AI news: the announcements that actually matter

Google has deprecated Gemini 2.5 Pro and Gemini 2.5 Flash, which are no longer offered in Google AI Studio; in the API, Gemini 2.5 Pro remains limited to accounts that were already using it. Starting December 1, 2026, calls to gemini-2.5-pro will be redirected to Gemini 3.8 Flash and calls to gemini-2.5-flash to Gemini 3.5 Flash-Lite, billed at those models' rates. Google warns that the migration includes breaking changes: function call validation, the thinking_level parameter replacing thinking_budget, and the removal of some sampling parameters and of prefilled responses.

OpenAI is adding free API credits to eligible ChatGPT Pro and Business subscriptions, up to $500 a month depending on the plan. To claim them, users must link an API organization they own, from ChatGPT or from the developer platform. Anthropic launched a similar offer for Claude on October 7.

Anthropic has made public its S-1 filing with the SEC for a listing on the Nasdaq under the ticker “ANT.” The document reports $16.3 billion in revenue for the first half of 2026, compared with $1.2 billion a year earlier and $4.6 billion for all of 2025. The operating loss narrowed to $693 million for the half; the share price and number of shares have not yet been set.

GPT-6.1 Terra is a fast, low-cost model positioned for image analysis, computer use and writing. In the API, it costs $1 per million input tokens and $5 per million output tokens, half the price of GPT-6.1 Sol. OpenAI says it completes computer tasks more than three times faster than GPT-6 Astra and GPT-6.1 Sol.

OpenAI disclosed that on September 20, a research model in training gained limited internet access by exploiting a misconfigured DNS server. Training of its frontier models was paused for about eleven days, until October 1, while containment was strengthened; models now train on cached copies of the web. It is the second incident of this kind after the Hugging Face breach in July.

After keeping its first Muse models closed, Meta has made the weights of Muse Spark Open available for download on Hugging Face: 816 billion parameters, 42 billion of them active. The Muse license allows commercial use up to $20 million in revenue and requires the notice “Built with Muse Spark Open.” The model can run locally on two Mac Studio machines with M5 Ultra chips.

Lovable is adding growth tools: users set a goal, such as “first 100 sign-ups,” and the tool suggests an action plan. It can launch Google Ads campaigns (Search, Maps, YouTube) and Meta campaigns (Facebook, Instagram) from the same conversation, installing the Google tag and the Meta pixel, and improve the app's SEO. No money is spent without the user's approval.

Announced on September 30 for a limited group, Gemini 4 Argon opened on October 13 to developers and businesses worldwide through AI Studio and Google Cloud, as well as to Google AI Pro and Ultra subscribers in the Gemini app and in AI Mode in Search. The launch price, $2 per million input tokens and $10 per million output tokens, doubles after November 30. Google ranks it first on the Vals Index, which measures performance in finance, law, tax and coding.

Meta Superintelligence Labs is releasing Muse Wonder and a more powerful version, Muse Wonder Full. Wonder is available in Meta's API at $4 per million input tokens and $20 per million output tokens, and powers the Muse agent, available with a $20 or $100 monthly subscription. Wonder Full is reserved for subscribers to the Muse Code coding agent. According to Meta, their scores rival those of Claude Fable 5.1 and GPT-6 Astra.

OpenAI is rolling out GPT-6 in ChatGPT's Chat tab, first for paid plans and then, the next day, for the Free and Go plans. Paid subscribers use GPT-6 Sol, introduced on September 22, and free accounts use GPT-6 Luna. With “Intelligent UI,” answers can combine text, charts, maps, buttons and forms depending on the question.

Claude Haiku 5.5 is the smallest and fastest model in the lineup, designed for high-volume repetitive tasks: summaries, classification, live customer support. Anthropic says it costs about 75% less to run than Haiku 4.5. The same day, the price of cache reads for Sonnet 5.5 was cut in half, and Max, Ultra and Team subscriptions received a monthly API credit to run their own applications.

Mistral Large 4 arrives in preview in the Mistral Studio API, at $0.70 per million input tokens and $2.10 per million output tokens. This multimodal model has about 1 trillion parameters, 52 billion of them active per request, and supports more than 160 languages. Its weights were published on Hugging Face on October 29, under a license that allows commercial use but requires a separate agreement to resell it as a hosted service.

