AI news, 7 November 2026: the cost of generating things fell again.

Three of the five stories in this window are price stories wearing a product announcement's clothes. Google took generated video to $0.07 a second and raised the clip ceiling to 15 seconds. Gradium published a 65MB speech model, Apache 2.0, that starts talking 28 milliseconds after you ask it to, on a CPU. AWS started letting one GPU hold up to seven models instead of one. The other two are about who is visiting your website and whether anyone is measuring accuracy on your accent: Cloudflare began treating AI agents as a visitor class of their own, and Hugging Face's speech leaderboard committed to reporting Indian-accented English separately from American.

Every number below was read at the source linked in its own section during this run, and the stories that did not survive that test are named near the end. AcquihireTech writes this roundup for Indian service businesses doing ₹10L+ a month, so each story gets the same question put to it: what does this change about how you run the business on Monday?

1. Google cut generated video to seven cents a second

What happened. On 5 November 2026 Google announced that Gemini Omni 1.2 Flash is generally available on the Gemini API and Google Cloud. It is priced per second of video, at $0.07 a second for 720p output, which Google states is 30 percent below Omni 1.1. Clip length is now settable in one-second increments from 3 to 15 seconds; Omni 1.1 stopped at 10. Drafts can be generated at 360p or 720p and upscaled to 1080p or 4K.

The rest of the release is about control rather than cost. Google lists better adherence to character, product, style, motion and audio references, stronger consistency from the first frame to the last and across extended clips, multi-turn conversational editing so a clip can be refined instead of regenerated from scratch, more natural dialogue with clearer lip-sync and multi-speaker support, and fewer jump cuts on complex camera moves. Access runs through Google AI Studio on the Gemini API side and the Vertex AI console plus an enterprise API on Google Cloud, where Provisioned Throughput is supported. It also turns up in Flow, in Google Vids at no extra cost with a Google account or Workspace subscription, in Google Ads, and in the Gemini app rolling out to Google AI Ultra subscribers. Every output carries an imperceptible SynthID watermark. Google says fast sports action, rapid script changes and complex physics are still being worked on, and the announcement of 5 November 2026 publishes no benchmark results at all.

Why it matters. Do the arithmetic, because it is the whole story. A full-length 15-second 720p clip costs $1.05 to generate. Ten variants of it cost ten dollars and change. At that price the generation step stops being a budget line and becomes a rounding error, which means the constraint moves entirely to the two things money cannot buy in bulk: knowing what the clip should say, and having someone competent look at it before it goes out.

What it means for a service business. If you are an interior firm or a clinic that pays an agency for short social videos, the honest read is that the shooting and rendering part of that quote is now contestable and the scripting and judgement part is not. Two practical consequences. First, if you buy video, ask your supplier what they are actually charging you for, and expect a straight answer about script, direction and review rather than render time. Second, if you make video, the cheap thing to do badly is now very cheap indeed, so your reviewer is the product. We would set a hard rule before you generate anything: no clip goes out without a named human who approved it, and the SynthID watermark means you should assume provenance is checkable whether or not you disclose it.

2. A 65MB speech model that runs on a phone, under Apache 2.0

What happened. A Hugging Face blog post dated 5 November 2026, written by Olesia Krizhevich of Gradium with Laurent Mazare, Eric Bezzam and Eustache Le Bihan, introduced Phonon, an on-device text-to-speech model in two sizes. Phonon 31M is about 65MB and roughly 30.8 million parameters, English-only, published as gradium/phonon-31m-en, and licensed Apache 2.0, which the post states plainly means you can use it free even in a commercial product. Phonon 40M is about 75MB and covers English, French, German, Spanish and Portuguese, but it and the other Phonon models are CC BY-NC: free for non-commercial projects, with commercial use priced by number of users.

