November 2025 AI News Archive: 12 Dated Records

November 2025 is easier to understand as a set of dated records than as a contest for the loudest launch. The month included new model families, image tools, a cloud agreement, two policy actions, a music licensing deal, and two security disclosures. Those events did not all carry the same kind of evidence. A product announcement says what a vendor released. A statute says what the law requires. An incident notice explains what an organization knew at a particular time.

This archive covers events whose operative date or public announcement fell between November 1 and November 30, 2025. The evidence was reviewed through July 18, 2026. Later corrections to a November source are labeled, but later launches and product changes are not inserted into the month. This is a selected source record, not a claim to contain every AI story published during November.

Each entry links to a company announcement, official technical repository, government notice, statutory text, or regulator publication. Vendor performance statements remain vendor statements. HUMAI did not reproduce the benchmarks, test the products, inspect the affected security datasets, or verify private deal terms.

How to use this archive

Use the date column to reconstruct sequence, then open the linked record before relying on a detail. For product work, check whether the source described a preview, a rollout, or broad access. For policy work, separate a proposal from a law that took effect. For security work, keep the date of the underlying activity apart from the date of public disclosure. That distinction prevents a later article update from silently changing the November timeline.

For follow-up research, copy five fields from a row before writing: the actor, the action, the date of the action, the source-publication date, and the status at that moment. Status words matter. “Announced,” “released,” “began rolling out,” “took effect,” “proposed,” and “disclosed” describe different levels of completion. Replacing all of them with “launched” can turn a plan into a finished deployment or a proposal into law.

A primary source is close to the event, but it is not automatically neutral. Companies know their own release scope and contract announcement; regulators know the text they administer. Those same sources also select favorable framing. This archive accepts a vendor page for the fact that its own product shipped and for the features it documented. It does not accept a vendor's benchmark chart as proof that competitors were inferior, or a party's press release as an independent valuation of a deal.

When a source changes later, save the original date and add the correction date. Do not replace the old record so cleanly that readers can no longer see what was known in November. The Mixpanel notice below demonstrates the rule: the incident, discovery, dataset handoff, public notice, and December clarification each have their own timestamp.

Compute and general models followed different clocks

November 3: OpenAI and AWS described a seven-year, $38 billion agreement. The joint announcement published by OpenAI said AWS would provide compute for OpenAI workloads, including Amazon EC2 UltraServers with hundreds of thousands of Nvidia GPUs. It also said the capacity targeted under the agreement was scheduled for deployment before the end of 2026, with room for later expansion. The November record is the agreement, not completion of that future buildout. The amount and deployment schedule came from the parties, not an independent audit of invoices or installed hardware.

November 12: GPT-5.1 began rolling out in ChatGPT. OpenAI's product notice introduced GPT-5.1 Instant and GPT-5.1 Thinking. It described adaptive reasoning changes, revised tone controls, and a gradual rollout that began with paid users. The notice also said API access would follow later that week, so the ChatGPT announcement should not be treated as simultaneous access in every product or account. OpenAI published a separate system card addendum on the same date. Claims about intelligence, speed, and conversational quality were the company's release claims.

November 18: Google introduced the first Gemini 3 model. Google's Gemini 3 announcement placed Gemini 3 Pro in the Gemini app, AI Studio, Vertex AI, Gemini CLI, and the new Antigravity development platform, with access conditions that differed by surface and plan. Google also described new generative interfaces and published its own benchmark results. The archived fact is that Google announced and started distributing Gemini 3 on November 18. The rankings in Google's charts are not presented here as independent comparisons.

November 24: Anthropic released Claude Opus 4.5. The release page listed access through Claude apps, the API, and three cloud platforms. Anthropic set the API price at launch at $5 per million input tokens and $25 per million output tokens, and added an effort control for trading cost and latency against capability. Those figures belong to the November 24 price record and should not be read as a current quote. Anthropic's coding, agent, computer-use, and token-efficiency results came from evaluations described by Anthropic and its system card.

Three visual releases solved different problems

November 19: Meta announced SAM 3 and SAM 3D. The Meta release record described SAM 3 as a model for detecting and tracking objects in images and video from text or visual prompts. SAM 3D addressed reconstruction of objects and people from a single image. Meta shared SAM 3 weights, an evaluation dataset, and a research paper, while SAM 3D received checkpoints and inference code. This was a segmentation and reconstruction release, not a general image generator, so folding it into one model leaderboard would erase the task boundary.

