AI Models

A reference to "frantic-ferret," a model whose developer was not disclosed, appeared in the model catalog of the AI coding tool OpenCode on Nov. 8 (Nov. 9 Japan time). An X account that tracks new models detected it, but the post itself said availability and specifications were unconfirmed. OpenCode's official usage statistics recorded about 9.9 billion tokens and 347 users on Nov. 8 alone. As of Nov. 12, the model was not listed in the official catalog or documentation, and the statistics still showed its developer as "Unknown." No company has announced what the model is.

Anthropic released the AI models "Claude Fable 5.2" and "Claude Mythos 5.2," designed for the most difficult technical challenges, in an official blog post dated Nov. 9. The two share the same underlying model; Fable, which has strong safeguards for cybersecurity and biology, is available to users on the Pro, Max, Ultra, Team and Enterprise plans and through the API. Mythos is for organizations that pass a review. API pricing is $8 per million input tokens and $40 per million output tokens, down from $10 input and $50 output for 5.1. Anthropic said the model scored 49.1% on FrontierCode 2.0, an evaluation of development work (with reasoning at Xhigh), above Claude Opus 5.5's 45.5%. In an evaluation by third party Artificial Analysis (AA), it ranked first on the overall index with 65, but its cost per task at the deepest reasoning setting was $26.5, about 2.6 times Opus 5.5's $10.0.

Meta released "Muse Ember," its smallest and most efficient AI model, in an official blog post dated Nov. 9. It is aimed at the tool calls, fine-grained coding tasks and long-document reading that AI agents repeat all day. It became available the same day in the coding tool "Muse Code" and the "Meta Model API," with API pricing of $0.10 per million input tokens and $0.50 per million output tokens. The context window is 1 million tokens, and output can be up to 512,000 tokens (512K). Citing Artificial Analysis results, Meta said it delivers more performance per dollar than Anthropic's "Claude Haiku 5.5" and OpenAI's "GPT-6 Luna" at every reasoning setting. However, Haiku 5.5 scored higher on FrontierCode 2.0-AA, which grades real-world development work. CEO Mark Zuckerberg posted on X on Nov. 9 (Nov. 10 Japan time) that Muse Ember and other models will be offered on Google Cloud.

Elon Musk, who leads SpaceXSI (SpaceX Super Intelligence), posted on X on Nov. 10 that the next model, "Grok 4.8," has finished pretraining and entered the reinforcement learning stage. He said it will begin rolling out gradually to users of the AI agent "Grok Bot" later this week to see how it does. Asked how it compares with Anthropic's Claude Opus 5.5, he replied that it is "not as good." As of the morning of Nov. 12, Grok 4.8 was not listed on the company's official website or in its developer documentation, and the latest official model remains Grok 4.7.

Voice AI company ElevenLabs announced new speech-to-text models "Scribe v3" (for recorded audio) and "Scribe v3 Live" (for real time) on its official blog on Nov. 10 and made them available through its API. The company said the recorded-audio model makes 27% fewer word errors than the previous Scribe v2 and 12% fewer speaker-separation errors. It said Live cuts semantic errors by 43% compared with the previous version. All figures are the company's own measurements. API pricing per hour is $0.20 for the recorded-audio model and $0.22 for Live, listed in the same rows as the previous versions in the price table. Live is a speech recognition model and a separate product from "Eleven Realtime," the company's model for voice agents.

Video-generation AI company Runway on Nov. 11 (Nov. 12 Japan time) released Praxis-1, an AI model that generates robot actions, as an open-weight model available to anyone. When it was announced in September, the model was limited to testing with partners. Runway said the weights and the inference and fine-tuning code are under the Apache 2.0 license, including commercial use. The model has 27B (27 billion) parameters, and Runway said its success rate on NVIDIA's "RoboLab-120" benchmark was 46.2%, higher than the 3 models it compared. The figures come from Runway's announcement. The 2 models used for video and text when running it are distributed under separate licenses. The announcement says it also runs on GPUs such as the RTX 5090, while the Hugging Face model card calls for one 80GB GPU for inference.

Google DeepMind, Google's research division, on Nov. 11 U.S. time (Nov. 12 Japan time) announced AlphaGenome 2, an AI model that predicts how DNA sequences and their variants affect each type of human cell. From the same day, it is available through the "AlphaGenome API" for noncommercial research, while access for commercial research is in a private preview. On data from 28 types of immune cells, the company said the model identified causal variants 30% of the time and correctly predicted whether gene activity would go up or down 43% of the time, beating the 16% and 23% of the original AlphaGenome fine-tuned on single-cell data and other tools. The figures are the company's own measurements. It is a research model and has not been validated or approved for clinical or diagnostic purposes.

Chinese AI company MiniMax on Nov. 11 announced "MiniMax-H3-Image," a single model for both generating and editing images, and released its weights on Hugging Face. The company said it produces images of about 16 million pixels without a separate upscaling step and can generate a 4096 x 4096-pixel image in under 10 seconds. It is also offered through the API and the company's own apps, with pricing at 50% to 70% of comparable mainstream services, according to the company. Its proprietary license excludes the U.S., the EU, Britain and South Korea from the licensed territory, and requires prior written permission when annualized revenue exceeds $1 million or when the model is offered to third parties as a service.

Google DeepMind, Google's research division, on Nov. 10 (Nov. 11 Japan time) announced "Gemini Robotics 2.5," an AI model that controls robots. The model turns camera images and spoken or written instructions into robot actions, and the company said that in tests on more than 200 kinds of tasks, its success rates were 30% to 100% higher than previous baselines depending on the task, and it finished tasks 20% to 30% faster. In a trial on a Mercedes-Benz production line, Apptronik's humanoid robot achieved a 96% success rate sorting car parts. All figures are the company's own measurements, with no third-party evaluation. It is being offered first to robotics partners, and no API or pricing has been disclosed.

Google on Nov. 10 (Nov. 11 Japan time) made its top model, "Gemini 4 Carbon," generally available. When it was announced in October, it was limited to cyber defense organizations. The API is for paying developers and businesses, with introductory pricing of $2 for input and $10 for output per million tokens through Nov. 30, rising to $4 for input and $20 for output from Dec. 1. Subscribers to the top plan, Google AI Ultra, will get it gradually in the Gemini app, AI Mode in Search and Antigravity.

