October 2025 was a loud month for artificial intelligence. Models shipped almost weekly, two companies put AI browsers in front of the public, and one of the largest corporate restructurings in tech history got signed on a Tuesday.
If you want the AI news today without the noise, this roundup sticks to announcements with dates, prices, and named sources. Same standard we hold for artificial intelligence news today every week. No predictions, no vibes, just what happened.
We cover this beat year-round in LLM news. Background on how we work sits on our about page. This is AI news today October 2025, roughly in order.
Model Releases: What Actually Shipped
Model drops are where AI news today usually starts, and October’s first big one came on October 15, when Anthropic released Claude Haiku 4.5. Pricing runs $1 per million input tokens and $5 per million output, about a third of what Sonnet 4 cost in May, at more than twice the speed.
It scored 73.3% on SWE-bench Verified, just ahead of Sonnet 4 at 72.7%, and Anthropic gave it to free Claude.ai users. Unusual move for a model this capable. Read the announcement.
IBM shipped Granite 4.0 on October 2, a hybrid Mamba and transformer family that needs far less memory than conventional large language models. It is Apache 2.0 licensed, cryptographically signed, and the first open model family to clear ISO 42001 certification after a months-long external audit.
The lineup is small on purpose: a 32B total parameter mixture of experts with 9B active, a 7B total with 1B active, and a 3B dense model. IBM built these for agentic workflows where latency and cost matter more than raw scale.
OpenAI added gpt-oss-safeguard on October 29. Two open-weight reasoning models, 120b and 20b, fine-tuned from gpt-oss for safety classification tasks and released under the same Apache 2.0 terms, with a short technical report attached.
Cognition shipped SWE-1.5 the same day, a frontier-size coding model served on Cerebras hardware at up to 950 tokens per second. Cognition claims that is 6x faster than Haiku 4.5 and 13x faster than Sonnet 4.5. It is live in Windsurf now.
NVIDIA used its GTC DC keynote on October 28 for a stack of open releases: Nemotron Nano 3, a 32B mixture of experts with 3.6B active parameters, plus the Nemotron Nano 2 VL vision-language model, a compact document parser, and a suite of RAG models.
Also in that batch, Llama 3.1 Nemotron Safety Guard 8B V3, which handles 23 safety categories across nine languages and reports 84.2% accuracy on harmful content classification.
Apple got into the act on October 15 with the M5 chip. Each GPU core carries a Neural Accelerator, and Apple claims over 4x the peak GPU AI compute of M4, with unified memory bandwidth up to 153GB/s. It ships in the 14-inch MacBook Pro, iPad Pro, and Vision Pro.
The pattern in this stretch of the AI news today is pricing, not benchmarks. Near-frontier coding moved from premium tiers to free tiers in about five months.
Product Launches: Browsers, Agents, Work
Product news is where the latest AI news today showed up most visibly, and OpenAI launched ChatGPT Atlas on October 21, a browser built on Chromium with ChatGPT parked in a sidebar. It started on macOS for Free, Plus, Pro, and Go users, with agent mode in preview for Plus, Pro, and Business plans.
Agent mode is the part that matters. ChatGPT gets a cursor in your browser and clicks through tasks while you watch, or while you do something else. Sam Altman framed the whole thing as letting people chat with a page. OpenAI’s launch post has the details.
Reuters counted roughly 800 million weekly ChatGPT users behind that push. Perplexity made its Comet browser free in the same window, so the field now includes Chrome, Edge, Brave’s Leo, Comet, and Atlas.
AI news today increasingly reads like this: distribution wars, not model wars. The theme of the AI updates today around products is placement: the model layer commoditized fast, so the fight moved to the surface where people actually type.
Google’s October batch was wide. Gemini Enterprise launched as a governed front door for workplace AI, with HCA Healthcare and Best Buy named as early adopters. A Gemini 2.5 Computer Use model went out through the Gemini API so agents can navigate websites and fill forms.
