AI news, 28 October 2026: a default moved, and four prices fell.

Google changed a default on 26 October: new paid-tier Gemini API projects now store every request they send, where they used not to. Anthropic cut 40% off Claude for US government buyers and extended that price to states, cities and tribal governments. It also published the first hard numbers from six drug-development partnerships, and the most useful one in the whole write-up is about reading documents, not about biology. ServiceNow wired its workflows into Gemini Enterprise over two open protocols. And Oracle said it will sell Google's TPUs, which no cloud outside Google has done before. One of these five changes something you ship this week. The other four change what you should be asking vendors.

Every date and figure below was read at the source linked in its own section during this run. Where a claim could not be opened at a primary source, it was cut rather than softened, and the cuts are listed near the end. AcquihireTech writes this roundup for Indian service businesses doing ₹10L+ a month, so each story is put through the same filter: what does this change about how you run the business on Monday?

1. New Gemini API projects now log your prompts by default

On 26 October 2026, Google flipped a default in the Gemini API. Requests to generateContent, streamGenerateContent and batchGenerateContent are now stored by default (store=true) for projects created on or after that date, which brings them in line with the Interactions API, where storage was already on. Projects created earlier keep whatever setting they have. The change is recorded in the Gemini API changelog, and the mechanics are set out on Google's Logs and datasets page, which states plainly that projects created on or after 26 October 2026 store requests by default and older ones do not unless logging was switched on.

The retention numbers are on the same page. Logs expire after 55 days by default. You can shorten that window to 7, 14 or 28 days, and 55 is the ceiling. Logs you deliberately save into a dataset do not expire at all. Storage is a paid-tier feature, so a free-tier key is not affected. There are two ways out: send store as false on an individual request, which needs a recent SDK, or turn the feature off for the whole project in the settings panel of the AI Studio logs dashboard, where generateContent and Interactions have separate toggles.

Why it matters. A default is a decision somebody else made for you, and this one moved in the direction of keeping more data for longer. Nothing was taken away. No existing project changed. But the next project you create behaves differently from the last one, and nobody is going to send you an email about it.

What it means for a service business. Think about what your prompts actually contain. In an Indian service business running AI anywhere near the enquiry pipeline, a single prompt routinely carries a client's full name, their phone number, the address of the site you are quoting, the number you quoted, and sometimes a line about why the last vendor was fired. That is the payload. If you spin up a fresh paid-tier project this week for a WhatsApp responder or a quote drafter, that payload is now retained for up to 55 days unless you say otherwise. Three things to do, and all three are ten-minute jobs. Check the creation date of every Gemini project you own, because the behaviour splits on 26 October. Set the retention window to 7 days on anything that touches client data, or switch logging off entirely. Then look at what you have told clients in writing about where their data goes, because if your proposal says data is not retained, that sentence is now doing work your configuration has to back up.

The wider point is duller and more important. Model vendors ship defaults, and defaults drift. The only durable protection is a habit: a dated note of what each AI vendor you use retains, for how long, and which setting controls it. Most of the teams AcquihireTech audits have no such note, and the reason is never carelessness. It is that nobody owns the question.

2. Anthropic took 40% off Claude's government price, and extended it to states and cities

Anthropic published an expansion of its partnership with the US General Services Administration on 27 October 2026. The agreement gives government partners a 40% discount on Claude Enterprise, Claude for Government and the Claude API. Anthropic's post says the discount applies across every product surface, including direct purchases and, in the coming months, purchases through cloud providers. There is no minimum spend commitment attached and agencies pay only for what they use. Buying runs through the GSA Multiple Award Schedule. Anthropic estimates the discounts will save the government a billion dollars, and it extends the same pricing beyond federal agencies to state, local, tribal and territorial governments. The post cites Utah piloting Claude Code across 10 state agencies, and promises a dedicated implementation-support programme in the coming weeks.

Why it matters. Two things are visible here that are usually hidden. The first is that list price is a starting position. A 40% cut with no minimum commitment is not a volume deal in the normal sense, because volume deals are what minimum commitments are for. The second is who the discount is aimed at: a buyer class that moves slowly, procures through a schedule, and once it standardises on something does not switch for years. Anthropic is buying installed base, and it is paying for it out of margin.

