AI news, 5 November 2026: India released its own foundation models.

The most consequential item in this window did not come from a US lab. On 3 November 2026 the IndiaAI Mission released three indigenous foundation models, two of them open source and built to run offline on hardware you already own. The same day, Menlo Ventures put a number on what everyone else is spending: $176.8 billion on deploying AI, up 5.4x in a year, with $1.70 going out the door on deployment for every $1 the industry spends building models. Anthropic quietly extended its text watermark to three more Claude models. OpenAI shipped two image models that return SVG. And AWS started handing AI agents root on a Linux desktop.

Every figure 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. India released three of its own foundation models, two of them open source

At a press event in New Delhi on the evening of 3 November 2026, the IndiaAI Mission under the Ministry of Electronics and Information Technology released three indigenous foundation models. EKA-1 comes from Soket Labs Technology and Research: an open, sparse hybrid language model with 26.3 billion parameters, trained on approximately 7.5 trillion tokens, spanning all 22 scheduled Indian languages alongside English, and built for maths, code, reasoning and agentic use. Soket's CEO, Abhishek Upperwal, said it was tailored to India's banking, financial services and cybersecurity landscape.

Dhīr 2B and Dhīr 5B come from ZenteiQ AiTech Innovations and are open source. They support English and 10 Indian languages, and the release describes them as secure, on-premise multilingual assistants for offline-capable, document-grounded automation, knowledge retrieval and citizen or enterprise services across defence, government, industrial and educational sectors, aimed particularly at low-connectivity or sensitive environments. ZenteiQ's CEO, Prof. Sashikumar, said they run across GPUs and TPUs and are light enough for on-device deployment on mobile phones with no cloud connection, keeping data on the device.

Both teams were selected under the IndiaAI Innovation Centre's Foundation Models pillar. The release credits subsidised national AI compute and support for ancillary research expenses with making the work possible, and MeitY Secretary S. Krishnan pointed at AIKosh as the route for researchers, students and the public to get at the models. Full details are in the PIB release of 3 November 2026.

Why it matters. Start with what the release does not contain: a single benchmark score. There is no MMLU number, no Indic-language eval, no comparison against any model you could otherwise download. That absence is the main thing to hold in mind, because a 26.3-billion-parameter model trained on 7.5 trillion tokens is a respectable size and tells you almost nothing about whether it reads a Marathi invoice better than what you are using now. The parameter count is an input, not a result. So treat the announcement as the arrival of options rather than the arrival of better options, and plan to measure the difference yourself.

What is genuinely new is the deployment shape. Two open-source models in the 2B and 5B range, explicitly built to run on-premise and offline, in Indian languages, is not something the frontier labs sell. Their products assume a network connection and a per-token bill.

What it means for a service business. There is one category of work where this changes the arithmetic immediately: documents you are not comfortable sending to an American API. Signed client contracts. Employee records. Patient intake forms in a clinic. Aadhaar-adjacent KYC paperwork. Up to now the honest answer on those has been either “send it anyway and hope the DPA covers you” or “keep doing it by hand”. A 2B model that runs on a laptop is a third answer.

The test to run this month is small and specific. Take 50 documents of one type that currently get read by a person, run them through Dhīr 2B on a machine in your office, and count how many fields come back correct. If the number is 40, you have a reviewed workflow worth building. If it is 15, you have saved yourself a project. Either way the download costs nothing and the compute is a laptop you already own, which is a materially different risk profile from signing up for a pilot. We would run this on a boring queue first at AcquihireTech, not a client-facing one, for the same reason you test a new courier on your own parcels.

2. Enterprises now spend $1.70 deploying AI for every $1 spent building it

Menlo Ventures published 2026: The State of Enterprise AI on 3 November 2026, drawing on a survey of nearly 400 US enterprise decision-makers responsible for more than $4 billion of AI spending. Enterprise AI deployment spending went from $33 billion to $177 billion in a single year, reaching $176.8 billion by September 2026, up 5.4x against 3.9x growth the year before. The line that reframes the market: for every $1 the industry spends building AI models, enterprises spend about $1.70 putting them to work.

Concentration tightened sharply. Anthropic, OpenAI and Google together took $122.7 billion, or 69% of deployment dollars, up from 50% a year earlier. Within enterprise LLM API spend, Anthropic moved from 40% to 55%, OpenAI from 27% to 31%, and Google fell from 21% to 9%. Open-source models account for 4% of spend despite 19% of usage. The application market is put at $77.3 billion, of which coding is $34.0 billion and chatbots $25.0 billion, and 55% of incremental AI spend came out of new budgets rather than existing ones.

On agents, the numbers are steeper than the narrative. Agents now generate 80% of all tokens on OpenRouter, roughly 7x human-driven usage, and agent traffic is up 37x since it overtook human traffic in February against 5x for human-driven usage. Among AI builders, 74% use agents to manage workflows and execute tasks, against 27% of startup deployments in 2025, and 32% already run several agents together. The supporting infrastructure moved with it: code sandbox adoption went from 11% to 40%, search from 5% to 30%, browser infrastructure from 4% to 14%. The full report is published by Menlo Ventures.

