As the agent platform race converges, it is time to go vendor-neutral | Software Talks AI
September 2026 · Issue #4 Monthly SaaS executive newsletter

As the agent platform race converges, it is time to go vendor-neutral

Welcome to another edition of Software Talks AI, Monterro's newsletter on AI for B2B SaaS leaders. We're back after a summer break, and what a summer it has been.

This month

It's proven to be a busy time for AI labs, as essentially all of them released new AI models. The agenda this time includes OpenAI's answer to Claude Cowork, big open-source releases, agents escaping sandboxes, and a discussion on vendor-neutral set-ups.

Back in May, we wrote that the agent platform race keeps converging. Well, over the summer it converged completely. OpenAI has served Cowork-like capabilities in the form of its Codex product for some time, but with the launch of ChatGPT Work this summer it is now competing head-on with Anthropic.

At the same time, the market itself has moved: Anthropic has passed OpenAI on quarterly revenue, open-weight models have proven to keep up with frontier models, and the gap between the frontrunners and regular users has continued to expand.

Top stories

What moved in AI this month, and what it means for B2B software leaders.

Story 01

Anthropic passes OpenAI on revenue, files for a record IPO

Anthropic's Q2 2026 revenue came in above $11.5 billion, 14 times higher than a year earlier, as the lab celebrates a profitable quarter. In August, its quarterly revenue passed OpenAI's, and its run rate passed $65 billion.

An IPO expected to match or exceed SpaceX's record $75 billion offering is being prepared, with shareholders expecting an October listing.

Read the full story →
Story 02

Summer of open-weights

This summer saw the release of some remarkably capable open-weight models: Moonshot's Kimi K3, Z.ai's GLM-5.2 and later GLM-5.3, and Alibaba's Qwen 3.8. The three of these are massive releases for the open-source community. Kimi K3 is the largest open-weight model ever released, as it performs at a level close to the best models by OpenAI and Anthropic. All three of these open-weight models are made by Chinese AI labs.

These models mean that frontier-adjacent capability now exists at price points and hosting terms that you control. Don't forget: in June this year, the US government decided to pull the Fable 5 model from Anthropic, and while the model has since returned, this displayed how US authorities can pull models at will. Open-weights enable not just compliant AI use, but also provide full control and stability over AI usage.

Read the full story →
Story 03

The gap between AI-capable firms tripled in five months

OpenAI published enterprise usage data drawn from more than ten million enterprise messages through its services. The top 10% of firms now produce 8.3 times the AI output per active user of the median firm, up from 2.6 times in January.

What separates them is not model access, it is plumbing: at frontier firms, 21% of weekly active users use plugins and 19% use skills, versus 9% and 3% at typical firms. Non-engineering agent use is growing fastest, up 108 times in legal and 41 times in sales year over year.

Read the full story →
Story 04

Agents escaped their sandboxes

Several major AI labs disclosed agent containment incidents this summer: Anthropic's models gained unauthorized access to three organizations' systems during security evaluations, OpenAI found its agents had compromised credentials at four companies, and the UK AI Security Institute logged 19 unauthorized actions across 122 test runs.

Read the full story →

Key metrics

Numbers worth keeping.

Growth +7,851%

Year-over-year growth in agentic AI traffic

Ecosystem 100,000+

Publicly listed MCP servers, allowing AI to communicate with systems

Ranking #4

Open-weight model Kimi K3 ranks fourth on the Artificial Analysis Intelligence Index v4.1, just a few points behind OpenAI and Anthropic models

Our take

ChatGPT Work vs Claude Cowork, and being vendor-neutral

This summer, OpenAI released ChatGPT Work as its Claude Cowork competitor. Many AI labs have now released Cowork competitors, as the competition includes the likes of Mistral Work, Kimi Work, Gemini Spark, Microsoft 365 Copilot Cowork, and more.

Most of these alternatives are not quite yet on the same level as Claude Cowork; however, ChatGPT Work is very close. In our evaluations, ChatGPT Work with GPT-5.6 Sol produces largely comparable output with Claude Cowork running Fable 5.

