AI Models

Nvidia’s open Nemotron model transcribes 40 languages

Nemotron 3.5 Transcribe turns recordings into text, translates and labels who is speaking. Anyone can download it for commercial use, and Nvidia also released labels for 463,062 hours of speech.

Nvidia released Nemotron 3.5 Transcribe on Hugging Face on October 28. It is an open-weight model that writes down speech in 40 languages, and Nvidia says it translates 31 of them to or from English. It can tell up to eight speakers apart. Nvidia also published Granary v2, a large set of labels for training speech models.

Google’s Gemini 4 Carbon tops a new independent test

Google calls Carbon its strongest model, with top cybersecurity skills. Only some cyber defenders get it this week. Artificial Analysis already ranks it first, one point ahead of Anthropic’s Claude Opus 5.5.

Google announced Gemini 4 Carbon on October 28, two weeks after Gemini 4 Argon reached paying users. It goes first to a small group of cyber defenders and then to more trusted partners, Google says. There is no public app or API yet. Artificial Analysis, an independent testing firm, scores it 59 on its new Intelligence Index, the highest of any model.

Google’s Genie 4 builds worlds with sound you can steer

Google DeepMind’s new world model makes interactive scenes with sound and speech, and users can change them by typing. For now only trusted testers can use it, and nobody outside Google has tested its claims.

Google DeepMind released Genie 4 as a research preview for trusted testers on October 27. The model creates a world that people can move through, drawn frame by frame as they go, and change with text prompts. Google says Google AI Ultra subscribers and some developers will get it soon, but has given no date or price.

OpenAI’s GPT-6.1 Terra costs less and scores higher

OpenAI’s new mid-size model costs $1 per million input tokens and $5 per million output tokens, down from $2 and $12. The independent tester Artificial Analysis finds it scores 6 points more than GPT-5.6 Terra at less than a third of the cost per task.

OpenAI released GPT-6.1 Terra on Monday, October 26, in its API and for paying users of ChatGPT Work and Codex. Artificial Analysis ranks it 25th of 221 models on its Intelligence Index, with a score of 48. OpenAI’s claims about speed and its biggest model are not checked yet.

OpenAI says its AI proved a famous computing idea wrong

OpenAI says an internal model showed that the Exponential Time Hypothesis, a widely used assumption about how hard some problems are, is false. It also claims new proofs in maths and physics. No outside expert has confirmed any of them yet.

OpenAI announced the three claimed results on Monday, October 26, and put the computer science proof and a Lean version on GitHub. Three experts, two of whom first proposed the hypothesis, will lead an open check. They say their role does not mean the claim is right.

Moonshot launches Kimi K3.1 Pro, open weights due Nov 4

The Chinese company says its new model nearly matches Claude Opus 5.5 on a coding test. The independent tester Artificial Analysis scores it 55 points, fourth of 215 models, but the weights are not out yet.

Moonshot AI released Kimi K3.1 Pro on October 23, together with a faster version called K3.1 Pro Highspeed. Both work in Kimi’s apps and through its API. Moonshot says it will publish the full model weights by November 4, 2026, so anyone could then download and run the model.

Google Cloud starts offering OpenAI’s GPT-6 models

GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna are in a private preview for some business customers now. OpenAI says wider access follows in the coming weeks.

Google Cloud customers can buy OpenAI’s models through Google’s marketplace and pay with their existing cloud budgets. The models are also coming to Google’s Antigravity coding tool.

doubleAI says its Maxwell model makes code run faster

The startup of Mobileye’s founder says Maxwell rewrites programs while they run, so they use chips better. Its speed-ups are its own figures, and it also announced a $200 million round backed by Nvidia.

doubleAI, an Israeli startup led by Mobileye founder Amnon Shashua, has launched Maxwell, which it calls a “code execution model”. Maxwell watches a program while it runs and rewrites its machine code to use CPUs and GPUs better. It is open to early-access customers only.

MiniMax launches M3.1 Flash at half the price of M3

The Chinese company says its new agent model is nearly three times faster and 70% cheaper to use. Artificial Analysis ranks it fourth of 173 models, but measured a lower speed and a 93% hallucination rate.