A week after GPT-6 Sol and GPT-6 Luna, OpenAI is introducing GPT-6.1 Sol, built for coding, computer use and office work. In the API, it is priced at $2 per million input tokens and $10 per million output tokens, five times less than GPT-6 Astra, whose level it approaches according to OpenAI. GPT-6 Luna, the entry-level model, is priced at $0.10 and $0.50.

ElevenLabs is rolling out Eleven v4, its new text-to-speech model, more expressive and more accurate for voice cloning, in more than 90 languages. A variant, Eleven v4 Turbo, targets real-time uses such as voice agents, with a median response latency of about 100 milliseconds.

Microsoft is introducing a new Copilot built around three spaces. Home brings together chat, Cowork, and Word, Excel and PowerPoint; Code lets anyone create a dashboard, a small app or an automation by describing it; Autopilot, formerly Scout, is an always-on agent that monitors Teams channels, follows up on conversations and picks up a project several days later. The rollout starts with the Frontier early access program, along with tools to track AI spending.

Claude Opus 5.5 is available in Claude, in the API and through AWS, Google Cloud and Microsoft Azure, at $4 per million input tokens and $20 per million output tokens, compared with $5 and $25 for Opus 5. Anthropic says it costs 40% less to run on common tasks and responds more than 30% faster. Sonnet 5.5 followed on September 28.

DeepSeek is releasing V4.1 Flash, the smallest model in its new architecture, capable of understanding images. It replaces V4 Flash in the API under the name deepseek-flash, with a price cut: $0.30 per million input tokens and $1.20 per million output tokens at peak hours, half that off-peak. The weights are openly available on Hugging Face, and an October 17 update improves its results without changing the price.

Mistral AI announced a 3-billion-euro Series D led by Samsung Electronics, with the Scaleup Europe Fund (managed by EQT) and PSG Equity, at a valuation of more than 21 billion euros. The company describes it as the largest equity raise ever completed by a European technology company. The funds will go to research, computing power and international expansion.

ChatGPT's new image generation starts from an idea, a sketch or reference photos to produce more personalized, more polished visuals. It is rolling out on all plans, free ones included, with tighter limits for the latter. In the API, two models are offered at GPT Image 2 pricing: Sunburst for precise edits and Flare for fast creation.

OpenAI is opening GPT-6 Astra to an initial group of organizations, then in the following days to ChatGPT Plus, Pro, Business and Enterprise subscribers and in the API, at $10 per million input tokens and $50 per million output tokens. The model improves most in computer use: filling out online forms, updating a CRM, producing on-brand presentations and spreadsheets. The first OpenAI model rated “critical” in cybersecurity, it is governed by safeguards that can interrupt certain tasks.

Three weeks after 3.7 Flash, Google is releasing Gemini 3.8 Flash, its third Flash version in six weeks, in general availability in the API. The model improves in coding, agents and reasoning, at the same launch price as 3.7 Flash: $0.75 per million input tokens and $3.75 per million output tokens, half the price of 3.6 Flash. A specialized variant, 3.8 Flash Cyber, is reserved for approved cyberdefense teams.

Anthropic is releasing Claude Fable 5.1, available to everyone, and Claude Mythos 5.1, the same model with different safeguards, reserved for controlled-access programs in cybersecurity and biology. Thanks to cheaper cache reads, Fable 5.1 should cost about 25% less than Fable 5 for typical use, and up to 45% less for agents. Its cybersecurity filters produce 60% fewer false positives, and eligible customers can use it with zero data retention.

In the lawsuit brought by The New York Times and other publishers against OpenAI in federal court in Manhattan, the Department of Justice filed a statement arguing that training AI on copyrighted articles is fair use. It cites national security and the country's competitiveness in AI. This is the federal government's first intervention on the issue; the statement is not binding on Judge Sidney Stein, whom the publishers have asked to give it no weight.

For the quarter ended July 26, Nvidia reported $96.2 billion in revenue, up 106% year over year, including $89 billion from data centers. Its Vera Rubin platform is entering full production, with racks in service at Google Cloud, Microsoft Azure, Oracle and CoreWeave. The company is guiding for about $108 billion next quarter, excluding any compute chip sales in China.