Gemini for Home replaced the Google Assistant on speakers and displays with a more conversational version. Google also added vibe coding to AI Studio, which wires up models and APIs from a plain description of the app you want.
Sundar Pichai confirmed at Dreamforce on October 17 that Gemini 3 is still due before the end of the year, after months of release-date speculation.
Research: Evidence Instead of Demos
The AI news today with the most scientific weight came from labs rather than product teams.
Google DeepMind’s Cell2Sentence-Scale 27B, built on the Gemma family, screened more than 4,000 drugs and flagged silmitasertib as a conditional amplifier of antigen presentation. That mechanism had never been reported in the literature. Lab tests in human neuroendocrine cells, a cell type the model had not seen in training, showed roughly a 50% increase when combined with low-dose interferon.
In plain terms: the model proposed a new way to make cold tumors visible to the immune system, then the wet lab checked it and it held up.
The same month brought Quantum Echoes, which Google calls the first algorithm to reach verifiable quantum advantage on real hardware. It computed the structure of a molecule about 13,000 times faster than the fastest classical supercomputers.
Google DeepMind also partnered with Commonwealth Fusion Systems on TORAX, a fast differentiable plasma simulator written in JAX. CFS is using it to run millions of virtual experiments before the SPARC tokamak ever switches on.
Genesis Therapeutics announced Pearl on October 28, a generative 3D foundation model for drug discovery that the company says clearly surpasses AlphaFold 3 on protein-ligand complex prediction. NVIDIA published CodonFM the same day, an RNA foundation model trained on 131 million protein-coding sequences from 22,000 species, available at 80M, 600M, and 1B parameters.
This is the AI news today October 2025 material that will still matter in five years. Benchmarks get argued over; a validated wet-lab result does not.
Money and Corporate Moves
The largest business story of the month, and the heaviest AI news today headline, landed October 28: Microsoft and OpenAI announced a restructuring that converts OpenAI into a public benefit corporation controlled by a nonprofit, drops the old restrictions on raising capital, and leaves Microsoft holding a 27% stake in OpenAI Group PBC.
Altman said on the livestream that an initial public offering looks like the most likely route. Reuters broke it. A $500 billion company can now raise and list without the nonprofit ceiling in the way.
No single item in the AI news today pile moved more capital than that. Anthropic told Reuters its annualized revenue run rate was approaching $7 billion in October, up from more than $5 billion in August, against a $183 billion valuation. Internal projections put 2026 targets between $20 billion and $26 billion.
Nvidia crossed a $5 trillion market capitalization during the month. Its GTC DC announcements included seven new Department of Energy supercomputers, a robotaxi venture with Stellantis and Foxconn, and 6G work with Nokia, Cisco, and T-Mobile.
Intel brought its Fab 52 plant in Arizona online, with data center products shipping early next year, and Microsoft kept technology access to OpenAI’s work through 2032.
Policy: Brussels Rewires the AI Act
Regulation is a standing item in any AI news today briefing now, and October delivered two real chapters.
On October 8 the European Commission launched the Apply AI Strategy and the AI in Science Strategy together, mobilizing roughly €1 billion. The Commission’s own numbers set the motive: only 13.5% of EU businesses use AI right now.
It wants an AI-first default across healthcare, energy, mobility, manufacturing, defense, and the public sector, plus support for SMEs that have been slow to adopt anything. The package also opened the AI Act Service Desk, a single place to check whether a company falls under the rules and what to do next.
The AI Board met on October 24 to review national implementation and the coming Digital Omnibus, which the Commission has scheduled for November 19.
Industry wants the high-risk obligations, set for August 2026, pushed back. Germany’s government sent a ten-page letter to Brussels on October 30 asking for a one-year delay to 2027, arguing the technical standards will not be finished in time. Civil society groups are pulling the other way, calling the simplification drive deregulation in disguise.
Rules for general-purpose AI models, in force since August 2, 2025, were not part of the delay conversation. Those stay where they are.