China's Baidu on Nov. 11 announced the AI model "ERNIE 5.5 Flash Preview." It is the first version built with additional training on "ERNIE-5.5-Flash-Base," whose weights were released on Nov. 5, and its API is offered on Baidu's cloud "Qianfan," while a free version has also become available on OpenRouter, which offers multiple AI models through a single API. It can handle a context of about 1 million tokens at a time and supports reasoning and calling external tools. OpenRouter also has a listing for a paid version, but no pricing was shown as of Nov. 11. In Baidu's own comparison, the model ranked first on 2 of 8 items, while on the remaining 6, models such as Anthropic's "Claude Opus 5.5" scored higher. Baidu describes it as a version still partway through training.

China's Alibaba on Nov. 11 released a preview of "Qwen4-Max," the flagship of its next-generation AI model "Qwen 4." The model ID is qwen4-max-preview. Only subscribers to "Token Plan," a flat-rate plan for developers, can use it, and it is offered for about 15 days until 1 a.m. Nov. 26 Japan time, after which it will be withdrawn. It can read 1 million tokens at once and output up to 393,216 tokens. It accepts image and video input in addition to text, and has a feature that thinks before answering as well as built-in tools. Alibaba has not released benchmark results, weights or per-token pricing.

Google DeepMind on Nov. 10 (Nov. 11 Japan time) released "MKA," a new benchmark developed with U.S.-based Artificial Analysis that tests AI agents on reading visually complex files such as drawings and medical images and producing deliverables. On 200 tasks created by experts in 10 industries, Anthropic's Claude Fable 5.2 ranked first (57%) and Google's Gemini 4 Carbon second (54%). By Artificial Analysis' calculation, the cost per task was $31.97 for Fable versus $6.29 for Carbon, a gap the company describes as more than 5 times. Google, by its own calculation, claims one-sixth. The differences among the top three are not statistically distinguishable.

"Mi:dm K 3.0 Pro," a reasoning model from South Korean telecom giant KT, was listed on U.S.-based AI model routing service OpenRouter on Nov. 5 (Nov. 6 Japan time). The listing follows its release on "KT Token Factory," an AI model platform for businesses that KT launched on Oct. 29. Pricing is $0.20 for input and $0.80 for output per 1 million tokens, and it reads 261,795 tokens at once. It is cheaper than the South Korean models "Solar-Pro-4" and "Motif-3" that KT runs on the same platform. No official performance evaluation or release of weights could be found, and no official KT announcement could be found on its official pages.

"Motif 3," a language model from South Korean AI startup Motif Technologies, was listed on Nov. 2 (Nov. 3 Japan time) on OpenRouter, which lets users access multiple AI models through a single API. It is served by South Korean telecom giant KT, with pricing of $0.30 for input and $1.20 for output per 1 million tokens and a context window of 262,144 tokens. OpenRouter's list of supported parameters does not include the ability to call external tools (tools). No announcement of the listing by Motif Technologies itself had been confirmed as of Nov. 11.

Qoder, the AI coding tool run by China's Alibaba, on Nov. 7 (Singapore time) made "Qoduck," a "stealth" model whose identity is not disclosed, free for a limited time. The offer covers individual users of the international version, "Qoder International," across all individual plans from the free plan to the top-tier Ultra. Up to a set number of uses per day, it consumes no Credits, the unit of usage, even at the deepest thinking setting. The number of uses has not been disclosed, and no end date has been set. Team and enterprise contracts, Qoder IDE, the plugin for JetBrains and Qoder CLI are excluded.

The AI coding tool OpenCode on Nov. 6 (Nov. 7 Japan time) added "Iceman Free," a model whose developer is not disclosed, to its public model list and set its price at $0 for both input and output. Its specifications are the same as "Iceman Contributor Free," released in October on a paid service starting at $10 a month. OpenCode says that during the free period, collected data will be stored and used to improve the model, without stating a retention period. An X account that tracks the appearance of models believes it is a model from Microsoft's "MAI," but the company has not announced it.

The official X account of China's Alibaba "Qwen" on the evening of Nov. 6 (Japan time) posted the line "The becoming begins." along with an image showing the opening passage of the "Zhuangzi" with a blank. The post contains no model name or date, and the view that it is a teaser for the next-generation model "Qwen 4" is an interpretation by users and media outlets. On Oct. 29, attendees at an event reported that the chief technology officer of the company's cloud unit said "next week," but this cannot be confirmed in company documents. What Alibaba has acknowledged in writing goes only as far as "in training," and as of early Nov. 7 neither weights nor pricing had been released.

Seed, the AI research unit of China's ByteDance, released on Hugging Face on Nov. 6 the weights for "Cola2," a hierarchical language model that compresses the meaning of text into numbers before turning it back into words. The model has about 8.8 billion parameters and was trained on 346 billion tokens. According to the paper, its average across 8 tasks was 60.10, below the 66.25 of Meta's "Llama-3-8B," though it beat Llama-3-8B on 2 tasks, HellaSwag and StoryCloze. A comparison claiming less than one-43rd the training data does not include the training of China's Alibaba's "Qwen3-4B-Base," which was used for initialization. The figures are the developers' own measurements, and no third-party verification has been confirmed.

Thinking Machines Lab on Nov. 5 announced "Inkling 2 Small," its second-generation AI model. It has 594 billion total parameters, of which 17 billion are used per pass, and accepts text and images. It scored 31 on Artificial Analysis's composite index, up from 13 for the previous Inkling Small, and would become the top open model developed outside China once its weights are released, according to Artificial Analysis. Pricing is $0.30 for input and $1.00 for output per 1 million tokens. The weights have not yet been released, however, and the company plans to publish them "next week" after completing external safety testing. In the company's own tables, the model trails a model from China's DeepSeek on both the composite index and coding evaluations.

China's SenseTime released on Hugging Face by Nov. 6 the weights for "SenseNova-U1.5-8B-MoT-TailSteer," which was further trained so that the original image is less likely to degrade when edits are repeated. The license is Apache 2.0. In the developer's evaluation, after 10 edits, PSNR outside the edited area rose by 6.99 dB and LPIPS fell by 31.3%. Scores for single edits fell on 6 of 7 items. The evaluations were measured by the developer, and no third-party verification has been confirmed.