The speed claim is the interesting one. On the Open TTS Leaderboard's CPU streaming benchmark, the post reports the 31M and 40M English models with the fastest 95th-percentile time-to-first-audio of any entry, at 28ms and 38ms, against 165ms for Pocket TTS. On intelligibility it reports first place in German and Spanish and a top-three placing in every language evaluated, competing with models up to 100 times larger. It ships more than 100 built-in voices and supports voice cloning and text-described voice design through Gradium Studio. Everything needed to try it is public: weights on Hugging Face with precomputed voices at gradium/phonon-voices, a browser demo in the gradium/phonon Space, the reference PyTorch implementation in Hugging Face Transformers from version 5.21, and Gradium's own inference framework ptts on GitHub with more than 10 CPU, GPU and NPU backends. The full write-up is on the Hugging Face blog.

Why it matters. Twenty-eight milliseconds to first audio, on a CPU, with no network call, is the difference between a voice agent that sounds like a conversation and one that sounds like a form being submitted. The dead air after a caller stops speaking is what makes most automated calls feel automated, and a large part of that gap has always been round-trip time to someone's API. A 65MB model removes the round trip entirely.

What it means for a service business. Say the obvious limitation first: there is no Hindi, no Tamil, no Indian language of any kind in either size, and the only Apache 2.0 build is English-only. If your callers want to be spoken to in Hindi, this does not replace your current vendor and nothing in the post suggests otherwise. Where it does land is the English-language leg of your call flow — outbound confirmations, appointment reminders, the "we have received your enquiry" callback — and the cost structure there changes shape, because a model this size runs on hardware you already rent. The test worth running this month is narrow and cheap: take the twenty sentences your system says most often, generate them with Phonon 31M on the box your automation already runs on, and listen. If they pass, that is a recurring per-character API bill that becomes a fixed one. Check the licence on the exact checkpoint you ship, because 31M and 40M are not under the same terms.

3. Cloudflare started treating the agent on your website as its own kind of visitor

What happened. On 5 November 2026 Cloudflare announced Trusted Agent Delegation, a coalition of ecosystem representatives convened to work through the problem of agent identity on the web and carry well-defined pieces of it to the IETF, with side meetings at the IETF and workshops planned. The post is explicit that the full delegation mechanism does not exist yet. What it sets out is the shape of the problem: a site should be able to tell when one agent is driving another inside a browser, and should be able to link an agent to the user it is acting for. Cloudflare proposes graduated identity levels (Nothing, Session, Account, Unique), with claims such as being over 18 supplied by operators or third parties, and notes that a machine-readable standard for site access policy is still missing.

Three capabilities are already in production. Security Analytics now shows Personal Assistants as their own traffic category, with a widget breaking out each agent's request volume, the paths it hit and the errors it got. AI crawler and agent access policies can be set per agent: allow one sitewide, block it on ad-bearing pages, or block it outright, with Search and Training policies configured separately. Account Abuse Protection, in early access, reports whether a personal agent made a login or sign-up request and lets you apply a WAF rule to a Hashed User ID, so one abusive account can be stopped without blocking that agent for everyone who uses it. Cloudflare names Web Bot Auth as its preferred mechanism for agents to sign requests and identify themselves and their operator, and references MOLE, x402, Privacy Pass and Private Access Tokens, Poppy, the IETF's AI preferences work, OAuth, WIMSE and Proof of Human. Business Insights is for Enterprise Bot Management customers and Account Abuse Protection is early access; the post gives no pricing and publishes no agent-traffic statistics, saying only that agentic traffic is growing quickly and may pass human browsing in the near future.

Why it matters. This is the first mainstream articulation of a problem most website owners have not noticed they have. Your bot controls were built on one assumption: traffic is either a human you want or a scraper you do not. An agent filling in your enquiry form on behalf of a real buyer is neither, and every tool you currently own will classify it as the second one.

What it means for a service business. You almost certainly do not need Enterprise Bot Management, and that is not the point. The point is that "block all bots" has quietly become a decision with a cost attached, and somebody should make it deliberately rather than inherit it from a plugin default. Three things worth doing without buying anything. Look at whether your WAF or hosting panel currently blocks unrecognised clients on your enquiry form, because that is where a buyer's agent will land. Separate in your own mind the three things you may want to allow or refuse — training on your content, inclusion in AI search answers, and an agent acting for a named buyer right now — since they are different questions and Cloudflare now configures them separately. And if you run anything with a login, note the direction of travel: the useful control is at the level of the user behind the agent, not the agent itself.