The Pushback
The skeptical strain of AI news today found its loudest voice in Andrej Karpathy, co-founder of OpenAI, who supplied the month’s most-quoted reality check. In a Dwarkesh Patel interview published October 21 he said plainly that the models are not there, put AGI at least a decade out, and called a lot of current agent output slop.
His argument was specific. Demos and leaderboards reward short tasks, while long-horizon planning, structured reasoning, and reliable tool use remain weak. Keep overselling agents, he warned, and software fills up with bad code and security holes.
That sits next to Jensen Huang calling 2025 the year of AI agents, and Dario Amodei predicting AI better than almost every human at almost everything by 2026 or 2027.
Both can be true at once. Capability is arriving faster than reliability, and cost is falling faster than either. That gap is the honest version of artificial intelligence news today.
Worth a Line
A few more AI updates today before the takeaways, all confirmed at the source.
DeepSeek published DeepSeek-OCR in late October, an optical character recognition model that packs information into images so a model reads fewer tokens. Cheap memory, if the approach holds up.
Walmart partnered with OpenAI so shoppers can check out inside ChatGPT, and started adding AI shopping tools to its own app alongside its Sparky chatbot.
Starbucks rolled out Green Dot, an AI barista assistant that works behind the counter, while its CEO made clear no robot is making your drink anytime soon.
GM said it wants Gemini technology in its cars as part of a plan to turn the vehicle into an intelligent assistant by 2028.
FAQs on AI News Today
What is the biggest AI news today from October 2025?
The Microsoft-OpenAI restructuring announced October 28 is the biggest by impact. It turns OpenAI into a public benefit corporation controlled by a nonprofit, removes limits on capital raising, hands Microsoft a 27% stake in OpenAI Group PBC, and puts an IPO on the table for a company valued around $500 billion. For AI news today October 2025, nothing else changed the map that much.
Which new AI models launched in October 2025?
The list runs long. IBM Granite 4.0 on October 2, Claude Haiku 4.5 and Apple M5 on October 15, Nemotron Nano 3 and CodonFM on October 28, then gpt-oss-safeguard and Cognition’s SWE-1.5 on October 29. The pricing news is the real AI updates today, and the latest AI news today on cost is Haiku 4.5 at $1 per million input tokens for near-frontier coding work.
What did Europe decide about AI regulation in October 2025?
The regulatory half of AI news today looked like this. Two strategies landed October 8, Apply AI and AI in Science, backed by about €1 billion. The AI Act Service Desk opened, the AI Board met October 24, and the Digital Omnibus proposal is due November 19. Germany wants high-risk rules delayed from August 2026 to 2027. General-purpose AI model obligations, live since August 2, 2025, were not on the table for delay.
Where can I follow AI news today without the hype?
Start with primary sources: company newsrooms, arXiv papers, regulatory dockets, and court filings. The latest AI news appears there first, while aggregators and social threads compress fast and drop the dates. We link the original announcement in every story on the LLM news desk, and we correct anything that turns out wrong, so tips are welcome.
Did AI research really advance in October 2025?
Yes, and the research half of AI news today held up better than usual. Not just on charts, either. Cell2Sentence-Scale proposed a cancer therapy pathway that held up in lab tests, Quantum Echoes hit verifiable quantum advantage on hardware, Pearl beat AlphaFold 3 on protein-ligand prediction in its report, and CodonFM moved RNA modeling forward with open weights. Independent replication is still pending on several of those, which is normal this soon after publication.
What This Month Adds Up To
The shape of the latest AI news today is easy to read once you line the dates up. Models got cheaper, interfaces went agentic, research results moved closer to wet-lab validation, and the money reorganized around an IPO nobody was sure would happen.
Frontier-class coding ability dropped to a third of its May price in five months. Browsers became the battleground. Europe spent a billion euros pushing adoption while arguing with itself about whether to slow its own rules down.
Bookmark this page for AI news today as it updates. Corrections, tips, and story pitches go through our contact page, and we answer.