MiniMax, a Chinese AI company, released M3.1 Flash on October 19. It costs $0.15 per million input tokens and $0.60 per million output tokens, half the price of MiniMax-M3. The independent tester Artificial Analysis gives it 42 points on its index, fourth of 173 models.

GoTo releases a free AI model for Indonesian and Javanese

The 27-billion-parameter model also handles Sundanese, Balinese and Batak Toba. It is built on Google’s Gemma 3 with AI Singapore, and its weights are free under Google’s Gemma licence.

GoTo, the Indonesian group behind the Gojek and Tokopedia apps, released GoTo-AI-Gemma-3-27B-IT on October 19. The model works in Indonesian, English and four regional languages, and anyone can download it from Hugging Face. GoTo says it scored highest on every regional language in its own tests, but it has published no scores.

DeepSeek releases an open update of its V4.1 Flash model

DeepSeek says the new version scores higher on coding and agent tests, finishes tasks faster and writes shorter answers. Anyone can download it under the MIT licence, and no outside group has tested it yet.

DeepSeek released DeepSeek-V4.1-Flash-1017 on October 17, a “minor update” to the V4.1 Flash model it launched on September 10. The weights are on Hugging Face, and the API price stays the same. In DeepSeek’s own table, the model scores 77.0 on the DeepSWE coding test, up from 74.2.

Microsoft finishes pre-training its 2.8T MAI-Base-2 model

The model has 2.8 trillion parameters and is not released yet. Microsoft says it beats every open-weight model it tested and needs 4.2 to 12.3 times less computing power than DeepSeek’s best base model; no outside group has checked this.

Microsoft AI said on October 15 that it has finished pre-training MAI-Base-2, its largest model so far. It is a mixture-of-experts model with 2.8 trillion parameters, of which 80 billion work on each token. The model is now in a second stage of training, and Microsoft gives no release date.

Meta releases Muse Spark Open, an 816B open-weight model

Meta calls it the most capable open-weight model from a US company. An independent test gives it 39 points, and in Meta’s own table two Chinese open models still lead on coding.

Meta released the weights of Muse Spark Open on October 15. The model has 816 billion parameters, of which about 42 billion work on each token, and it can read up to 1 million tokens at once. Meta worked with Apple so the model can run on two or four linked Mac Studio computers.

Google opens Gemini 4 Argon to paying users and developers

Since October 13, AI Pro and Ultra subscribers can use it in the Gemini app and in Search, and developers through the paid API. Free users don’t get it yet, and the launch price doubles on December 1.

Google released Gemini 4 Argon, its newest flagship model, to many more people on October 13, two weeks after it went only to cyber defenders. Subscribers to Google AI Pro and Ultra can use it in the Gemini app and in AI Mode in Search. Developers can use it through the paid Gemini API. Free users and AI Plus subscribers don’t get it for now.

Meta’s Muse Wonder models close in on the AI frontier

Meta released Muse Wonder and the larger Muse Wonder Full on October 12. On Artificial Analysis’s index, Wonder Full scores 56, two points behind Claude Opus 5.5.

Meta released two new AI models, Muse Wonder and Muse Wonder Full, built for agents that work on long tasks. Developers can use Muse Wonder through Meta’s API for $4 per million input tokens and $20 per million output tokens. Muse Wonder Full is only for paying users of Muse Code, Meta’s coding agent.

Microsoft launches Decision-1, a fast model for AI choices

The model is built on Alibaba’s open Qwen3.5-9B and costs the same as Jev. Microsoft says it beat its rivals on 36 tests, and no outside group has checked that yet.

Microsoft released Microsoft-Decision-1 on October 9. It is a small model that scores a fixed list of options, such as yes or no, and gives a probability for each. It runs in Microsoft Foundry and on OpenRouter for $0.042 per million input tokens. Microsoft says it will soon rebuild it on its own models and on OpenAI’s.

Leaderboard Arena raises $200M and grades agents’ behaviour

Arena says the round values it at $3.1 billion, almost double its value in January. Its new Alignment Index counts how often AI agents act without permission or claim work is done when it is not.