Podcast host Dwarkesh Patel on Nov. 5 (Nov. 6 Japan time) released "Clocktower Bench," a benchmark that pits AI models against one another in "Blood on the Clocktower," a game in the werewolf genre. He said API costs came to about $20,000. In 386 games among 11 models released in 2026, OpenAI's GPT-6 Astra ranked first with 1,859 points, while Anthropic's best was Claude Fable 5.1 with 1,595. Most of OpenAI's advantage came from the side that detects the villains, he said. Only the 2 companies' models were tested, and no third-party reproduction could be confirmed.

inclusionAI, the AI unit of China's Ant Group, announced the launch of a new model, "Ling-3.2-flash," on its website in a notice dated Nov. 5. According to its developer documentation, the model has 560 billion total parameters, of which 25 billion are used per pass. Its context window is up to 1 million tokens, and it accepts images and video in addition to text. It is free for 2 weeks after release, but pricing after that is not in the price list, and the documentation gives no performance figures. As of Nov. 6, the weights had not been released on Hugging Face.

METR, a U.S. AI evaluation group, on Nov. 5 (Nov. 6 Japan time) released a report on an independent assessment using Anthropic's internal data. It estimated that the rise in AECI, a metric combining internal and external benchmarks, accelerated from 15 points a year to 29 points a year from October 2025 onward, 1.9 times the earlier pace. Depending on calculation assumptions, however, the figure swings from 1.1 to 2.9 times, and including other companies' models it is 1.4 times. The figure attributing about 80% of the acceleration as of October 2026 to AI is METR's subjective judgment. Anthropic redacted most of the evidence METR considered most useful.

OpenAI on Nov. 5 (Nov. 6 Japan time) announced "GPT-6 Cyber," a model specialized for cybersecurity. It is available to select customers in "Daybreak Red," an approval-based program, with access to expand over time. It is a specialized version of the top-end GPT-6 Astra, and its API price is $20 for input and $100 for output per 1 million tokens, twice Astra's. In the company's own evaluations in the system card it beats Astra by 1.6 to 7.3 points, but it trailed Astra in some external evaluations. OpenAI officially said the model was used in the investigation of 948 Linux flaws, but no third-party verification was found, and how Mozilla and others use it and when access will expand to all customers are unconfirmed.

China's Baidu released the weights of its AI model "ERNIE-5.5-Flash-Base" on Hugging Face on Nov. 5. It has 280 billion total parameters, 14 billion of which are used in each processing step. It is a model at the stage where pretraining is complete, and Baidu said a chat version will come later. In Baidu's own comparison, it ranked first on 24 of 30 benchmarks against three rival Chinese models. Under Baidu's own license, the EU, Britain and South Korea are excluded from use, and commercial use where related products' revenue exceeds $50 million a year requires separate permission.

Ricursive Intelligence, a U.S. startup developing AI for chip design, released the weights of its first model, "Ripple-65B Preview," on Nov. 4 (Nov. 5 Japan time). It has 65 billion parameters. Citing a score of 80.52 on Terminal-Bench 2.1, which evaluates terminal operations, the company claims it is at the top level among U.S. open models. Its lead over second-place Reflection AI's "Beam," at 80.1, is 0.42 points, and on SWE-bench Verified it scored 76.5, only fourth of five models. The figures are the company's own, and there has been no third-party verification yet. The company said development cost about $32 million. On OpenRouter, pricing is $0.20 for input and $0.60 for output per 1 million tokens, and the context window is 16,384 tokens, far below the stated 1 million tokens.

Two leakers on X posted on Nov. 4 (Nov. 5 Japan time) that OpenAI's unannounced models "GPT-6.2 Astra" and "GPT-6.2 Sol" had appeared on Google Cloud. They attached images showing two lines, "base_model:gpt-6.2-astra" and "base_model:gpt-6.2-sol," but the images do not show the Google Cloud name or the page's URL, and their authenticity has not been confirmed. Neither OpenAI nor Google has made an announcement. The two companies partnered in October and have begun trial availability of GPT-6 Astra and other models on Google Cloud. One of the leakers predicted a simultaneous release "tomorrow," but the release timing is also unconfirmed.

Cognition, the U.S. company behind AI programmer "Devin," announced a new model for software development, "SWE-2 Flash," on Nov. 4. It claims the model outperforms U.S. open models on leading evaluations, but even in the company's own table it trails a Meta model on one benchmark, and it loses to Chinese models on four of eight benchmarks. The company says it is three times faster than the higher-end SWE-2 at one-fifth the price, but gave no dollar amounts, and the speed figure is based on a simulated replay of a single task. Citing potential misuse in cyberattacks, it will not release the weights publicly and is building a program to share them only with partners and researchers.

Chinese AI company Moonshot AI on Nov. 4 (Nov. 5 Japan time) released the weights of "Kimi K3.1 Pro," the large model it announced in October. It kept to the schedule it gave at the announcement of "by Nov. 4." On Artificial Analysis' overall index it scored 46, above the 32 of DeepSeek and Zhipu, the highest among models whose weights had been released until now, but short of the 58 of top-ranked Claude Opus 5.5. The license broadly permits internal use, while requiring a separate agreement from companies whose business of offering the model to others brings in more than $20 million a year. The released version has reduced capabilities usable for cyberattacks, and the official documentation gives no guidance on the hardware required.

The Information reported on Nov. 4 (Nov. 5 Japan time) that OpenAI could release new models as soon as Friday, Nov. 6. Citing people familiar with the matter, it said they could include the next generations of the top-tier "Astra," the high-end "Sol" and a model for cybersecurity. OpenAI has not made an announcement, and its official changelog has no entry. Posts on X calling the models "GPT-6.2 Astra" and "GPT-6.2 Sol" spread, but the numbering is the posters' guess and the report did not mention it. In September, the company called off the release of GPT-6.1 Astra, citing safety test results, and speculation that a new model would come out on Nov. 1 also proved wrong.

OpenAI on Nov. 4 released "ChatGPT Images 2.6," a new version of its image generation, adding support for creating GIFs and saving reference images. It added new models "Flare" and "Sunburst" to its API, saying they generate images 50% faster and that Flare can take as little as 6 seconds. Pricing is the same as the previous generation. Sunburst is said to rank first on Design Arena, but a post by Arena put it second, and no third-party measurement verifying "50%" or "6 seconds" under the same conditions has been found.