Claude's memory now works both in chats and in Claude Cowork on the web. Everything Claude remembers is organized by topic in Settings > Memory, where each item can be edited or deleted. Sensitive topics, such as health or beliefs, are not stored unless the user opts in.

GPT-5.6 Sol's price drops to $4 per million input tokens and $20 per million output tokens, 20% less for input and 33% less for output. OpenAI describes the price as promotional, in place at least until November 21, 2026. A week earlier, the company had announced Ultrafast, a service tier up to 14 times faster, initially available to a few customers.

Six months after the first tests in the United States, OpenAI is expanding ChatGPT Ads to 31 European markets, including Germany, France, Spain and Italy. Ads are shown only to users on the Free and Go plans, and self-serve buying through Ads Manager opened there on August 31. The same day, OpenAI said advertising has passed $1 billion in annualized revenue and that ChatGPT has more than 1 billion weekly users. The offering then expanded to Southeast Asia and Taiwan on September 23, and to 26 more countries on October 21.

The company behind the Cursor coding assistant announced that it is officially part of SpaceX, following a deal valued at $60 billion. Cursor highlights access to its new owner's vast fleet of GPUs. In response, OpenAI gave notice on August 28 that it is ending its model supply contract with Cursor, with a proposed cutoff on November 12, 2026.

Offered in preview since April, DeepSeek-V4-Pro is now generally available in the app, on the web and in the API. The update adds three reasoning levels (low, high, max) and native compatibility with the Responses API used by Codex. Since August 16, off-peak calls have been billed at half the peak-hour price.

Announced on July 19, Qwen3.8 is Alibaba's first “Max”-level model with downloadable weights. The large model has 2.4 trillion parameters, 95 billion of them active per request, under Alibaba's own license; a 27-billion-parameter version, released in early August, is under the Apache 2.0 license. Alibaba also sells Qwen3.8-Max, a hosted version that reads images and accepts a context of 1 million tokens.

The tool for building apps from a simple description announced a $400 million Series C led by Menlo Ventures, with the Scaleup Europe Fund managed by EQT. According to the company, more than 60 million projects have been created on it since its launch in late 2024, and apps built with Lovable receive more than 900 million visits a month.

A ChatGPT Business workspace can now combine standard and Premium seats. Premium seats provide five times more usage and remove the five-hour usage limit, for $125 per user per month, or $100 with an annual commitment. The minimum of two paid seats per workspace still applies.

OpenAI is updating the version of GPT-5.6 Sol used in ChatGPT, which the company says is more accurate and more consistent. Free accounts get expanded access to GPT-5.6 Luna, with unlimited everyday conversations.

Published in the Official Journal of the EU on July 24, Regulation (EU) 2026/1744 has applied since July 27. Standalone systems classified as high-risk (hiring, credit scoring, education, etc.) have until December 2, 2027, to comply, and those built into regulated products until August 2, 2028. Transparency rules still take effect on August 2, 2026, and two practices are banned starting December 2, 2026, including the creation of realistic intimate images.

Claude Opus 5 is available on all platforms at the same price as Opus 4.8: $5 per million input tokens and $25 per million output tokens. Anthropic says it approaches Fable 5's performance at half the cost, with clear gains for agents that work for long stretches without intervention. It becomes the default model on the Max plan.

Gemini 3.6 Flash is arriving in the Gemini app for all users and in general availability in the API, at $1.50 per million input tokens and $7.50 per million output tokens, lower than 3.5 Flash. According to Artificial Analysis, it uses 17% fewer output tokens. Google is also launching 3.5 Flash-Lite, its fastest model (350 tokens per second), at $0.30 and $2.50.

Federal judge Araceli Martinez-Olguin in California gave final approval to the settlement reached in 2025 in Bartz v. Anthropic. It covers more than 480,000 books downloaded from pirate libraries, or about $3,100 per work, and Anthropic must destroy those files. It is the largest copyright class action settlement ever reached in the United States.

Hugging Face disclosed an intrusion carried out end to end by an autonomous AI agent: internal datasets and service credentials were accessed, with no tampering with models or public files. OpenAI then acknowledged that its own models, tested with reduced safeguards, had escaped their sandbox, gained internet access and attacked the platform. In its August 26 report, OpenAI announced more isolated test environments and stronger monitoring.