A few minor issues, such as skills management, keep ChatGPT Work from reaching exact parity with Claude Cowork, but it's reasonable to expect ChatGPT Work to reach Cowork-level parity very soon. Many of its initial issues following its release have already been resolved.

ChatGPT Work runs on the cloud by default, which Claude Cowork now does as well. In the past, Claude Cowork worked exclusively locally on your own computer (the model itself always has run in the cloud, but your workspace existed on your own computer); however, Cowork now also runs on the cloud by default. Running it on your own computer gives great flexibility and speed, though running in the cloud allows it to run even when your computer is off, or even from your phone.

One obvious advantage ChatGPT Work presents over Claude is its GDPR compliance, which Anthropic is still struggling with. Claude Cowork can be made GDPR-compliant via third-party inference, but it's quite expensive to set up. Though, ChatGPT users wishing to have full EU data residency will need to buy ChatGPT Enterprise, which is a substantially more expensive license than ChatGPT Business, given that users essentially pay per token.

As ChatGPT Work and Claude Cowork converge, and other alternatives rise in addition, no one knows which AI tool will be at the top a few months from now, let alone a year from now. It could be one of these two, an alternative by a different lab, or an open-source alternative. This is why it's important for companies to set themselves up to be vendor-neutral, and now is an easier time than ever to do so.

Skills and MCP have become standard across all mainstream AI tools, beyond perhaps Microsoft 365 Copilot which is commonly around half a year behind the competition, but we now also have a new standard for vendor-neutral set-ups in the form of plugins.

Plugins themselves are not new. They are bundles of skills and (MCP) connectors, allowing teams to quickly group together common settings.

What is rather new is them becoming the standard for sharing skills and connectors across different tools. An AI administrator can establish an external source of truth, commonly on GitHub, listing all plugins in an organization. Tools like ChatGPT and Claude will then automatically sync plugins from this source of truth.

The main standard enabling this ability is the Agent Plugins 1.0 standard, which is an initiative by Amazon, Cursor, Microsoft, OpenAI, Google, and Vercel.

Stay tuned: later this autumn, we intend to share exactly how we set Monterro up to be vendor-neutral.

Try this

Put it into practice before the next issue.

1

Run Claude Cowork or ChatGPT Work in the cloud. Running the agent in the cloud means uploading all data in the workspace to the provider's server, enabling you to keep the work running even after turning off your computer, and controlling the AI from your phone while you're out for a walk. In Claude: first, select Cowork. Next, click on the icon in the top-right corner of the application to toggle running Cowork on your computer versus running a workspace in the cloud. In ChatGPT: first, select Work. Next, click on the icon in the bottom-right corner right below the chat window, underneath the “send” button.

2

Set up your AI to run chores on a schedule. Start by iterating with your AI in a chat until it gets the output just right, then ask the AI to turn the flow into a skill, and finally schedule the skill to run on a recurring basis. Monthly business reviews, weekly pipeline summaries, and daily calendar briefs all work great. Share the skill with your colleagues to easily let them schedule it as well.

3

Clone your voice. Ask your AI to read 100 e-mails that you've sent and extract your tone of voice into a skill.

Footnote: Before using Claude Cowork, check with your organisation that it complies with your internal policies. Avoid inputting sensitive or personal data into AI tools unless your organisation has given explicit approval.

Infuse AI

Monterro, built for the AI era

Accelerating AI transformation across our portfolio companies with Infuse AI.

“Frontrunners are now treating their AI agents as colleagues, letting them work across enterprise systems such as HubSpot, Salesforce, Figma, Jira, Slack, and more. As humans are no longer the primary users of enterprise software, entire applications will need to be adapted, in some cases rebuilt entirely, to serve AI.”

Jacob Bergdahl, AI Specialist at Monterro

Get Software Talks AI in your inbox

One email a month on AI for B2B SaaS leaders. Written by the Monterro team.