What it means for a service business. You will not get the GSA rate. What you can take from this is narrower and still worth money. First, if an AI vendor quotes you list price and you have any volume at all, ask what the committed-use rate is, because the published number is evidently negotiable for buyers with leverage. Second, and this is the one that costs real money to get wrong, do not architect your systems around one vendor's current pricing. Per-unit model costs have been falling all year, and vendors are now discounting hard to lock in buyers who are slow to move. The right design answer is boring: keep the prompt, the retrieval and the business logic in your own code, and treat the model as a swappable component behind one interface. If moving from one provider to another is a week of work rather than a quarter, every price cut in the market is yours to collect.

3. Anthropic's drug-development results, and the one finding that transfers to your business

Also on 27 October, Anthropic published early results from deep research partnerships with six drug developers, run over several months. The post is careful that the work is early and mostly preclinical, and the laboratory results are genuinely striking. Generate Biomedicines used Claude Code to propose shorter genetic designs for a CAR T-cell therapy; one design was about 70% shorter than the standard version, let immune cells make roughly 10 times as many receptors, and stayed active longer. Divergence Labs generated roughly 35,000 antibody designs in about a month, largely autonomously, and found three that bound IgE and 20 that could dislodge it. BigHat Biosciences ran about 1,010 variants through an automated lab in roughly 10 weeks; the best bound more than 10 times more tightly, one lead stayed in solution at a temperature 18°C higher, and another clumped more than 100 times more slowly. Anthropic also acquired Mithril Medicine, a one-person company working on rare forms of muscular dystrophy, and says a candidate went from idea to promising mouse results in under 18 months.

None of that transfers to a renovation firm in Gurugram. One result does. Enveda used Claude on 1,061 pages of manufacturing documentation and regulatory filings. Claude caught all 11 issues the human reviewers had found, and flagged seven more they had missed. Review time went from more than 25 hours to under three, and a regulatory document that took about three months to write by hand was drafted in four hours. Anthropic's own figure for the review-time change is 89%.

Why it matters. Read that result again and the headline number is the wrong one to be impressed by. Twenty-five hours becoming three is an efficiency story, and efficiency stories are easy to sell and easy to overstate. The seven issues the humans missed are a different kind of claim. They say the previous process had a defect rate nobody had measured, in a regulated domain, with expensive reviewers. That is not a speed finding. That is a quality finding, and quality findings are the ones that survive contact with an actual business.

What it means for a service business. The shape of the Enveda task is everywhere in Indian service work: a thick document set, a checklist of known failure modes, and a reviewer who is senior, expensive and bored by page 40. A tender response. A vendor contract before signature. A GST filing pack. An insurance claim bundle. A set of site measurements against a bill of quantities. If you want to try this, copy the structure rather than the technology choice. Take a document set you have already reviewed by hand and know the answers to. Write down the failure modes you were looking for. Run the model over it and compare three things: what it found that you found, what it missed, and what it found that you had not. The third column is the one that tells you whether this is worth building, and it is the column almost nobody measures, because it requires admitting the old process let things through.

4. ServiceNow plugged its workflows into Gemini Enterprise over MCP and A2A

ServiceNow and Google Cloud announced on 27 October that ServiceNow's industry workflows, exposed through its Action Fabric, now connect to Gemini Enterprise, with Gemini Enterprise acting as the interface a person actually types into. The integrations named are retail store operations (Retail Operations plus Field Service Management, so a reported equipment fault can get a technician dispatched), customer relationship management (deals, quotes, account summaries, follow-ups), and IT service and operations (ITSM, ITOM and the CMDB, for incident triage, change-risk assessment and alert investigation). Retail, CRM and IT are generally available to mutual customers now. The telecommunications integration is expected to be generally available in early 2027. The connectors will ship as plugins in Google Cloud Marketplace. The release states the integration runs on Model Context Protocol servers and Agent2Agent connectivity, and quotes ServiceNow's John Aisien saying “Conversational AI is only as useful as the action it drives”.

Why it matters. The protocols are the story, not the partnership. Two large enterprise vendors with every commercial reason to build a proprietary bridge instead shipped on MCP and A2A, which are open. That is the second time in a month this roundup has had cause to mention MCP, and the direction is consistent: the plumbing for connecting an AI system to a business system is standardising, which means it is getting cheaper. Aisien's line is also the most honest sentence in either company's release. A chatbot that cannot complete the task is a support cost wearing a product's clothes.

