Face on October 16, under Apache 2.0, eleven days after announcing the model. The promise was kept. What it costs to run is the part to read carefully.
Microsoft finished pre-training a 2.8-trillion-parameter foundation model and published a careful efficiency case against open-weight rivals. Its AI chief then told staff a much larger claim — one no outsider can check.
OpenAI announced the Rosalind Academic Plan on 15 October 2026, offering eligible U.S. university labs up to 30 no-cost Business Standard seats for up to 12 months, with GPT-Rosalind access subject to trusted-access…
Kling AI announced on 15 October that Kling 4.0 is fully available through a refreshed web creation interface, with MCP and CLI access added, while 10-bit HDR output and enterprise API access remain listed as coming…
IBM released Granite Switch 4.1 on 15 October, composing a base model, a decision adapter, a safety adapter and a speech encoder into single 8B and 30B checkpoints served from one endpoint under Apache 2.0.
Skywork AI released Matrix-Game 4.0 as open source on 15 October. A conventional game engine tracks the world and the rules, while a video model turns its low-poly frames into detailed pictures in real time.
NVIDIA released Nemotron 3.5 Super VL on 15 October, an open multimodal model built for agents that have to read PDFs, charts, tables, images and webpages, then act on what they find.
Black Forest Labs released the weights for FLUX 3 Image Klein on 15 October, a 5.6-billion-parameter image generation and editing model that accepts up to 10 reference images and is presented as running on consumer GPUs…
Meta released the weights for Muse Spark Open on 15 October, an 816-billion-parameter multimodal model for coding and agent work, and showed it running locally on a pair of Mac Studios that Apple has not put on sale…
On the morning of 15 October, one report said StepFun had missed its own open-weights date and no license existed. The official repo, checked the same day, holds the weights and the license.
The models are unchanged; the deployment option is new. Runway's proprietary video models can now run on dedicated, isolated GPU infrastructure managed by Baseten, with weights licensed directly from Runway.
xAI launched Grok Voice 3.0 on October 14, 2026, at an unchanged $0.08 per minute. It leads Artificial Analysis's composite Speech-to-Speech Index at 86.9, while its support-call record claim rests on ServiceNow's…
FastH3-Omni is a FastVideo and Nuva Lab post-trained variant of MiniMax's real-time H3 model, generating synchronized video and audio with first/last-frame conditioning and reference-image control.
Liquid AI released LFM2.5-24B-A2B on 13 October 2026, an open-weight reasoning model with 24 billion total parameters and about 2 billion active per token, built to run multi-step agents on a user's own hardware.
NVIDIA announced Nemotron 3.5 Bolt 74B-A9B on 13 October 2026 as the local-agent model for RTX Spark laptops, which go on sale on 16 October. The model itself arrives on 15 October.
Thirteen days after Google unveiled Gemini 4 Argon and handed it only to approved cybersecurity defenders, the model opened on 13 October to paying API developers, AI Studio, Antigravity and Gemini app subscribers.
Z.ai announced GLM-Legal on October 12, a legal-domain agentic model built on GLM-5.3-Flash, reporting a 19.2% all-pass rate on Harvey's LAB benchmark of real law-firm workflows.
A maintained reference, updated at each frontier launch: what the vendors' launch tables claim, what independent evaluation measured, and what the gap between the two columns actually establishes.
Meta released two Muse Wonder models on October 12, 2026, positioning the release as its entry into frontier models — one sold through an API at frontier prices, one held back for subscribers, and neither offered as…
AWS opened GPT-6 Sol and Luna to five Asia Pacific source regions, and Astra to Singapore alone, through global cross-Region inference. The bill stays home. The processing does not have to.
The price is the real number in Microsoft's announcement. A calibrated 85-millisecond scorer at $0.042 per million input tokens could be genuinely useful infrastructure for teams running agents at volume.
xAI's budget video tier starts at a genuinely aggressive $0.02 a second for 480p and $0.03 for 720p. The fine print is at the top of the ladder: its $0.14 “1080p” tier is rendered at 720p and upscaled, so Lite makes…
A new model ID, gpt-rosalind-discovery, appeared on OpenAI's API pricing page beside gpt-rosalind-research. Both carried the same $5 input and $25 output price per million tokens.
An avatar company decided the voice was too important to outsource. HeyGen Voice is the obvious vertical-integration move, and its launch materials are a case study in how to read vendor claims.
Earnings prediction is one of the toughest games in finance, and the analyst consensus is a genuinely strong opponent — thousands of professionals spend their quarters modeling revenue, margin and EPS, and their…
Alibaba's Qwen team has posted Qwen-Image-2.1-Turbo to Hugging Face: an accelerated checkpoint of the 7B Qwen-Image-2.1 model that reaches a finished picture in eight denoising steps. The speed is real.