4. AWS let one GPU hold up to seven models

What happened. On 6 November 2026 the AWS Machine Learning Blog, in a post by Shantanu Tripathi, Apoorva Chandra and Dmitry Soldatkin, announced that SageMaker Inference supports GPU partitioning through NVIDIA Multi-Instance GPU. One physical GPU divides into isolated partitions, each with a fixed share of compute and its own dedicated memory, enforced in hardware, and each partition can hold a different model.

The constraints are as important as the feature. It works only on endpoints that use inference components, and not on single-model endpoints, serverless endpoints or heterogeneous instance pools. The partition count per GPU depends on the accelerator and never exceeds seven: P4d with 40GB A100s and P5 with 80GB H100s both go to seven, G7 with the 32GB RTX PRO 4500 Blackwell goes to two, and G7e with the 96GB RTX PRO 6000 Blackwell Server Edition goes to four. The layout is fixed when the endpoint is created, and a model needing multiple GPUs still needs whole devices. AWS's worked example puts mig-4g.20gb, mig-2g.10gb and mig-1g.5gb on each of the eight A100s in an ml.p4d.24xlarge for 24 partitions, eight of each type, and gives an 8B model at FP16, about 16GB of weights, serving an 8,000-token context inside a 40GB partition. Carving the accelerators added roughly 17 seconds to endpoint provisioning in the authors' own testing. Enabling it means adding AcceleratorPartitionLayouts to a production variant in CreateEndpointConfig and AcceleratorPartitionType to an inference component's compute resource requirements; both fields are optional and there are no new API operations, resource types or IAM actions. The stated motivation is that small models, such as fine-tuned variants, embedding models, rerankers and classifiers, leave most of an accelerator idle. Salesforce is quoted saying the feature would significantly improve GPU utilisation, without a number attached. The post publishes no utilisation or cost figures.

Why it matters. This is the mechanical answer to a question every buyer of AI services should be asking: why does inference cost what it costs? Until now the smallest unit you could rent was a whole accelerator, so a reranker or a classifier got a 40GB card whether it needed one or not. AWS's own worked example puts 24 partitions on the eight A100s in a single instance, and its sizing note has an 8B model at FP16 serving an 8,000-token context inside one 40GB slice. Nothing about model quality changed; the economics of serving several small models did.

What it means for a service business. You will not touch a SageMaker endpoint, and you do not need to. Use it as a question for whoever bills you. If an automation supplier is charging you for dedicated GPU capacity to run a document classifier or an embedding model for your own search, this is the thing they should be doing with that card, and it is reasonable to ask whether they are. The broader pattern is worth holding onto, because it keeps repeating in this roundup: the layer where costs are falling fastest is the plumbing, not the model. That is also the layer where a vendor's margin is least visible to you.

5. Hugging Face's speech leaderboard will start reporting Indian-accented English separately

What happened. A Hugging Face post dated 6 November 2026 by Eric Bezzam, Ashley Mo and Aoi Otani, with Liva AI credited as collaborating partner, added Egyptian and Nigerian English to the Open ASR Leaderboard. The datasets are Liva AI Conversational EG, 1.34 hours of Egyptian English, and Liva AI Conversational NG, 1.55 hours of Nigerian English, both conversational and spontaneous, with punctuated and cased transcriptions that keep the disfluencies in. The structural change matters more than the two datasets: the leaderboard's Private tab will now break results down by accent using Kachru's Three Circles, reporting US against non-US, and Inner Circle — the United States, the United Kingdom, Australia, Canada — against the Outer and Expanding Circles, where India, Nigeria and Egypt sit.

Two limits, stated plainly because the post states them by omission. Those durations are small by the standards of speech benchmarks. And the post names no models, publishes no word error rates and reports no results, so this is a change in what will be measured, not a finding about which recogniser handles Indian English best.