What it means for a service business. You are not buying ServiceNow. You are, however, buying or already running a CRM, a ticketing tool, an accounting system and a WhatsApp inbox, and the question to put to each vendor has changed. It used to be “do you have an API”. Ask instead whether they expose an MCP server, or plan to, and when. The difference is practical: an API means you write and maintain the integration, while an MCP server means an agent can discover and use the tool without you hand-coding each call. One of those is a project. The other is a configuration. If your main systems answer no with no date, that is a fact worth knowing before your next renewal, not after.

5. Oracle will sell Google's TPUs, starting in 2027

At Oracle AI World in Las Vegas on 27 October, Oracle and Google announced that Google's Tensor Processing Units will be offered through Oracle Cloud Infrastructure. Oracle describes this as the first time Google TPUs will be available on a hyperscale cloud platform outside Google, with OCI supplying the power, cooling and datacentre networking. The announcement covers Google's eighth-generation TPUs in two flavours: TPU 8t for training, which Google says delivers nearly three times the compute performance per pod of the previous generation, and TPU 8i for inference, aimed at the repetitive work of many specialised agents. Availability is calendar year 2027, ramping over a multi-year period. The release gives no chip counts, cluster sizes or deployment volumes. Mahesh Thiagarajan, who runs OCI, frames it as giving customers “multiple leading AI compute architectures in one cloud”.

Why it matters. Inference supply sets the floor under what agents cost, and until now the only way to rent a TPU was from Google. A second hyperscaler selling the same silicon is competitive pressure on GPU inference pricing, which is the line item that decides whether an always-on agent is affordable. The honest caveat is the date. This is 2027 hardware with a multi-year ramp, so it changes nothing in your budget this quarter, and the per-pod performance figure is Google's own.

What it means for a service business. One concrete implication: be careful what you sign. If a vendor offers you a two or three-year commitment priced off today's inference costs, you are betting against a market where supply is widening and published per-token prices have been falling all year, which this roundup documented on 26 October from the vendors' own pricing pages. Shorter terms cost slightly more per month and are worth it. The second implication is about architecture again, and it is the same answer as story two: the cheaper compute gets, the more valuable it is to be able to move.

The window, and the stories we cut

The window for this edition is 26 to 28 October. All five stories were read at a primary source during this run: two posts on Anthropic's own newsroom and research pages, two press releases on Google Cloud's press corner, and Google's own developer changelog and logging documentation for story one. No story here rests on secondary reporting, and no figure was taken from an aggregator.

Six candidates were cut, and one of them was the biggest headline of the day. Google DeepMind published Genie 4 on 27 October, a world model that generates interactive simulations, natively generates audio for the first time in this line, and can be steered mid-simulation in plain language. It verifies cleanly. It is also a research preview for trusted testers, with a waitlist, coming to Google AI Ultra subscribers and early-access Gemini API partners, and there is no decision an operator can make differently on Monday because of it. So it is named here rather than given a section, which is the honest placement. Apple is reported to be holding an event in Los Angeles today at which M6 Macs are expected, and one aggregator carried a claim about time-to-first-token for local language models against the M1; Apple's newsroom did not render for this pipeline and the claim was not read at an Apple source, so it is out. A reported date of 4 November for Moonshot's Kimi K3.1 Pro open weights was not read on a Moonshot channel. Oracle was also reported at the same event to be putting ChatGPT and Codex inside its applications, which is plausible and was not read at an Oracle source in this run, because oracle.com returns a 403 to this pipeline. A claim that an Alibaba Qwen model attempted to send a file to a Beijing endpoint is contested in the same reporting that carries it. And a political item about renaming “artificial intelligence” to “super intelligence” appears only in a live blog with no confirmable date.

The short version of this edition: check the creation date on every Gemini project you own and set its retention window deliberately. Write down, dated, what each AI vendor you use retains and which setting controls it. Ask your next AI vendor for the committed-use rate instead of accepting list price. Pick one document set you have already reviewed by hand and measure what a model finds that you missed. And ask your CRM and ticketing vendors whether they expose an MCP server, and when.