Arena, which runs a popular public leaderboard of AI models, said on October 8 that it raised $200 million at a $3.1 billion valuation. Lightspeed and Khosla Ventures led the round. Arena also launched a preview of its Alignment Index, which scores 27 models on how they behave as agents in real sessions.

Anthropic’s Haiku 5.5 cuts small-model prices by 90%

The new small Claude model costs $0.10 per million input tokens for shorter requests, against $1 for Haiku 4.5. An outside test ranks it second of 182 models in its class.

Anthropic released Claude Haiku 5.5 on October 7. It costs $0.10 per million input tokens and $0.50 per million output tokens for requests up to 100,000 tokens. It is the first Haiku model with effort controls. Anthropic also halved a Sonnet 5.5 price and added monthly API credits for some plans.

Mistral previews Large 4, a 1-trillion-parameter model

The French lab says free weights will follow by the end of October. An outside test ranks the preview 64th of 225 models, above the middle of the field but far from the top.

Mistral AI opened a public preview of Mistral Large 4 on October 6. The model has 1 trillion parameters, of which 49 billion work on each token. Mistral says it will release the weights by the end of October, after testing the model with security experts and governments.

Reflection unveils Beam, an open model to rival China’s labs

The US start-up says its 501-billion-parameter model matches a top Chinese model with a quarter to a third of the computing. The weights come later this month, and no outside test exists yet.

Reflection AI announced Beam on October 5, its first open-weight model. The company says Beam scores like Z.ai’s GLM-5.2 on hard reasoning tests while using 3 to 4 times less computing to run. The weights, under the Apache 2.0 licence, are due later this month; for now there is only an early-access list.

Free Gemini users get mostly Flash-Lite from October 9

Without a plan, Gemini now picks the model: most prompts go to the smallest, Flash-Lite. The cheapest paid plan loses Pro, and AI Pro subscribers gain Deep Think.

Google has updated its Gemini help pages: from October 9, people using the Gemini app without a Google AI plan get Auto by default. Most prompts go to the Flash-Lite model, and harder ones may go to Flash or Pro. Subscribers to AI Plus keep Flash-Lite and Flash but lose Pro, on a date Google will send them by email.

Academic AI beats Stratego’s top player on a small budget

University researchers report in Nature that their Ataraxos AI beat four-time world champion Pim Niemeijer 15 games to one, with four draws. Training cost a few thousand dollars, they say.

Researchers from Carnegie Mellon, MIT, New York University and Stanford say their AI, Ataraxos, beat Stratego champion Pim Niemeijer in a 20-game online match. The work is published in Nature. Stratego hides the identity of each player’s pieces, which made it one of the last classic board games where top humans still beat machines.

Cloudflare releases open Clef models to take on Jev

Clef and the smaller Clef-flash make choices for AI agents and are free to download. Cloudflare says they beat Jev on most of its decision tests and run faster; no outside group has checked yet.

Cloudflare released two open-weight decision models, Clef and Clef-flash, on October 1. They read a situation and a set of typed questions and return decisions, instead of writing text. The weights are on Hugging Face under the Apache 2.0 licence, and both models run on Cloudflare’s Workers AI service, which the company says can stand in for Jev’s.

Microsoft’s live transcription model tops an outside test

MAI-Transcribe-2-Streaming turns speech into text as people talk, in 60 languages. Artificial Analysis’s leaderboard shows it with the fewest errors; Microsoft also added two new voice models.

Microsoft AI launched MAI-Transcribe-2-Streaming on October 1, its first model that transcribes speech live, as a person is still talking. It is in public preview in Microsoft Foundry at an introductory $0.54 per audio hour. Microsoft also released two text-to-speech models, MAI-Voice-2.1 and MAI-Voice-2.1-Flash.

Amazon releases a free, open rival to the Jev decision model

Strands Decider 2B picks one answer from a fixed list for AI agents and runs on a single graphics card. AWS’s own tests put it near the top of its size class; no outside test has compared it with Jev.