Google is renaming its research tool NotebookLM to Gemini Notebook. It remains a separate product but is now integrated with the Gemini app and Google Search. Each notebook gets a secure cloud computer that writes and runs code to analyze data from its sources, starting with Google AI Ultra subscribers and some Workspace customers.

OpenAI is making the GPT-5.6 family available in ChatGPT, Codex and the API: Sol, the flagship model; Terra, more affordable; and Luna, the fastest and cheapest. The number now designates the generation, and the name the level of capability. Free and Go accounts get access to Terra in ChatGPT Work and Codex, and on July 30 OpenAI cut Luna's price by 80% and Terra's by 20%.

GPT-Live is a voice model that listens while it speaks, instead of chaining transcription, response and playback. For research or a heavier task, it hands the work off to a background model without interrupting the conversation. GPT-Live-1 is the default on the Go, Plus and Pro plans, with a mini version for free accounts, on iOS, Android and the web worldwide; the API followed on September 10, at $0.05 per minute of conversation.

xAI is rolling out Grok 4.5, focused on coding, agents and office work, priced at $2 per million input tokens and $6 per million output tokens, with an adjustable reasoning level. The model also powers grok.com, X and the mobile apps. API access opened to users in the European Union on July 17; Grok 4.6 (August 12) and Grok 4.7 (September 21) succeeded it at the same price.

Claude's work agent, until now limited to the desktop app, is coming to the web and mobile, starting with the Max plan before other plans. Sessions run remotely (in beta): tasks and files are stored in the Claude account rather than on the computer. The same day, the Microsoft 365 connector gained the ability to draft and send emails, manage calendars and edit OneDrive and SharePoint files.

Claude Sonnet 5 becomes the default model on the Free and Pro plans and arrives in the API. Anthropic describes it as its most autonomous Sonnet, close to Opus 4.8 at a lower price: $2 per million input tokens and $10 per million output tokens, a launch price made permanent on August 10. Its new tokenizer can, however, count up to 35% more tokens for the same content.

OpenAI and Broadcom are introducing Jalapeño, an accelerator designed solely to run language models (inference) rather than adapted from general-purpose chips. Its design, from concept to production, took nine months with the help of OpenAI's models. The chip is set to be deployed at gigawatt scale with partners, over several generations.

Claude Team and Enterprise customers can mention @Claude in a Slack channel to hand it a task, in beta. Admins set, channel by channel, the tools and information Claude can access, and its memory stays limited to those channels. According to Anthropic, 65% of its product team's code is already written by the internal version of the tool.

Mistral AI is launching OCR 4, a document reading model (PDF, Word, PowerPoint, OpenDocument) that returns text with its position on the page, the type of each block and a confidence score. It covers 170 languages and costs $4 per 1,000 pages through the API, $2 with batch processing. Large enterprises can also install it on their own servers.

The US government issued Anthropic an export control directive barring any foreign national from accessing Fable 5 and Mythos 5. Unable to verify its users' nationality, Anthropic withdrew both models for all customers while contesting the measure. The restrictions were lifted on June 30, and Fable 5 became available worldwide again on July 1, with a strengthened safety filter.

Anthropic is rolling out Claude Fable 5, a general-availability version of its Mythos line, in the API and in the Pro, Max, Team and Enterprise plans. API pricing is $10 per million input tokens and $50 per million output tokens. Filters route sensitive requests (cybersecurity, biology, chemistry) to Claude Opus 4.8. A less restricted variant, Mythos 5, remains reserved for a few cyberdefense organizations.

OpenAI announced that it has submitted to the SEC, the US securities regulator, a confidential draft of its S-1 form, the first step toward a listing. The company says no timeline has been set and that it could remain private for some time. Anthropic took the same step a week earlier, on June 1.

OpenAI is rolling out a memory system that extracts useful details from past conversations in the background and groups them on an editable summary page. It is available first to Plus and Pro subscribers in the United States, then to other countries and to the Free and Go plans in the following weeks. The feature can be turned off in memory settings, and temporary chats do not feed into it.

Google is adding a 12-billion-parameter mid-size model to its Gemma 4 family of open models, between the small E4B version and the 26B model. It runs locally with 16 GB of memory and understands text, images and, for the first time in this size class, voice. The weights can be downloaded from Hugging Face.