Three developments this week make it easier to change AI suppliers, and two pieces explain why that option matters.
- Claude in Google Workspace. Anthropic's assistant now edits Docs, Sheets and Slides in place, so decide who gets it before staff install it themselves.
- Mistral Large 4. A second serious open-weight model from outside China arrived in two days, which strengthens your hand in model negotiations.
- Personal Agent Protocol. Meta, Sierra, Stripe and Shopify proposed an open standard that lets a business set what a customer's agent may do.
- Provider switching. Ramp data shows a record 8% of businesses changed AI model providers in September, and two of the largest buyers are reportedly cutting back on Claude.
- The same stack. Investor Gil Dibner argues AI startups now share one architecture, so advantage has to come from knowing the customer.
- Inference economics. Tomasz Tunguz projects inference spending will pass the database market, which makes every AI-enabled product a reseller of someone else's compute.
- Every's token audit. Every cut its agent's token costs by more than 80% by removing waste, after a model-routing scheme made costs worse.
The common thread: keep your options open, and know what your AI actually costs.
Developments
Technology
Claude Moves Into Google Docs, Sheets and Slides (October 6, 2026)
Anthropic released a Claude add-on for Google Workspace in public beta on October 6, available on all paid Claude plans. Claude reads the open file and edits it from a sidebar: text and headings in Docs, formulas, pivot tables and charts in Sheets, and slides built from a deck's own layouts. By default each change appears as an approval card; a second mode applies edits without stopping. Access follows existing Google sharing permissions, and admins can deploy by domain or group.
So what
The assistant now works where documents already live, which removes the copy-and-paste step that kept many teams from using it on real files. If your company runs on Google Workspace, decide who gets the add-on before employees install it themselves. Keep the approval mode as the default and pilot it on one finance or board-reporting workflow. It is a beta, so check its work on anything that leaves the building.
Confidence: Established fact. Anthropic's own announcement; reliability in production has not been independently tested.
Go deeper Anthropic · Superhuman AI
CompetitionDeveloping
Mistral Previews a One-Trillion-Parameter Open-Weight Model (October 6, 2026)
Mistral put Large 4 into public preview on October 6 and says it will release the weights by the end of the month. The model has one trillion parameters, was trained on 3,800 Nvidia Grace Blackwell GPUs in Mistral's European data centers, and is priced at $1.36 per million input tokens and $4.18 per million output tokens. It follows Reflection AI's Beam preview a day earlier. Mistral's chief scientist told The Deep View the pitch is control: a model you hold cannot be withdrawn by its vendor.
So what
Credible open-weight options outside China are arriving quickly, and that strengthens your hand in any model negotiation. You do not need to self-host to benefit. Ask your engineering lead which workloads could run on an open model at acceptable quality, and bring that list to your next model contract renewal. Wait for independent tests before moving anything customer-facing.
Hype check
Every performance figure comes from Mistral, including its claim to lead all non-Chinese open models. On Mistral's own coding chart, as reported by The Rundown AI, China's Kimi K3 still ranks first.
Confidence: Watch this. The launch and pricing are confirmed; the performance claims have not been independently verified.
Go deeper Mistral · The Deep View
Customer BehaviorNew Angle
Meta, Sierra, Stripe and Shopify Back an Open Standard for Customer-Owned Agents (October 6, 2026)
Meta and Sierra announced the Personal Agent Protocol on October 6, an open standard for how a consumer's AI agent deals with a business. Sierra's post names Genesys, Rocket, Shopify, Stripe and Walmart as partners and says Instinct will join. The customer decides whether an agent gets read-only or write access to an account; the business sees when an agent is acting and sets what it may do. It builds on OAuth and MCP, and a v0.1 specification is due later this month.
So what
Agents acting for your customers are coming whether you plan for them or not, and this is an early attempt to let the business set the terms. There is no specification yet, so there is nothing to build today. Have your product lead read the announcement and list which account actions you would allow an agent to take. Note that the partner list does not include OpenAI, Google or Amazon.
Confidence: Established fact. The announcement is confirmed; adoption beyond the named partners is an open question.
Go deeper Sierra · The Neuron
Customer Behavior
Ramp: A Record 8% of Businesses Switched AI Model Providers in September (October 6, 2026)
Ramp's spending data shows the monthly share of firms switching AI model providers, averaged over three months, reached a record 8% in September. Ramp's Ryan Stevens writes that switching had declined for a year until OpenAI's June price cuts and model launches. Separately, The Information reported on October 5 that Microsoft cut a planned $1 billion of internal Anthropic spending by more than a third and that Meta's Claude Code users fell from about 60,000 to about 30,000.
So what
Buyers are treating models as replaceable, and the largest ones are acting on it. That favors anyone who kept their options open. Check how hard it would be to move your highest-volume AI workload to a second provider, and avoid multi-year commitments that assume one vendor stays ahead. If you sell AI features, expect your customers to ask the same question about you.
Confidence: Informed perspective. Ramp does not define switching or describe its sample, and the Microsoft and Meta figures rest on one report that neither company has confirmed.
Go deeper Ramp · The Deep View
Perspectives
Competition
Gil Dibner: Every AI Startup Is Building the Same Stack, So Stop Pitching It (October 6, 2026)
Angular Ventures' Gil Dibner argues that AI startups have converged on the same components: agent loops, connectors, memory, sandboxes and web search. He compares it to 2005, when every web company needed a database and a login page. Pitch decks now read alike too, since founders draft them with the same tools. His conclusion is that the stack is table stakes and that advantage sits in domain expertise, customer understanding and a clear value proposition.
So what
This is a useful test for your own positioning and for any AI vendor or investment you evaluate. If the pitch leads with architecture, ask what the team knows about the customer that competitors do not. Investors should weight customer references and team interviews above questions about memory systems or model choice. Review your own sales deck with the same eye this week.
Confidence: Informed perspective. One investor's reading of his own deal flow.
Go deeper Angular Ventures · TLDR Founders
BusinessNew Angle
Tomasz Tunguz Says Inference Is Becoming Software's Largest Cost Line (October 5, 2026)
Tunguz estimates companies spent about $25 billion running AI models in 2025 and will spend roughly $130 billion this year and $350 billion in 2027, nearly double the database market. He argues this turns every application into an inference reseller. Usage charges could exceed seat fees, the classic 72% software gross margin may erode, and customers who bring their own model keys leave vendors with high margins on much smaller contracts.
So what
If you sell software with AI inside, your gross margin now depends on a supplier's price list. Ask your CFO to report inference as its own cost line and to model three cases: prices fall, prices hold, and usage doubles. Investors should ask portfolio companies for gross margin with and without inference. Treat the forecasts as direction, not precision.
Confidence: Informed perspective. The market-size figures are one analyst's projections.
Go deeper Tomasz Tunguz · TLDR Founders