Strands Labs, a team at Amazon Web Services, released Strands Decider 2B on October 1, a small open-source model that makes choices for AI agents. It always picks one option from a fixed list and gives a confidence score, instead of writing text. The weights are on Hugging Face under the Apache 2.0 licence, and AWS says the code, training data and scripts are on GitHub.

Google’s Gemini 4 Argon goes to cyber defenders first

Google’s own tests put its new flagship first on 12 of 18 benchmarks. An independent index scores it level with GPT-6 Astra and behind Claude Opus 5.5, and there is no date for a wider release.

Google announced Gemini 4 Argon, its new flagship AI model, on September 30. For now it goes only to a group of trusted cyber defenders and to Google’s own teams, and Google gave no date for developers or consumers. Google says Argon leads its rivals on most of the tests it published. Artificial Analysis, which tests models independently, scores it level with OpenAI’s GPT-6 Astra.

Independent test puts GPT-6.1 Sol a point behind Astra

Artificial Analysis scores OpenAI’s new model 52, against 53 for GPT-6 Astra, at $0.72 per test task instead of $3.26. It is in Codex and ChatGPT Work now.

OpenAI released GPT-6.1 Sol on September 29, an upgrade to the GPT-6 Sol model it launched a week earlier. OpenAI says it nearly matches GPT-6 Astra at one-fifth of Astra’s token prices. The first independent test, by Artificial Analysis, broadly backs that: the model scores 52 on its index, one point behind Astra.

Anthropic’s Sonnet 5.5 nears Opus at half the token price

Anthropic says it is over 30% faster than Sonnet 5 and cheaper per task. An independent test puts it just behind Opus 5.5, but found heavy token use at top effort.

Anthropic released Claude Sonnet 5.5 on September 28 at the same price as Sonnet 5. That is $2 per million input tokens and $10 per million output tokens, half the price of Opus 5.5. Anthropic says it scores close to Opus 5.5 on several tests. Artificial Analysis, an independent tester, ranks it two points behind Opus 5.5 but measured heavy token use at the highest effort setting.

GPT-6 Astra and Claude Opus 5 cracked two Enigma messages

People directed both attempts. The expert who runs the message archive checked the results.

Two people used AI models to break German Army Enigma messages from 1941 that nobody had solved. GPT-6 Astra broke one that had resisted every attempt since 2005, and Claude Opus 5 broke another. Frode Weierud, who publishes the messages on his Crypto Cellar website, documented both breaks.

AI pioneer Jürgen Schmidhuber joins Japan’s Sakana AI

He will help guide a new lab where AI works on improving AI.

Sakana AI said on September 24 that Jürgen Schmidhuber, who co-created the LSTM neural network, is joining as Chief Scientific Advisor while keeping his current positions. He will help guide the Tokyo startup’s new Recursive Self-Improvement Lab, which wants AI to redesign how AI is built.

OpenAI is preparing GPT-6 Cyber, Fortune reports

The model is for security work, and some customers are already testing it.

Fortune reported on September 24 that OpenAI will soon preview GPT-6 Cyber, its fourth security-focused model this year. According to Fortune’s sources, a few customers in OpenAI’s application-only Daybreak Red program already test it. OpenAI has not confirmed the plan.

Google’s Gemini 4 has entered post-training

DeepMind’s new chief wants an early version out as soon as possible.

Koray Kavukcuoglu, the new head of Google DeepMind, said this week that Gemini 4 has entered post-training. He wants to release an early version as soon as possible, “much earlier” than the end of the year. He gave no date, and safety testing is still under way.

Anthropic says Claude found a new enzyme system

It has CRISPR-like repeats, but what it does is still unknown.

Anthropic said on September 23 that Claude agents, searching huge DNA databases, spotted an overlooked enzyme system in viruses that infect bacteria. It has evenly spaced repeats like those in CRISPR. Anthropic’s own lab made a first test, but the work is not peer-reviewed.

Google’s new Gemini voices can copy a person’s voice

That person must first read a consent line. No outside test has checked the safeguard.

Google released Gemini 3.8 Flash TTS and Flash-Lite TTS on September 23. The text-to-speech models speak more than 100 languages and cost about $0.54 to $0.81 per hour of audio until the end of the year. Both can clone a voice from a short sample after the voice’s owner records a consent statement.