November 20: Google announced Nano Banana Pro. Google's launch post identified it as an image generation and editing model built on Gemini 3 Pro. The initial product record emphasized text rendering, multilingual text, reference-image use, editing controls, and output up to 4K. It also named several Google surfaces where access started. Availability limits and later integrations changed after November, so this archive keeps only the launch description and does not convert it into a current feature matrix.

November 25: Black Forest Labs released FLUX.2. The company's release article and official inference repository provide the dated record. The repository identifies FLUX.2 [dev] as a 32-billion-parameter model for text-to-image generation plus single-reference and multi-reference editing, released under the FLUX.2-dev non-commercial license. It also warns that the reference script requires substantial memory. Later FLUX.2 family additions are outside the November release and are not backdated into this entry.

Policy and contract records need narrow wording

November 10: New York's algorithmic pricing disclosure requirement took effect. The New York attorney general published a consumer alert on November 5 ahead of the effective date. The underlying General Business Law section 349-a requires a clear notice when an entity sets a consumer price using personalized algorithmic pricing covered by the statute. It does not ban every use of dynamic pricing, and the text contains exceptions. The archive therefore records a disclosure duty, not a general prohibition on algorithms setting prices.

November 19: the European Commission published a proposal to adjust parts of the AI Act. The Digital Omnibus on AI page describes targeted simplification measures and links to the legislative proposal and annexes. Publication placed a proposal into the policy process; it did not itself rewrite the AI Act on that date. Anyone reconstructing compliance obligations should use the enacted law and subsequent legislative outcome, not treat the November proposal as binding text.

November 25: Warner Music Group and Suno announced a partnership that also settled their litigation. Warner's press release said future Suno models would use licensed arrangements and that participating artists and songwriters would control uses of their names, images, likenesses, voices, and compositions. It also stated plans to deprecate then-current models in 2026, restrict downloads to paid accounts, cap paid downloads, and transfer Songkick to Suno. These were announced terms and future plans. The release did not supply the private settlement agreement or prove that every relevant rightsholder had licensed every future model.

Security disclosures ran on two dates at once

November 13: Anthropic disclosed activity it said it had detected in mid-September. In its public report, Anthropic attributed the activity with high confidence to a Chinese state-sponsored group. The company said the attackers manipulated Claude Code while targeting roughly 30 organizations, with successful intrusion in a small number of cases. Anthropic called it the first documented large-scale cyberattack executed without substantial human intervention. That characterization, target count, attribution, and assessment of automation are Anthropic's findings. The detection date does not establish when the campaign began.

Anthropic edited the page on November 14. It added another link to the full report and corrected the request-rate wording from “thousands of requests per second” to “thousands of requests, often multiple per second.” This archive uses the corrected statement. The dated November record remains the public disclosure, while mid-September is only the point when Anthropic says it detected suspicious activity.

November 26: OpenAI disclosed an incident in Mixpanel's systems. OpenAI's incident notice says Mixpanel became aware on November 9 that an attacker had exported a dataset, then shared the affected dataset with OpenAI on November 25. OpenAI published the notice the next day. It said the exposed fields could include names, email addresses, coarse locations, browser and operating-system details, referring sites, and organization or user IDs. It also said chats, API requests, passwords, credentials, API keys, payment details, and government IDs were not exposed. A December 19 clarification expanded the description of affected users; that correction is retained because it changes how the November disclosure should be read.

What the record supports

The sequence shows several kinds of change arriving in the same month: compute commitments, product rollouts, technical releases, a statute taking effect, a legislative proposal, a private-sector licensing arrangement, and post-incident reporting. It does not establish which model was best, which deal mattered most, whether planned infrastructure was completed, or whether announced product access remained unchanged.

For a research note, cite the event entry and preserve the source date. For a purchase or compliance decision, move from this archive to current product documentation or the law in force. For security analysis, keep the incident, discovery, notification, and publication dates as separate fields. That is the practical value of the archive: it preserves what was publicly recorded in November without pretending that November's record is the present.