Why it matters. The sample is US enterprises and the report publishes no India-specific figures, so do not read the absolute numbers as a forecast for Delhi or Mumbai. The ratios travel better than the totals. And one of those ratios is quietly bad news for a particular kind of pitch: if the industry spends $1 building and customers spend $1.70 deploying, then most of the money in AI is being spent on integration, change management, review workflows and the plumbing around a model call. That is service work. It is also exactly the work that gets left out of a proposal that leads with the model's name.

The open-source figure deserves a second look too. Four per cent of spend against 19% of usage is not a contradiction; it is the price difference showing up in a chart. People run open models for the cheap, high-volume, repetitive half of the workload and pay the frontier labs for the half that has to be right.

What it means for a service business. Two concrete moves. First, if you quote AI projects, split the line items so the deployment work is visible: a client who thinks they are buying a model will query every hour you bill around it, and a client who can see that integration is the majority of the work will not. Second, go and check your own ratio. Add up what you paid in model tokens last quarter against what you paid in people's time wiring those calls into something usable. If the second number is not comfortably larger than the first, you are either unusually efficient or you have not finished the job.

3. Anthropic's text watermark reached three more Claude models

Anthropic's platform release notes dated 3 November 2026 record that text generated by Claude Mythos 5, Claude Opus 4.7 and Claude Sonnet 4.6 now carries Anthropic's text watermark on the listed platforms. That list covers the Claude API, Claude, Claude Code, Claude Cowork and Claude Tag, worldwide. This is the older-model half of a rollout Anthropic described in August: models launched in the EU on or after 2 August 2026 were marked at launch, and earlier ones are being added under an EU transition period. Anthropic's support documentation says the Amazon Bedrock rollout for these three was expected to complete by 10 November.

Mechanically, the watermark is Anthropic's implementation of Google DeepMind's SynthID-Text approach. It changes how the model picks among similarly good word choices so the choices follow a pattern checkable with a key. No tokens are added, no hidden characters are inserted, and it carries no information about the user or the organisation. Anthropic says it has seen no effect on quality, creativity or readability, and no meaningful impact on speed or cost.

There is a trap in the detail, and it is the kind that bites developers rather than writers. On Claude Sonnet 4.6, text generated for a request that sets temperature, top_p or top_k to a non-default value might not carry the watermark, and Anthropic's own documentation tells you to omit all three if you rely on it, including any a framework or client library sets on your behalf. Many client libraries set temperature by default.

Why it matters. Detection is not a thing you can do. The detection API is in private preview for eligible organisations: regulators, law enforcement, media, fact-checkers, researchers, educational organisations, EU civil society groups and enterprises with their own compliance obligations. So the practical state of affairs is asymmetric. Your output may be markable, the key to check it sits with the vendor and a short list of institutions, and you have no way to audit your own content. Anthropic is also clear about the limits: the mark does not apply where only one word is correct, such as exact facts and most code, light editing probably will not remove it while a full word-by-word rewrite will, and detection is less reliable on short passages.

Note what the watermark does and does not establish. It shows Claude was likely involved somewhere. It cannot separate Claude-written text from heavily edited text, cannot confirm human authorship, and says nothing about other AI systems. Anyone selling you AI-detection certainty on the back of this is overreading it.

What it means for a service business. If you sell content, assume from today that AI-assisted drafts are markable at source and write your client contracts accordingly, which mostly means not promising what you cannot verify. A clause saying “no AI was used” is now a claim a third party might one day be able to check, and you cannot. Replace it with a description of your actual process: what the model drafts, who edits, who signs off. That is both honest and defensible. If you build on the API, add a one-line check to your code review: does anything in the stack set temperature on a Sonnet 4.6 call? You may be dropping the watermark without meaning to, which cuts both ways depending on what you promised.

4. OpenAI shipped two image models that return editable SVG

OpenAI's API changelog dated 4 November 2026 introduced two models for generating and editing images: gpt-image-2.6-sunburst and gpt-image-2.6-flare. The guide positions Sunburst for detailed visual work and Flare for fast iteration. Three things are new beyond the models themselves. Output now includes SVG and animated GIF through the Image API. A complexity field accepts low, medium or high, defaults to high, and carries no separate fee; lower suits focused edits, higher helps with dense text and complex layouts at the cost of latency and a greater chance of changes you did not ask for. And markup-guided editing now extends to supported GPT Image 2.5 and 2.6 models.

Markup-guided editing is the one worth understanding properly. You post to /v1/images/edits with the original image first and a transparent PNG annotation layer as markup. The annotation layer has to match the first image's width and height after orientation and must contain transparent pixels. Markup applies to the first input image; additional images act as references. You cannot combine markup with mask or prompt_rewriter. Input images and annotation layers cap at 16,777,216 pixels each, the annotation file at 50 MB, with recommended sizes of 1024x1024, 1536x1024 and 1024x1536. The details are in OpenAI's image generation guide.