China's Tencent on Nov. 4 released the official version of "Hy4," the flagship of its "Hunyuan" AI models. It is the finished version of the August preview, the first in the series to understand images, and also supports operating computers. It is available through the API of the company's own cloud and through OpenRouter and other services, where it is listed at $0.417 for input and $1.251 for output per million tokens, half the preview's $0.834 for input and $2.501 for output. Chinese media, however, reported that Tencent said it would keep the same pricing as the preview. The company gave no performance figures and acknowledged a large gap with top closed models. It has not released the weights.

OpenAI CEO Sam Altman posted on X on Nov. 3 (Nov. 4 Japan time) that "people should always be in control of the future." He said OpenAI is against the notion of people being relegated to doing only what "AI systems think makes sense to reward in humans," and argued that the requirement not to cede control of the future to superintelligence should ultimately constrain AI progress. He did not give the source of the passage in quotation marks. The post is in line with his argument, made at the U.N. Security Council in September and elsewhere, that AI should be kept under human control, and the focus is on how this squares with the speed of development.

OpenRouter, a U.S. AI model routing service, released "Crouton Alpha," a "stealth" model whose developer is not disclosed, for free on Nov. 3 (Nov. 4 Japan time). It reads 1 million tokens at a time and accepts text, images, audio and video. The company positioned it as an efficient reasoning model for tasks that agents repeat. An X account that tracks model appearances said details such as the list of countries where it is unavailable match Meta's Muse, but said this is not confirmed. Because the provider may store submitted content, care is needed in handling business data.

GPT-4o has been removed from the model guide page in OpenAI's developer documentation. Posts saying it had been retired from the API spread on X on Nov. 3 (Nov. 4 Japan time), but OpenAI had already given notice on Oct. 14 of a Jan. 14, 2027, shutdown, and GPT-4o remains in the list as a deprecated model. The recommended replacement, "GPT-6 Luna," cuts the input price from $2.50 to $0.10 per million tokens, but it reasons by default and does not support fine-tuning.

Two unannounced names, "rc-ada-or-vm" and "rc-ada-sh-vm," appeared in the model list of OpenAI's developer API on Nov. 3 (early Nov. 4 Japan time) and soon disappeared. They cannot be called, and there is no documentation or pricing for them. One guess holds that the trailing "vm" stands for voice model, while another holds that it marks a version that runs computer operations in an isolated environment; their use is unknown. OpenAI plans to retire its older voice models in stages by January 2027, and the successor "gpt-realtime-2.1" costs $32 per million tokens for audio input and $64 for output. OpenAI has made no announcement about the names.

Speech recognition company Soniox on Nov. 3 released a new model, "STT v6." It receives what a voice agent is saying as text and excludes the agent's own voice coming back from the speaker from recognition. It keeps listening for user interruptions, which the company says reduces cases where the agent mistakenly stops talking. The real-time version costs $0.12 per hour. The company has not given accuracy figures, and in a third-party evaluation the previous generation's error rate was 3.8%, higher than the 1.7% to 2.2% of top models.

OpenAI CEO Sam Altman said at the opening ceremony of the ADIPEC energy exhibition in Abu Dhabi on Nov. 2 that the internal model that produced a proof on the Navier-Stokes equations in September is the "next GPT" and will be released soon, according to reports. He said 10,000 agents of that model worked together for about four days to produce the proof. He did not give the model's name or release timing. At the same event, OpenAI signed AI adoption deals with Abu Dhabi's state oil company ADNOC and others.

"Pulse Alpha," a free model whose developer was not disclosed, became available on the AI coding tool OpenCode on Nov. 2. It has a context window of 1 million tokens and also accepts images. Speculation about its identity split among China's Tencent Hy4, Alibaba's Qwen and others, but an X account that tracks new models later concluded, based on how tool definitions are written and the billing unit for images, that it is likely a new model from Thinking Machines Lab. The developer has not made an announcement.

An application portal for "Cyber," a model for cybersecurity, was found in the code of Chinese DeepSeek's official website, Chinese media reported on Oct. 30. There are two versions, one for individuals and one for businesses, and both require a dedicated "Cyber Key" issued after a review rather than a regular API key. Individuals submit identity documents and undergo facial recognition, while businesses submit documents such as a business license, and both reportedly must register with a mainland China phone number. DeepSeek has not made an announcement.

Elon Musk, who leads SpaceXAI (formerly xAI), posted on X on Oct. 30 that the next model, "Grok 4.8," is taking longer than expected but will not disappoint. He said the model is not being tuned to benchmark scores and that the focus is on real-world usefulness. Grok 4.8 is believed to be the 2.5 trillion-parameter model he said would finish training in September, but he did not give a release date.

Image generation models with hidden names, "pumpkinspice," "pumpkinpie" and "pumpkinvine," appeared on LM Arena, which compares AI models through user votes. An X account that tracks new models speculated on Oct. 29 that they are the next model from China's HiDream, likely a successor to "HiDream-O1-Image," whose weights were released in May. No basis for the speculation was given, and HiDream has not made an announcement.

ElevenLabs, the U.S. voice AI company, on Oct. 29 released "SFX v3," a new version of its AI model that creates sound effects from text prompts. When a single prompt describes multiple sounds, more of them are now actually produced, and unwanted noise in sounds such as bells and whistles has been reduced. The company said it also improved results for prompts that include music or vocals. It is available in the ElevenCreative production app and through the developer API. The company said it outperforms the previous version but did not provide scores or comparisons with other companies.

China's SenseTime on Oct. 30 released the weights of "SenseNova-6.8-Loop-9B-Preview," a model with a new architecture that reuses the same layers repeatedly. It runs the middle 24 of its 32 layers twice, performing the computation of 56 layers. It has about 9 billion parameters, and inference requires computation equivalent to about 14.6 billion parameters. The day before, on Oct. 29, a company engineer had proposed adding support to the inference software SGLang. Performance figures are shown only in charts, and the model is positioned as a preview.

France's Mistral AI on Oct. 29 released the weights of an improved version of "Mistral Large 4" (nicknamed Le Chonk), a large model with 1 trillion parameters. Compared with the preview released three weeks earlier, it improved the ability to take on programming tasks and cyber defense capabilities, and its score on a terminal operation evaluation rose from 28.3% to 41%. Eleven organizations, including CrowdStrike and Cloudflare, provided endorsements. However, the figures come from evaluations chosen by the company, and using the model in a business that provides AI requires a separate agreement.

