AI Marketing News, 28 Sep 2026: GPT-6 Sol at Half Price, Claude Opus 5.5, 93% of GTM Teams Run Agents
OpenAI halves GPT-6 Sol pricing, Anthropic ships Opus 5.5 and Claude Marketplace, ChatGPT Ads hits $1B, and a third of GTM teams can't count their agents.
LeanData: 93% of GTM teams run AI agents, and nearly a third can’t say how many touch their records. LeanData’s 2026 State of AI Go-to-Market Readiness Report, covered by MarTech on September 25, surveyed 157 B2B practitioners in May 2026, 36% of them in RevOps. 93% have deployed at least one AI agent. Nearly a third could not say how many agents act on their CRM records, and 30% have found agent actions with no audit trail. 27% have seen more than one tool or agent contact the same prospect, and 17% have seen a marketing sequence fire while a rep was working the deal. 19% say nobody owns GTM AI strategy, and RevOps owns it at only 18% of organisations.
Importance for marketers: The double-touch and the sequence that fires mid-deal are routing failures, and they happen because each agent writes to the CRM with its own logic and no shared lock on the record. Put every agent that can contact a prospect behind one suppression check, an open-opportunity flag and a last-touch timestamp that HubSpot or Salesforce enforces before any send, then give RevOps ownership of the agent inventory the same way it owns field permissions. The logging and spend controls that make that inventory auditable are in Token budgets and audit logs for GTM agents.
OpenAI cuts API prices 50% with GPT-6 Sol and Luna. Released September 22, GPT-6 Sol costs $2 per million input tokens and $10 per million output, and Luna costs $0.10 and $0.50. OpenAI says Sol scores 33.2% on Zapier’s AutomationBench, which tests sales, marketing and operations workflows across 47 tools, at $0.27 per task, and claims that beats Claude Opus 5 at 9% of its cost. OpenAI also says Sol makes about half as many factual errors as its predecessor. Cached reads are now discounted 90%. Both models are live in ChatGPT Work and Codex on paid plans, and Free and Go users get Luna in the desktop app.
Importance for marketers: At Luna’s price, classifying a 50,000-row Clay table by ICP fit or persona costs a few dollars, so the case for running a cheaper model on the bulk step and a frontier model only on the final judgement is now a line-item argument. Split every enrichment and scoring workflow into those two stages and set the model per step in n8n or the Agents API. With the 90% cache discount, the long system prompt that holds your ICP definition and messaging rules costs almost nothing after the first call, so put more context in it.
Anthropic releases Claude Opus 5.5 at $4 and $20 per million tokens. Opus 5.5 shipped September 22 as the first model in the Claude 5.5 family. Anthropic says it matches Claude Fable 5.1 on most work and costs 40% less than Opus 5 on typical workloads. List pricing drops from $5 and $25 to $4 input and $20 output, cache reads fall 60% to $0.20, and a fast mode runs up to 2.5x quicker at $8 and $40. Output is over 30% faster than Opus 5. On AutomationBench it scores 40.0%, just behind GPT-6 Astra at 41.4%. It is available on the Claude Platform, AWS, Google Cloud and Azure, with Sonnet 5.5 and Haiku 5.5 due in the coming weeks.
Importance for marketers: Read the two launches together. On the same business-workflow benchmark, Opus 5.5 finishes more tasks than Sol, and Sol costs a fraction as much per task. That leaves you with a routing decision per workflow, and neither vendor wins across the board. Account research, deal summaries and anything a rep will forward to a buyer belong on the stronger model. List hygiene and first-pass classification belong on the cheap one. Re-run your own 20-case test set on both before you move production traffic, because each vendor’s launch post picks the comparison that flatters its own model.
Anthropic launches Claude Marketplace and opens its plugin directory to developers. Announced September 23, Claude Marketplace brings plugins, connectors, Claude-powered agents and service partners into one place, listing more than 2,000 connectors and plugins including Salesforce, Atlassian, Notion, Google and Microsoft. Enterprise customers can spend part of their committed Anthropic spend on partner software such as Snowflake, Harvey, Cursor and CrowdStrike. On September 25 Anthropic opened plugin submissions to paid-plan developers, with automatic validation, safety scanning and usage analytics. Claude now supports the MCP 2.0 spec plus MCP Apps, which render interactive UI inside a chat, and Enterprise Managed Auth, which provisions OAuth for enterprise users without a per-user login.
Importance for marketers: Enterprise Managed Auth removes the blocker that kept most sales teams off connectors: IT can switch on the CRM, intent and enrichment plugins for every seat at once, so the whole sales team can work from the assistant, not just a few power users. G2 and Zig.ai shipped MCP access to intent data and sequencing inside Claude and ChatGPT the same week. Decide now which system of record each plugin can write to, because a rep who can queue a sequence from chat will bypass the approval step you built in the sales engagement tool.
