Issue #25 | October 31, 2026 | Weekly Intelligence Briefing

Instinct, an AI agent that lives in your text messages, raised $1 billion in September at a $10 billion valuation, per TechCrunch.

This week TechCrunch listed Instinct among the agents sending orders to DoorDash. CEO Tony Xu put the count at "several hundred orders in total" from those third-party agents.

DoorDash's own SEC filing counts 993 million orders for the quarter. The agents are raising money faster than they're placing orders.

That gap runs through the week. Shopping agents can push your prices up, and Anthropic's IPO filing shows up to $567 billion in compute contracts behind the model you pay for.

The useful work shipped quieter. Google sent its newest model to security teams first, OpenAI priced a midrange model at half of its Sol tier, and a free security platform landed on GitHub.

Below: 5 signals, the Hype Check Spotlight, and the 1 tool worth your Monday.

Source: Bloomberg Tech

Bloomberg ran a segment on October 26 on AI agents already changing the price of goods. The detailed evidence sits in Groundwork Collaborative's August 4 testimony by Lindsay Owens on AI surveillance pricing.

Half of Walmart's app users use its Sparky assistant. Shoppers who use it spend about 35% more than those who don't, per a Walmart earnings call cited in the testimony.

Researchers found AI models favored the company over the user, steering people toward pricier products and even payday loans, the testimony says.

Pricing intermediaries surveyed by the FTC claimed 2 to 5% revenue growth from these practices. Target's app charges more while you sit in its store parking lots.

In the survey the testimony cites, 72% of shoppers would accept smaller discounts rather than hand over personal data for pricing. More than 40 bills in 24 states and cities target the practice this year.

An agent that knows your budget can price to it, and that agent answers to whoever sets the price.

→ If your team buys supplies through an AI assistant, compare 5 recent orders against a logged-out browser price this week.

→ If you sell online, check whether your platform offers personalized pricing, and read your state's pending bills before you switch it on.

→ Watch for assistant features that ask for your purchase history. That data is what the testimony says sets the price.

Read the full signal.

Source: AI Business

Anthropic made its S-1 filing public on Monday, October 26, and expects to list as soon as next month. The figures below come from the filing, as reported by The Wall Street Journal.

First-half 2026 revenue reached $16.3 billion, up from $1.1 billion a year earlier. The net loss for the half was $80.7 billion, and about $72 billion of it sits in accounting charges.

Anthropic spent close to $9 billion on compute and cloud infrastructure in the half, up from $2.3 billion. Its long-term infrastructure contracts total up to $567 billion.

About 20% of revenue came from 2 customers. The founders keep 50.1% of voting power, and a separate stock class elects a board majority through the Long-Term Benefit Trust.

A vendor carrying up to $567 billion in compute contracts has to recover that money from customers, and API pricing is where it shows up first.

→ Price your Claude usage at 2x today's rate in next year's budget, and check the workflow still pays.

→ Keep 1 workflow running on a second model provider, so a price change doesn't stop your work.

→ Kashyap Kompella, CEO of RPA2AI Research, told AI Business that pressure to recover infrastructure costs could change pricing or usage limits. Plan for both.

Read the full signal.

Source: Google DeepMind

Google announced Gemini 4 Carbon, its next frontier model, and is releasing it first to a subset of cyber defenders through its Fairwind Program. Developers, businesses and consumers get it later, with no date given.

Google reports 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploitation and 71.0% on CWE-bench for patching.

On Gray Swan's indirect prompt injection benchmark, Google reports a 0.0% attack success rate for Carbon.

The introductory price is $2 per million input tokens and $10 per million output, with cached input 95% off. After the introductory period, it's $4 and $20.

Google is testing guardrails on its newest frontier model before the public gets it, so the $2 price matters to you once access opens.

→ Don't plan a workflow around Carbon this quarter. Access sits with the Fairwind cohort.

→ If you pay a security firm, ask whether it's in the Fairwind Program and what Carbon has found for clients.

→ Budget at the $4 and $20 rates. The $2 price ends with the introductory period.

Read the full signal.

Source: simonwillison.net

OpenAI released GPT-6.1 Terra, a midrange API model, written up by Simon Willison on October 26. OpenAI's own pricing page lists it between Sol and Luna.

Terra costs $1 per million input tokens and $5 per million output, with cached input at $0.05, per OpenAI's pricing page on October 30.

GPT-6.1 Sol costs $2 and $10, so Terra runs at half of Sol on input and output. GPT-6 Luna sits at $0.10 and $0.50.

Batch pricing halves Terra again, to $0.50 per million input tokens and $2.50 per million output.

Terra halves the per-token price of Sol, and the saving counts once the work you move to it passes your own quality check.

→ Rerun 20 recent tasks from your current model on Terra and compare the outputs before you switch.