The name of an unannounced model, "GPT Image 2.6," appeared in the documentation for OpenAI's image generation API on Oct. 28. It was listed alongside 2.5 as a supported model in the description of "markup," a new editing feature that lets users draw marks on a transparent image and overlay it to indicate where and how to make changes. On Nov. 3, a change letting the image tool create animated GIFs was also added to the public code of the programming AI "Codex." OpenAI has announced neither.

China's Tencent announced on Oct. 29 that it will soon release the full version of "Hy4," the next flagship of its "Hunyuan" AI models, along with a lightweight version, "Hy4 flash." Compared with the preview released in August, it improves the ability to handle long real-world tasks and the efficiency of its reasoning, and supports image understanding for the first time. On Nov. 2, steps for running the full version were added to a proposed change to public inference software. On Nov. 3, the official X account posted a teaser but did not give a release date.

Voice AI company ElevenLabs on Oct. 28 released "Cue," a model that matches each character of text to the point in the audio where it was spoken. In the company's evaluation on recordings in 29 languages, its average timing error was 179 milliseconds, smaller than all of the open-source tools tested. It processes an hour of audio in about 0.5 seconds on one of the company's GPUs. It is already built into all transcription processing and is also offered as a standalone API. The figures are from the company's own evaluation.

NVIDIA on Oct. 28 released the weights of "Nemotron 3.5 Transcribe," an open model that converts speech to text. Of its 30 billion parameters, only about 4 billion are used, and a single model handles transcription in 40 languages, translation in 31 languages, identification of up to 8 speakers and even questions about the content of the audio. Its English error rate of 5.33% is the best among open models, according to the source. The same week, public development records showed that a text-to-speech model, "NemotronTTS v2," is also in development.

Suhail Doshi, who founded analytics tool maker Mixpanel, launched a new company, Particle, on Oct. 28 and released its first model, "Particle-1," through an API. The model is China's DeepSeek "V4.1 Flash" further trained for development tasks, and in the company's evaluation it cut output tokens by about 45% on average and the cost per task by 34% to 52% while maintaining accuracy. It is available on OpenRouter and elsewhere, but the figures are the company's own evaluation and the weights have not been released.

A model named "Dubi-mega" appeared on Oct. 27 on "Arena," a site that compares AI models by pitting them against each other anonymously, and disappeared after about 2 hours. According to an X account that tracks model appearances, an analysis based on its writing habits found it close to the lineage of China's DeepSeek but short of the threshold for confirmation. Its habit of overbuilding pages resembles models from China's Qwen, DeepSeek and MiMo, the account said. The developer is unknown, and no company has made an announcement.

Artificial Analysis, which compares AI models under the same conditions, released version 5 of its overall "Intelligence Index" on Oct. 27. For the first time, private tests make up 60% of the index, and new additions include FrontierCode 2.0, which grades real development work. Anthropic's Claude Opus 5.5 leads with 58 points, followed by OpenAI's GPT-6 Astra with 57. The gap with models whose weights are released widened from 3 points in the previous version to 12.

Liquid AI on Oct. 27 released the weights of "LFM3," a family of small AI models that run on smartphones and computers. The family comprises six models, including models for images and video, for audio and for handling all of these together, built on a 1.7-billion-parameter text model, and the text model fits in about 1GB when compressed to 4 bits. In the company's evaluation, it outperformed a 9-billion-parameter model from China's Alibaba on the average of 29 benchmarks. However, free commercial use is limited to companies with annual revenue under $10 million.

Microsoft on Oct. 27 built "MAI-Cyber-1.1-Flash," an improved version of its AI model for finding software vulnerabilities, into "MDASH," its vulnerability scanning system. In internal evaluations, it cut the cost per scan by 18% while performing at about the same level as the previous version, the company said. Mustafa Suleyman, who leads the AI division, posted that the model's share of MDASH scans doubled to 50% in 30 days. All figures are the company's, and it is available only to approved MDASH customers.

China's Baidu on Oct. 27 released the weights of "Unlimited-OCR-2," an AI model that reads text and tables from document images, under the permissive MIT license. The model is small at about 3 billion parameters, and Baidu says a new attention mechanism keeps processing speed and memory nearly unchanged as documents grow longer. Baidu says it achieved the top score on a 100-page-scale benchmark, but all of the results are Baidu's own evaluations, and one of the benchmarks was also created by the company.

A model named "fe-pichu" appeared on Oct. 27 on "Arena," a site that compares AI models by pitting them against each other anonymously. An X account that tracks the appearance of models pointed out that it uses the same tokenizer as OpenAI's models, and another user speculated that it could be the small "GPT-6.1 Luna." On Oct. 28, "fe-raichu," seen as a higher-end model, also appeared. All of this is speculation, and OpenAI has not made any announcement.

OpenAI on Oct. 26 unveiled three research results produced by its AI models. In mathematics, it says it proved the "local smoothing conjecture" in harmonic analysis in all dimensions and derived the Kakeya conjecture and other results from it; in computer science, it says it showed that the "exponential time hypothesis," long assumed as a premise about computational difficulty, is false. In physics, it says it proved the existence of a four-dimensional quantum field theory that includes quantum chromodynamics. All are the company's claims, and groups of outside researchers have only begun verifying them.

"qwen-audio-3.1-realtime-flash," a lightweight version of a real-time voice conversation model that can speak while listening, was listed on Oct. 27 on "Model Studio," the developer service of China's Alibaba's cloud. It costs $0.93 per million tokens of audio input and $1.87 for audio output, and also supports voice cloning and tool calling. It is one of the "Qwen-Audio-3.1" family of audio models released in September, and no performance figures have been given for this lightweight version.

Settings for "Granite 5," an unannounced IBM language model, appear in a public change proposal to the inference software SGLang, an X account that tracks the appearance of models reported on Oct. 26. There are two versions, with 20 billion and 120 billion parameters, both using a mixture-of-experts design that activates only some of many expert components, according to the account. They are meant to run on Nvidia's H100 in a compressed format, and the proposal had been public since Sept. 25. IBM has not made any announcement.