Why it matters. SVG output is the quiet one and the most useful. A raster image of a logo or an icon is a dead end: your designer redraws it. A vector comes back as paths somebody can open in Figma or Illustrator and actually change, which moves generated imagery from mood board to production asset. Animated GIF matters less than it sounds, because GIF is a poor format for anything with gradients, but it does remove a round trip for a simple loop.

The guide publishes no price for either model, so the cost per image is unknown until you check the pricing page, and nothing here should be planned on an assumed rate.

What it means for a service business. If you run social or ads for clients, the thing to test first is markup-guided editing, because it matches how feedback already arrives. A client circles something in a screenshot and writes “make this bit smaller”. That is now expressible as an API call rather than a conversation with a designer. The honest caveat is the setup cost: a transparent PNG layer matching the original's exact dimensions is not something a non-technical account manager produces by hand, so this is a build, not a feature you switch on. Start by checking how many revision rounds your last ten creative jobs actually took. If the answer is one or two, this solves a problem you do not have.

5. AWS gave AI agents root on a Linux desktop

On 3 November 2026 AWS announced that Amazon WorkSpaces for AI agents now supports Linux and Windows 11 cloud desktops in addition to Windows Server. The Linux option gives agents full root access to the operating system. Windows 11 lets them work inside enterprise applications. AWS frames it as choosing the operating system each agent's workflows and applications need, letting agents join the environments a workforce already uses without re-platforming, and running developer tasks such as builds, tests and scripts on Linux desktops. The new desktop types carry the same controls as the Windows Server ones: network isolation, IAM-based access and existing compliance boundaries. The announcement names no regions and no pricing.

Why it matters. Read the phrase “full root access to the operating system” twice. This is not a capability announcement so much as a governance one. The reason it is useful is the reason it is dangerous: an agent with root can install what it needs, which is exactly what makes a coding agent productive and exactly what you would refuse a new contractor on day one. AWS's answer is that the blast radius is a disposable cloud desktop with IAM around it, and that is a reasonable answer. It only works if somebody has actually scoped the IAM role rather than attaching something broad to get the demo running.

What it means for a service business. Most ₹10L+/month service businesses will not provision an agent desktop this quarter, and that is fine. The item still earns its place, because it tells you what to put in your vendor questionnaire. When an automation supplier proposes an agent that touches your systems, ask three questions: what identity does it run as, what can that identity reach, and what happens to the machine when the task ends. A supplier who answers “it runs as us” has told you something important. The useful default, whether you are on AWS or a laptop in the office, is that an agent gets its own identity, its own scoped permissions and a machine you can throw away, rather than a seat on a box that also holds your CRM.

The window, and the stories we cut

The window for this edition is 3 to 5 November 2026. Nothing dated 5 November could be opened at a primary source at the time of writing, which is expected: the US day had not finished when this went out. All five stories are dated 3 or 4 November, and they come from five different organisations, which is a healthier spread than the 2 November edition managed.

Four candidates were cut for failing verification rather than for being dull, and the pattern in them is worth naming. Several aggregators are currently presenting OpenAI's GPT-6.1 Terra launch as this week's news; it is real, but AWS's own release note confirms it reached general availability on Amazon Bedrock on 26 October, a week and a half outside this window. The same aggregators are recirculating the White House accord on frontier AI as current, and every outlet covering it dates the signing to the end of September. Salesforce's agreement to acquire the voice-AI firm Cresta is being reported with a $2.7 billion price tag, which traces back to The Information citing one unnamed source; no Salesforce announcement or filing that could be opened in this run discloses a price, so the number does not get repeated here. For contrast, the Salesforce press release exhibit on EDGAR for its separate Braze acquisition, filed 20 October 2026, states the terms plainly at about $5.66 billion in enterprise value and $47.50 per share in cash. That is what a confirmed deal price looks like. Reporting that Anthropic signed its first publisher licence with the Associated Press for election data rests on Adweek's two anonymous sources with neither company commenting on the record, which is not enough for a factual claim.

Three verified items were left out for being too narrow rather than unsound. GitHub shipped Copilot inside Xcode's built-in coding assistant on 4 November, in public preview and requiring the Xcode 27.3 beta on macOS Tahoe 26.6 or later. On 3 November GitHub's Copilot usage metrics API began reporting bring-your-own-key usage separately from Copilot-hosted usage, which is genuinely useful if you are trying to work out what your developers' AI costs actually are, with the caveat that local BYOK counts depend on client telemetry and are therefore a lower bound. And Google added Guided Learning to Classroom assignments on 4 November, which is a school story rather than a business one.

The short version. Download Dhīr 2B this week and run 50 real documents of one type through it on a machine in your office, then count the correct fields before you believe anything about it. Work out your own ratio of model spend to integration labour, and if integration is not the bigger number, the job is not finished. Take “no AI was used” out of your content contracts and put your actual process in instead. Check whether anything in your code sets temperature on a Sonnet 4.6 call. And before any supplier's agent touches your systems, make them tell you what identity it runs as and what that identity can reach.