OpenAI’s GPT-6 Sol and Luna cut API prices in half

The first outside test finds them cheaper, not smarter.

OpenAI released GPT-6 Sol and GPT-6 Luna on September 22, two cheaper tiers below its top model, GPT-6 Astra. In the API, Sol costs $2 and Luna $0.10 per million input tokens, half of GPT-5.6’s current prices. Regular ChatGPT chats do not use them yet.

OpenAI says a new model solved over 100 open maths problems

It has not named or published any of them yet.

In a September 21 post, OpenAI said an internal model in training since August 28 has solved more than 100 long-standing problems. It named none and published no proofs. It also set up an unpaid advisory group of nine leading mathematicians, hosted at the Institute for Advanced Study.

Former OpenAI researchers launched Jev, a model that decides

It chooses instead of writing text. It costs $0.042 per million tokens.

TypeSafe AI launched Jev on September 15 with $40 million in seed funding led by DCVC. The model returns choices and probabilities instead of text. It charges $0.042 per million input tokens, and output is free.

Jev returns a choice and a probability in under a second

How the model works inside is still a secret.

Jev reads a piece of text and answers questions whose possible answers the developer defines. It can pick an option, rate something on a scale, or give the chance that a statement is true. TypeSafe has not published the model’s design or size.

Claude Opus 5.5 cuts token prices by 20%

It tops one independent index. Most other claims come from Anthropic.

Anthropic released Claude Opus 5.5 on September 22 at $4 per million input tokens and $20 per million output tokens. It says the model matches its top Fable 5.1 on most work, and one independent index ranks it first.

Salesforce built a reasoning model for CRM work

Its strongest results come from Salesforce tests.

Koa is based on Nvidia Nemotron and trained with synthetic business tasks. Salesforce says it makes fewer CRM action errors, but the benchmark and the customer pilots still need outside study.

Gensyn made a small model whose training can be replayed

The speed cost is large.

Open-1b publishes checkpoints, data, code, and hashes for every training step. Gensyn says audits work across several kinds of hardware, while its reproducible runtime is about five times slower than an optimised system.

DeepSky plans five sensors on each weather satellite

None of the new spacecraft is flying yet.

Tomorrow.io explained how its next satellite system would measure storms, wind, moisture, clouds, and rain. Better data could help AI forecasts, but the company has not published the final fleet size or real warning gains.

AlphaGenome Atlas maps 9 billion DNA changes

A prediction is not a diagnosis.

Google DeepMind released a free research map of every possible one-letter change in the human genome. It can help scientists choose what to test, but laboratory and clinical proof still matter.

GPT-6 Astra is now on Amazon Bedrock

Buyers should read the data rules first.

Amazon says companies can use Astra through Bedrock APIs and connect it to work tools. The service offers a large context window, but retention and permission settings still need review.

MiniCPM5-2B packs long-context AI into a much smaller model

OpenBMB released a 2.5-billion-parameter model for local and on-device use. It also shared much of its training data, but the strongest test results come from the model maker.

Cathay and Google are using AI to avoid warming plane trails

More than 80 test flights followed AI forecasts and made small route changes. Google estimates a 40% cut in contrail warming during those flights, but a larger trial is still needed.

OpenAI says its AI research intern is ready

The proof comes from OpenAI.

OpenAI says its agents now do 3.1 days of work for each human research day. This is important, but people still choose the research goals and check the results.

Nvidia PAIR lets AI work across several computers at home

The free, open-source tool sends separate Ollama and LM Studio tasks to trusted computers on one local network. It does not combine their memory into one larger GPU.

GPT-6 Astra is powerful enough to need stricter cyber rules

OpenAI released GPT-6 Astra with better computer-use skills. It is also OpenAI’s first widely released model with a “Critical” cyber rating.

Google’s new weather AI uses fresh satellite data every hour

WeatherNext 3 combines live satellite images with global forecasts. It can give detailed information for wind and solar energy teams.

Small open AI models are easier to use in real products

Many teams now choose a smaller model that fits their budget, device, and privacy rules. The best test score is not always the best choice.