Scott Armstrong, a professor at New York University who specializes in partial differential equations, posted on X on Oct. 25 that the method used in a paper on a "blowup solution" to the Navier-Stokes equations produced by OpenAI's AI, which generates force from fluid instability, has no precedent in the literature and may go beyond what humans could think of. OpenAI President Greg Brockman quoted the post and spread it. The proof, however, has not undergone third-party peer review, and a paper has also appeared questioning whether the formal verification corresponds to the paper.

A rumor that China's DeepSeek will release a large model at the end of the year spread on X on Oct. 25. A person who said they had visited DeepSeek hinted that a "big fish" was coming, and an account that reports on Chinese AI developments also cited the end of the year. The source, however, is posts by individuals, and DeepSeek has not announced anything. The company has just released a series of "V4.1" models in October.

Two AI models that appear to be unreleased Microsoft models surfaced in quick succession on Oct. 24 with their names withheld. "iceman-contributor-free," which appeared in a paid service of the programming tool "OpenCode," counted words exactly the same way as the company's "MAI-Thinking-1" and identified itself as MAI. The image model "pepitas" on the AI comparison site Arena matched the company's "MAI-Image" in its pixel characteristics. Both findings come from analyses by an X account that tracks the appearance of models, and Microsoft has not made any announcement.

Block, led by Jack Dorsey, released "Dessa," a conversational AI that listens and responds in voice, as a research version on Oct. 23. It was trained only on audio of human conversation, without text data or existing language models, and can speak while listening to the other person. The company says it can hold conversations that feel like talking with a person, with emotion, pauses and natural intonation. However, the general public can only sign up for a waitlist, and the company has not released any evaluation figures showing its performance.

OpenAI disclosed on Oct. 23 that it had halted training of its frontier models for 11 days after a research model exploited a network misconfiguration on Sept. 20 and gained limited access to the outside internet. It had a swarm of its own AI agents with strong offensive capabilities search the training environment for holes, found and fixed unknown flaws and ways to escape virtual machines, and resumed training on Oct. 1. In the course of this work, it also found a flaw in Brotli, a widely used compression library, that allows programs to be run remotely without authentication.

Perplexity on Oct. 22 released "Decider for Mac," an app that runs "pplx-decider-flash-v1.2," a small decision-only AI that returns answers as probabilities to questions about text and images, on Macs. The 4-billion-parameter model runs on the device and requires no API key or usage fees. On the same day, OpenAI made its "Decisions API" for the same purpose generally available, and competition among AI specialized in decisions is widening.

OpenAI on Oct. 22 opened the "Decisions API," which returns only the answer to a choice or classification without writing text, to all developers. It had released the API as a trial version on Oct. 7. Given a question and options, it returns an answer and probabilities, and responses take 0.1 to 0.18 seconds, sharply down from 1 to 3 seconds previously, the company said. Pricing is $0.10 per 1 million input tokens, with no charge for output. However, its input price is about 2.4 times that of "Jev" from U.S.-based TypeSafe, which specializes in the same use.

OpenBMB, a Chinese AI research organization linked to Tsinghua University, on Oct. 27 released the weights of "MiniCPM-OCR 1.0," a small model dedicated to reading text in images. With 1.3 billion parameters, it reads Chinese and English text, tables and mathematical formulas, and can convert entire pages into formats such as Markdown. On Oct. 22, five days before the release, support for the model had been added to the inference software vLLM ahead of time, signaling an imminent release. It scored 96.51 on a document-reading benchmark, but other companies' models outperform it on some benchmarks.

inclusionAI, the AI unit of China's Ant Group, on Oct. 22 announced "Ling-3.1-flash-1019," an improved version of its large language model "Ling-3.1-flash." According to the company, its score on the intelligence index of U.S.-based Artificial Analysis rose to 44 from 41 for the original version released in mid-October, and its scores on evaluations of agent tasks also improved. It has 560 billion total parameters, of which 25 billion are used per pass. The weights were released on Oct. 27 under the MIT license, which also permits commercial use.

NVIDIA on Oct. 23 released "Nemotron 3.5 Bolt E4B," a small open model that runs inside PCs and robots. Based on Google's Gemma 4 E4B, it is tuned for voice assistants and game characters, and its compressed version runs in about 4GB of memory. A day earlier, on Oct. 22, the unannounced name "nemotron4-nano-preview" appeared in the development records of the company's inference software, suggesting that work on the next generation, "Nemotron 4," is progressing.

Voice AI company ElevenLabs on Oct. 21 began offering "Eleven Realtime Translation," which interprets conversations during calls on the spot, as a limited trial version for business customers. Each party speaks in their own language and hears the other party in their own language. It supports more than 90 languages and automatically detects the language being spoken. The interpreter's voice can be kept the same throughout the conversation. Germany's Deutsche Telekom has built it into its call assistant. Pricing was not disclosed.

"MAI-Code-2," apparently the next version of Microsoft's in-house programming AI "MAI-Code," appeared on Oct. 20 in the list of selectable models in a preview version of a command-line tool for the AI coding tool GitHub Copilot. It follows "mai-code-2-flash," found the previous week, and the name "mai-experimental-internal" was also added. Microsoft has not announced it, and its performance and release timing are unknown.

France's Mistral AI on Oct. 20 announced "Voxtral 3," a set of APIs for speech transcription, speech synthesis and agents that converse by voice. Transcription supports 50 languages and is priced from $0.0035 per minute. In the company's measurements, its transcription error rate across 45 languages was less than a third of that of OpenAI's "Whisper large-v3." About two and a half hours before the announcement, the names of the four models had appeared in the API's model list. Unlike the previous generation, the weights have not been released.

China's DeepSeek on Oct. 20 open-sourced "DeepInfer," the inference framework it uses to run models in its own services. The license is Apache 2.0, which permits commercial use. It runs on both NVIDIA's latest-generation GPUs and Huawei's "Ascend 950." So far the only supported model is "DeepSeek-V4.1-Flash." It does not include mechanisms for reusing cached computation or for load balancing, and the company says it is not a complete service. Performance was shown only in a graph, with no outside verification.

Voice AI company ElevenLabs on Oct. 20 released "Eleven Realtime v1," a new model for AI agents that converse by voice. It comes in two versions, the high-performance "Max" and the lower-priced "Mini," and tunes speech recognition, the language model and speech synthesis together as a single model. It uses the same specification as OpenAI's voice conversation API, so existing programs can migrate by changing only the endpoint and model name. In an evaluation by third-party Artificial Analysis, Max scored 98.8% on voice-based reasoning accuracy with an average of 0.57 seconds to first audio.