→ Send overnight jobs through Batch at $0.50 and $2.50.

→ Track output tokens. On Terra, output costs 5x input.

Read the full signal.

Source: AutoGPT Blog

AutoGPT compared 5 AI agent development services for customer service teams: Deloitte, IBM Consulting, Azumo, 10Clouds and Markovate. Its benchmark section sets vendor claims against independent data.

Vendor-published deflection rates cluster at 70 to 80%, per Digital Applied's 2026 data cited by AutoGPT. Decagon claims 80%, and Intercom's Fin reports 67% across 7,000+ customers.

The independent Zendesk enterprise median is 41.2%, with the top quartile at 58.7%. Gartner finds 14% of issues reach full self-service resolution.

Intercom's own pricing page lists Fin at $0.99 per outcome, checked October 30.

Plan your support budget at the 41.2% median, and treat any vendor quote above 58.7% as a claim to test.

→ Ask each vendor for deflection rates by ticket category from a past client, as AutoGPT advises.

→ Price the outcomes. At $0.99 each, 1,000 Fin outcomes cost $990.

→ Pilot on your top 5 ticket categories before you sign a build contract.

Read the full signal.

Source: TechCrunch

TechCrunch's October 24 roundup lists the agents you reach by text, from Instinct to Poke. DoorDash measured this category's orders in the hundreds.

Community adoption. Szn says on its website that it has served more than 7,500 businesses across 16 countries and completed over 10,000 jobs, per TechCrunch. Those are the company's own figures.

Pricing model. TechCrunch lists free tiers and monthly subscriptions across the category. We couldn't confirm a price on a vendor's own page this week, so no price appears here.

Benchmark data. TechCrunch's roundup carries no head-to-head test. The closest usage number is DoorDash's count: several hundred agent orders in a quarter of 993 million.

Expert sentiment. Investors are paying up. Instinct raised $350 million at a $2.5 billion valuation, then $1 billion at $10 billion in September.

Town raised a $55 million Series A led by Andreessen Horowitz and Forerunner. Cognition bought Poke's parent, The Interaction Company of California, in a deal valued in the low nine figures.

Release maturity. Trust features differ. Maverick asks for your permission before taking action, and Ollie is one of the first family-focused assistants to reach SOC 2 compliance, per TechCrunch.

The verdict: funding is running ahead of usage in this category.

Try an agent for personal errands, and keep business logins out of it until the security record catches up.

Read the full signal.

Source: Medium

ARGUS is an open source security operations platform released under the MIT license. It's self-hosted and multi-tenant, so the cost is your own server time.

It runs 7 specialist AI agents with read-only tools, plus an action broker with 27 typed actions across services including AWS, Azure, Okta, Cloudflare, CrowdStrike and SentinelOne.

Autonomous response ships off by default. Switched on, its playbooks start in shadow mode, use reversible actions, stay under hourly caps and roll back unless you keep the change.

The weak spots: the GitHub repository showed 116 stars and 53 commits on October 30, so it's a young project. Setup needs Python 3.11 or 3.12, or Docker, and a person who can run a server.

Worth a weekend pilot in shadow mode if you already pay for CrowdStrike or SentinelOne and have a technical person. Leave autonomous response off until you've read its logs.

Read the full signal.

Agents that want your money, your login and your ad budget.

DoorDash counted several hundred AI agent orders in a quarter of 993 million. CEO Tony Xu gave the agent count on the earnings call. DoorDash's 8-K puts total orders up 28% to 993 million and Marketplace GOV up 36% to $33.9 billion. Reported by TechCrunch, with totals from DoorDash's SEC filing.

OpenAI's Dots assistant starts at $100 a month and wants your inbox. Dots works in the background for ChatGPT Pro subscribers and above, connected to your calendar, email and computer. TechCrunch's tester found it flags calendar conflicts, though sometimes late. Reported by TechCrunch.

Google is taking AI Mode ads global as Search growth slows to 12%. Google Search revenue grew 12% in the third quarter, down from 17% and 19% in the 2 quarters before. Ads in AI Mode now expand to international markets, with multi-turn ads in the US. Reported by TechCrunch.

Bank of America says you own whatever your AI agent does with your login. The bank emailed customers not to share banking credentials with AI tools or agents. Its email reads: "keep in mind that you’re responsible for activity resulting from that access." Reported by TechCrunch.

The full week's signals, detailed breakdowns, and action items are on the site. If this issue earned its place in your inbox, forward it to whoever signs off on your AI budget.

This issue went out to subscribers Saturday. If you want next week's before it hits the web, subscribe at metadatamarketer.com/subscribe

Test. Cut. Share.

Moe Sbaiti, The AI Profit Wire https://metadatamarketer.com

Disclosure: some tools referenced in this newsletter are affiliate partners. Full disclosure and analysis at metadatamarketer.com.