Why it matters. Speech recognition vendors quote a single accuracy number. That number is overwhelmingly earned on American English, and anyone who has watched a transcription tool work on a Delhi sales call already knows the gap is real without being able to cite it. A public leaderboard splitting Inner Circle from Outer Circle results is the beginning of being able to cite it.

What it means for a service business. This one is directly actionable and costs nothing. Stop accepting a single word error rate from any voice or transcription vendor, and stop quoting one in your own proposals. Ask which accents the figure was measured on, and if the answer is vague, run your own test: take ten real recordings of your own calls, with your own staff and your own customers, transcribe them, and count the errors that would actually have broken something: a wrong phone number, a wrong name, a wrong amount. Ten calls is enough to tell you whether a vendor's number travels, and our position is that no voice deployment should go live without that test, because a figure earned on American English has no obligation to hold on an Indian accent over a bad line.

The window, and the stories we cut

The window for this edition is 5 to 6 November 2026. Nothing dated 7 November could be opened at a primary source, which is expected on two counts: 7 November is a Saturday, and the US day had not finished when this went out. The five stories come from five different originators (Google, Gradium, Cloudflare, AWS, and Liva AI with the Hugging Face team), though Hugging Face is the publisher of two of them. None appeared in the 5 November edition.

The biggest cut is also the most frustrating. OpenAI is being widely reported as having launched GPT-6 Cyber on 5 November and ChatGPT Images 2.6 on 6 November, and both are probably real. Neither appears here, because openai.com returned HTTP 403 to every automated request made during this run and the developer-community thread for the images release returned a 404, so no OpenAI page could be opened to confirm anything. Aggregators are circulating specific figures for GPT-6 Cyber — a Critical rating under OpenAI's Preparedness Framework and 48.5 percent of ExploitGym tasks solved against 32.1 percent for GPT-5.6 Cyber. Those numbers may well be right. They are not repeated as fact here, because this roundup's rule is that a figure gets published only if the source was opened in the run, and a model with a Critical cyber rating is exactly the kind of claim that should not reach you through a newsletter quoting a newsletter.

One genuinely verified story was left out for being about someone else's business rather than yours. NVIDIA announced Vera on 5 November, a CPU built for agent workloads, and the numbers are substantial. In NVIDIA's own 100-megawatt data-centre comparison it claims 1.6x more agents per watt than the fastest x86 CPU, and 2x faster agents than the most energy-efficient one. The partner results are the striking part: Daytona brought 1,000 agent sandboxes online in 10.8 seconds against 88.4 seconds on x86, Perplexity ran agentic coding 1.5x faster, and DeepInfra reported up to 2.2x the speed of the best x86 processor on agentic orchestration. All of that is NVIDIA's own framing of partner tests, and all of it matters to whoever sells you agent capacity rather than to you this quarter.

Three more were verified only to the level of a title and a date on a publisher's own index, which is not enough to write a section on. NVIDIA published Cosmos 3 open models for interactive world simulation on 5 November and the Allen Institute published ToyBlocks, a controlled world for robot learning, on 6 November; both are robotics and simulation research. GitHub's changelog for 6 November carries Copilot code review API management updates, a Copilot impact dashboard that now reports chat and VS Code agent usage, and general availability of @GitHub in Slack and Teams — developer tooling housekeeping. Cloudflare's other post that week, on the post-quantum future of web privacy, is a good piece and is not about AI.

The short version, in the order we would do it. Run twenty of your most-repeated English sentences through Phonon 31M on the machine your automation already uses, and check the licence on the exact checkpoint before you ship it. Price a 15-second video at $1.05 of generation and work out who in your business is the named reviewer, because that role is now the expensive one. Check whether your WAF or hosting panel is blocking unrecognised clients on your enquiry form. Ask your automation supplier what identity and what hardware your small models run on. And take ten real recordings of your own calls and count the errors your transcription vendor actually makes on them, rather than the ones in their marketing.