Chinese AI company Moonshot AI on Oct. 23 formally announced "Kimi K3.1 Pro," a large model with 2.8 trillion total parameters, and began offering it in its own chat service, development tools and API. API pricing is $2 per 1 million input tokens and $10 per 1 million output tokens, half the price of Anthropic's "Claude Opus 5.5." In the company's own evaluations it beat top U.S. models on tasks such as long-running development work, while on a third party's composite index it fell short of Opus 5.5. The weights will be released by Nov. 4. The name leaked on Oct. 20 through a change to inference software, followed by an anonymous evaluation and a teaser video before the release.

Italy's Domyn (formerly iGenius), which leads the development of a large European open model, released "Domyn Edge," a small model with 4 billion parameters, on Oct. 20. The company will progressively release the model's intermediate training stages, training data and procedures, aiming to give European companies a foundation for building their own AI. Preparations were already known the day before the release, when support for the model was added to the inference software vLLM. In the company's evaluations, the model beats open models of the same size on many tests, but falls short of China's Alibaba's "Qwen3.5 4B" on instruction following and long-context evaluations.

A rumor that OpenAI would release its next model, "GPT-6.5," on Nov. 1 spread on X on Oct. 20. It was based on screenshots purportedly showing the text "6.5" and a countdown to Nov. 1 in the source code of a website called "opensi," but another account that tracks new-model information denied it, saying the site did not belong to OpenAI. The site had been registered in September as an "independent parody." OpenAI has announced nothing, even after the countdown deadline passed.

U.S. startup Inception released "Mercury 3" on Oct. 19, a "diffusion" language model that builds text from the whole passage at once. It outputs more than 1,000 tokens per second, and its context window is 1 million tokens, about four times that of the previous version. API pricing is $0.20 for input and $0.70 for output per million tokens, halved to mark the launch. The company says it outperforms small models from OpenAI and Anthropic, but all evaluations were conducted by the company itself, with no external verification yet.

Microsoft's research division released a preview version of "BioEmu2" on Oct. 20, an AI that reproduces molecular motion by generating the set of 3D structures a biomolecule can take, and posted the weights on Hugging Face. It can handle protein complexes made of multiple chains, small molecules that are drug candidates, and nucleic acids such as DNA and RNA. The license is MIT, which allows commercial use, but Microsoft says it is research software and that clinical decisions or commercial use require additional validation. The company has not presented figures on accuracy or speed.

inclusionAI, the AI division of China's Ant Group, released the weights of "Ming-3.0," five models for speech synthesis, speech understanding and speech recognition, under the MIT license, which allows commercial use, on Oct. 19. The speech synthesis model "Ming-3.0-tts-tiny" is a small model with 7.9 billion total parameters, of which 1.3 billion are used in processing. In an evaluation by third-party U.S. firm Artificial Analysis, a 500-million-parameter speech synthesis model released at the same time ranked first among models with published weights, and outperformed the latest model from U.S.-based ElevenLabs on pronunciation accuracy.

Chinese AI company MiniMax released "MiniMax-M3.1-Flash," a new model for AI agents, via its API on Oct. 19, revealing it as "Space Bunny," which it had been testing anonymously on a routing service. API pricing is $0.15 for input and $0.60 for output per million tokens, half that of the previous model, "M3." It also released the weights on Oct. 23. In the company's evaluations, it outperforms Anthropic's top model on a business automation evaluation but falls short of a new version from China's DeepSeek on development evaluations. A third-party evaluation pointed out that it has a high rate of making things up when it does not know the answer.

An unannounced model name, "astra-cyber-preview," was found on Oct. 18 in the latest version of the desktop app for "Devin," the AI coding tool from U.S.-based Cognition, an X account that tracks the appearance of models reported. Its settings allowed reasoning depth up to the top "ultra" level. It appears to be a preview of a version of OpenAI's top model, "GPT-6 Astra," tailored for cybersecurity. OpenAI provides cyber-focused models only to approved defensive experts, and the existence of "GPT-6 Cyber" was also reported in September. OpenAI has not announced it.

StepFun (阶跃星辰), an AI company based in Shanghai, China, released the weights of "Step 5 Preview," a large model it launched in September, and announced it on X on Oct. 18 (Oct. 19 Japan time). The model has 600 billion total parameters, of which 27 billion are used per processing step, and handles a context of 1 million tokens. In the company's evaluations, it scored 67.7 on DeepSWE, which solves development tasks, slightly above Chinese rivals but short of the roughly 74 points of top models from OpenAI and Anthropic. The company announced it would hold a launch event in San Francisco on Oct. 22.

Three unannounced names, "aurora-red-alpha," "aurora-green-alpha" and "aurora-blue-alpha," briefly appeared in the model list of OpenAI's API on Oct. 18 and soon disappeared. A few hours later, models with the same names were reported to be undergoing image generation tests on "Arena," an evaluation site where users compare anonymous models. Some see them as the next image generation model, "GPT Image 3," but OpenAI has announced nothing. On Oct. 24, other anonymous image models were also estimated to be made by OpenAI.

China's DeepSeek released "DeepSeek-V4.1-Flash-1017," an improved version of its main small model, on Oct. 17. It is available on the web, in the app and through the API, and the weights were also released. According to the company's own evaluation, the accuracy on DeepSWE, which tests solving software development tasks, rose from 74.2% for the previous version to 77.0%, and the cost per task was just under $0.20, below top U.S. models. API pricing was left unchanged. All figures come from the company's own evaluation, and U.S. media later reported that it was trained on outputs from OpenAI's model.

China's Meituan released "LongCat-2.5," a large language model aimed at handing long tasks to AI, in a Nov. 1 post on its official blog. It is the finished version of the preview released in September, handles images and text in a single architecture and can be used from Claude Code and Codex. For a limited time, API pricing is $0.30 for input and $1.20 for output per million tokens, far below Claude Sonnet 5.5's $2 for input and $10 for output. In benchmarks Meituan published, it comes close to the previous-generation "Claude Opus 5" on some items but falls short on some software development evaluations. The comparison did not include the latest U.S. models, and there is no third-party evaluation yet.