ChatGPT Ads reaches a $1B annualised run rate and expands to seven Asian markets. OpenAI said on September 23 that ChatGPT Ads hit a $1 billion annualised run rate in under 200 days with tens of thousands of advertisers. The same day it began rolling out ads in Indonesia, Malaysia, the Philippines, Singapore, Thailand, Vietnam and Taiwan, taking availability past 60 countries. Advertisers buy through OpenAI’s Ads Solutions team, through agencies including dentsu, Havas, Omnicom Media, Publicis and WPP, or self-serve in Ads Manager. Ads appear only to Free and Go users, and Plus, Pro and Enterprise remain ad-free. OpenAI says new formats, optimisation tools and measurement arrive over the next few months.
Importance for marketers: Anyone using ChatGPT on a company Enterprise licence never sees an ad, so the larger the target account, the more of its buying committee sits outside your reach. Aim ChatGPT Ads at mid-market ICP tiers and at the practitioners who research on personal Free accounts, and keep enterprise accounts on LinkedIn and ABM display. APAC teams now get the channel on the same terms as the US. Run a one-market test through HubSpot’s ChatGPT Ads connection before 2027 budgets lock, so you have pipeline data rather than a run-rate headline to argue from.
OpenAI tests a year-long ChatGPT cookie that links chats to advertiser site conversions. MarTech reported on September 23 that a researcher known as Buchodi found a ChatGPT cookie named __obi with a lifespan of up to a year. The same identifier appears in data sent back to OpenAI from advertiser sites running OpenAI’s ad pixel, alongside page and conversion data. OpenAI classifies it as an analytics cookie, and it could reportedly be set even when a user had declined marketing cookies. The mechanism would let OpenAI tie a ChatGPT session to a later purchase or form fill on the advertiser’s site.
Importance for marketers: If you have installed OpenAI’s pixel, your consent banner is now part of this story. A pixel that sends a cross-site identifier after a user declined marketing cookies puts the compliance exposure on your site, whatever category OpenAI assigns the cookie. Gate the pixel behind the marketing consent category, and send conversions server-side from HubSpot or Salesforce for users who opt in. That route gives you a stage-level match the pixel never will.
Field experiment: forcing searches into Google AI Mode cut external clicks by 18.8 points. A randomised study of 1,100 US Chrome users, covered by Search Engine Land on September 23, redirected one group’s Google searches into AI Mode for seven days. The share of searches leading to an outside website fell 18.8 percentage points against normal Google use, and participants ran 0.92 fewer Google sessions a day. Clicks to Reddit fell 21.2 points and to news sites 12.5 points. Satisfaction, trust and sense of control all dropped, and use of Bing, DuckDuckGo or Yahoo rose 11.2 points. Before the test AI Mode accounted for 0.6% of participants’ searches, so the result shows what happens when AI Mode becomes the default rather than an opt-in.
Importance for marketers: Google has not made AI Mode the default, but this is the first controlled estimate of the traffic you lose if it does: close to a fifth of external clicks, with the heaviest searchers losing the most. Use the 18.8-point drop as the downside case in your 2027 organic forecast, and move content effort toward the pages AI answers retrieve from: comparisons, pricing and integration docs. The gap between answers seen and visits recorded is covered in The attribution gap nobody’s measuring.
Microsoft Ads API adds LinkedIn segment targeting and a search term to landing page report. Search Engine Land reported on September 25 that the Microsoft Advertising API now supports LinkedIn profile segment targeting. The new Search Term Landing Page Report puts the query, the headline served, the final URL and performance metrics in one row, so you can see where intent, ad message and landing page diverge. Performance Max campaigns gain campaign-level device exclusions for computers, smartphones or tablets. Microsoft has not given specific availability dates beyond its API release notes.
Importance for marketers: Programmatic access to LinkedIn segments means you can sync company and job-function targeting from your ICP tiers into Bing, Copilot and MSN inventory without touching the UI. Pull the Tier 1 account list and target titles from Clay into a scheduled n8n job that updates Microsoft Ads audiences weekly. Use the landing page report to find AI-matched queries that land on the wrong page, which matters most in low-volume B2B accounts where Smart Bidding has too few conversions to correct the mismatch itself.
Google Ads adds IP address and timestamp matching to Customer Match. Documented on September 25, advertisers can now upload a user’s IP address and interaction timestamp alongside existing Customer Match identifiers such as email, phone and mailing address. Unlike those identifiers, IP and timestamp must be uploaded unhashed. The file format now takes eight columns with required English headers. The new options are unavailable for users in the European Economic Area, the UK and Switzerland, so advertisers must exclude those users and cover the sharing in their disclosures and consent. Google has also changed its help pages to describe customer data as information advertisers “collected” rather than “shared.”
Importance for marketers: B2B lists match badly on business email, and a form fill with an IP and timestamp gives Google a second route to the same person. Add both fields to the lead object at capture, strip EEA, UK and Swiss records in the export step, and rebuild your retargeting and exclusion lists from the enriched file. An unhashed IP is personal data under most US state privacy laws, so update your privacy notice before the first upload.