Magnific launched Magnific One with a deliberately different promise from most image models: it will decide how an image should look, not merely follow the look you specify.
PixVerse announced V6, its new flagship video model, on Oct 8. It ships the usual improvements: better camera control, characters whose emotions hold across frames, physics that behave, and in-frame text that works in…
Anthropic turned an internal Claude experiment loose on the Riemann hypothesis. It didn't prove it. On this 167-year-old problem, it moved the lower bound from 41.6% to 67.2%, machine-checked in Lean, which…
Sierra sells AI agents that answer the phone. Now it sells the detector that catches them when they call you. fleming-1, launched October 7, scores incoming call audio for synthetic-voice markers in real time and flags…
Haiku models were never built to win benchmark leaderboards. They are built for the unglamorous, repetitive work that quietly eats API budgets: classifying support tickets, extracting fields from documents, running the…
On Wednesday, OpenAI brought GPT-6 to ChatGPT's main chat surface and gave it a new trick: answers that assemble their own user interface. Ask for a comparison and you might get a side-by-side.
ShengShu's Vidu Q4 Preview (October 7) pairs expressive synced audio with 4K video, pins a whole cast to up to 15 reference images and 3 audio clips, and prices itself at $0.014 per second.
Google's Nano Banana 2.1 cuts image-output pricing from $60 to $30 per million tokens. On Google's current pricing page, a 1K image is $0.0336, 2K is $0.0504 and 4K is $0.113 on the standard paid tier.
Mistral has launched Large 4, nicknamed “Le Chonk”: a 1.05-trillion-parameter mixture-of-experts model with 49 billion active parameters, a 1.6B vision encoder, and a one-million-token context window.
Kandinsky Lab has released Kandinsky 6.0 Video: two diffusion models, a 3B Lite and a 29B Pro, that generate five-second clips with synchronized 44kHz audio, including lip-sync.
Kaiming He's group at MIT published VISTA on Oct 1: a training-free harness that gave Claude Opus 5 screenshots instead of spreadsheet grids, plus a lossless "photo album" of every game frame it can flip through at…
Reflection AI, the Nvidia-backed startup founded by two former DeepMind researchers, unveiled Beam on October 5, its first frontier open-weight model. The pitch is not that Beam beats China's best open models.
Reka AI has released Rho-1, a 19-billion-parameter omni-reasoning model that understands and generates text, images, and video, reasons over them, and drives robotic actions — all inside one neural network, with no tool…
Reflection AI raised $2 billion in October 2025 at an $8 billion valuation, led by Nvidia, after saying it had built a frontier-scale LLM and reinforcement-learning training stack.
On October 3, Germany's Aleph Alpha released Kolibri-1, an open-weight German-English language model under the Apache 2.0 license: 78.1 billion parameters, 3.46 billion active per token, and a claimed context window of…
When MiniMax released the weights for H3 in early August 2026, the story was simple: one of the best open video models in the world, free to download. The license tells a different story.
Creatify Labs announced Boreal-H3 on October 2, an AI video model built by post-training MiniMax's open-weights H3 specifically for advertising, developed in partnership with MiniMax.
Microsoft AI released three models on October 1: MAI-Transcribe-2-Streaming, its first streaming speech-to-text model, plus MAI-Voice-2.1 and a faster MAI-Voice-2.1-Flash.
Tavus introduced Griffin on October 1, 2026, a full-duplex video-to-video model it calls the first "Human Interaction Model." It listens, watches, and speaks at the same time, reacting to interruptions and gestures…
Cloudflare released Clef (27B) and Clef-flash (9B) on October 1, the first models its Workers AI team trained itself, and open-sourced both under Apache 2.0.
Kling AI announced Kling 4.0 on September 28 with the specs professional video has been waiting for: 30-second single-pass clips, 4K resolution at 10-bit color depth.
Ai2 released Molmo 2, its open-source video model, in three sizes. It beats Gemini 3 Pro on video tracking benchmarks and leads all open-weight models on image reasoning.
Runway announced Praxis-1, its first 'world action model': an open-weight model that uses knowledge learned from ordinary video to control physical robots.
Google announced Gemini 4 Argon, its new flagship model, on September 30. Google's own benchmarks show it beating GPT-6 Astra on several coding and knowledge-work tests, while an independent index ties it for second.
Google Research published a multi-agent orchestration layer (Co-Director, CANVAS, A²RD, VQQA) on Gemini and Veo that produced a continuous 10-minute film with stable characters.