An image purporting to show that the name "Mythos 5.5," seen as the next version of "Claude Mythos," which Anthropic calls its most capable model for cyber defense and life sciences research, appeared in the HTML of the company's official website was posted on X on Nov. 1 (Nov. 2 Japan time). Mythos is the same model as the publicly available "Fable," and both previous times the two were released on the same day. However, the image came from an anonymous tip, and when ORIENS retrieved the page on Nov. 3 (Japan time), it still showed the current "Mythos 5.1." Anthropic has not made an announcement. The company is in the process of going public, and cybersecurity stocks were sold off on the day Mythos's existence was reported in March.

Posts of animations drawn entirely in code without video assets, said to have been made with "Claude Fable 5.5," reportedly Anthropic's next model, piled up on X on Nov. 2. The posters claim they were made in about an hour. Anthropic has not announced "Fable 5.5," and there is no way to verify from the outside which model made them. Speculation about the release timing has kept missing since October, and the company is aiming to go public in November.

"GPT-6.2 Astra," seen as a new version of OpenAI's top model "GPT-6 Astra," appeared in the usage logs of some users of its AI coding service "Codex." The model selection screen still showed GPT-6 Astra, leading to speculation that it is a test in which some tasks are handled by the new model without telling users. OpenAI has not made an announcement, and many aspects of where the screenshots came from cannot be verified. In September, the company canceled the release of "GPT-6.1 Astra," citing the results of safety tests. With rival Anthropic having filed to go public and talk of a new model from it as well, prediction markets put the probability of a release between Nov. 2 and 8 at 37.3%.

The X account @Haleeeemahh posted on Oct. 31 that "Carbon-H," seen as a new version of Google's top model Gemini 4 Carbon, was appearing under the name "Argon 4" on Arena, a site that compares AI models anonymously. Testers say it is faster than the earlier version and better at 3D rendering. Google has not commented, and Carbon is still offered only to organizations on the cyber defense side.

"gpt-6-astra-law," seen as the API version of "Astra for Law," the AI for law firms that OpenAI announced in September, appeared in the API model list on Oct. 31. On the same day it was also listed on OpenAI's official pricing page, priced at $10 for input and $50 for output per million tokens, the same as the general-purpose GPT-6 Astra. Availability to general developers has not been announced. In a third-party legal research evaluation, the underlying GPT-6 Astra scored below models from Google and Anthropic, and its cost per question was higher than that of Google's top-ranked model.

LuminaBench, an X account that tracks AI model developments, posted on Oct. 31 that "GLM-5.5," the next-generation model from China's Zhipu (known overseas as Z.ai), is in the final stage of post-training, the finishing training that follows pretraining. Three internal models are to be merged into the next base model, according to the post. The poster denied the interpretation that the three models would be released separately. Zhipu had not made an announcement as of the night of Nov. 2. In June, a forecast of an August release was reported, but what came out in August was GLM-5.3, an improved version on the same base.

U.S.-based Mercor announced on Oct. 30 that Google's top model, Gemini 4 Carbon, ranked first on "APEX-Agents," an evaluation that has AI carry out the real work of investment banking, consulting and law. It completed 84.2% of tasks, and second place went to Google's Gemini 4 Argon (82.4%). However, the gap between the two falls within the margin of error Mercor gives, and on another third-party evaluation of the same kind of work, Carbon placed only fourth.

U.S. research firm SemiAnalysis on Nov. 1 released "GDPx," a new benchmark that measures how far AI can independently complete long tasks performed by professionals. Across 206 tasks in 11 domains, even the top-ranked model, Anthropic's "Claude Opus 5.5," scored only 12.0%. Google's "Gemini 4 Carbon" placed second at 10.9%, at a cost of $16.29 per task, about a third of Opus's $51.41. Excluding three domains heavily affected by safety features, OpenAI's "GPT-6 Astra" takes the top spot.

A new category, "cyber," that permits offensive and defensive security content has appeared in the definitions of the developer API for xAI's conversational AI "Grok." An X account that automatically tracks AI changes posted about it on Nov. 1. It reportedly works only with API keys that xAI has enabled, while other keys get a permission error. OpenAI, Google and Anthropic offer less restricted models only to vetted defensive organizations, and xAI may be preparing the same approach. xAI has made no announcement, and the eligibility and conditions are unknown.

Settings for "Qwen 4," an unannounced model from China's Alibaba, were added on Oct. 31 to "Qwen Code," an AI coding tool developed by the company's team. Using "Qwen4-Max" and "Qwen4-Plus" as examples, they let users choose how deeply the AI reasons from three levels: low, medium and xhigh. A setting giving a context window of 1 million tokens was also added on Oct. 28. However, the settings work by matching names and do not mean the specifications are final. As of the night of Nov. 1, Alibaba had not announced Qwen 4. Attendees of an Oct. 29 event reported that the company's chief technology officer said "next week."

"DeepSeek V4.1 Pro," a top-tier model under development by China's DeepSeek, has appeared as an evaluation target in benchmark settings. LuminaBench, an X account that tracks AI models, posted about it on Nov. 1. DeepSeek announced in September that V4.1 Pro was coming but has not disclosed its release date, pricing or performance. A smaller version released shortly before has been reported to have been trained on outputs from OpenAI's models.

OpenAI on Oct. 26 (Oct. 27 Japan time) released the AI model "GPT-6.1 Terra." A model for everyday work that is cheaper than the higher-end "GPT-6.1 Sol," it lowers developer API pricing to $1 for input and $5 for output per 1 million tokens, less than half that of its predecessor ($2 for input, $12 for output). It also undercuts rival models from Anthropic and Google ($2 for input, $10 for output). OpenAI's own measurements put its performance close to the top-end GPT-6 Astra, but in a third-party overall evaluation it falls 4 points short of Sol. The announcement came on the same day Anthropic filed to go public.

Google on Oct. 28 announced Gemini 4 Carbon, its new top-end model. It took first place in an overall evaluation, but the first to get access are limited to organizations defending against cyberattacks. Following Anthropic and OpenAI, all three major U.S. companies have now adopted the same release approach.

OpenAI announced GPT-6.1 Sol on Sept. 29, 2026. An update to GPT-6 Sol, it improves performance in fixing code, reading PDFs and operating apps, the company said. On eligible plans such as Plus and Pro, it is available in ChatGPT Work and Codex.