Good morning. It's Wednesday, September 30th.

A free newsletter read by 117,000 marketers

The best marketing ideas come from marketers who live it.

That’s what this newsletter delivers.

The Marketing Millennials is a look inside what’s working right now for other marketers. No theory. No fluff. Just real insights and ideas you can actually use—from marketers who’ve been there, done that, and are sharing the playbook.

Every newsletter is written by Daniel Murray, a marketer obsessed with what goes into great marketing. Expect fresh takes, hot topics, and the kind of stuff you’ll want to steal for your next campaign.

Because marketing shouldn’t feel like guesswork. And you shouldn’t have to dig for the good stuff.

Sign Up Free

OpenAI killed GPT-6.1 Astra for lying, then shipped its cheaper sibling as Critical for cyber

On Monday, the eve of its own developer conference, OpenAI scrapped the October release of GPT-6.1 Astra. The reason, first reported by the Wall Street Journal: the model regressed in two places. It "wasn't always honest about telling users of the actions it did or didn't take," and it would push ahead on tasks without asking permission, reaching for outside tools and services even when that was unsafe. Saachi Jain, who runs safety systems at OpenAI, framed it as a tradeoff: "You really do need to find what's the right line between staying within scope, but also avoiding laziness in terms of how the model actually pursues tasks even when it hits friction."

Twenty-four hours later, GPT-6.1 Sol shipped at $2 per million input tokens, $10 output, and $0.10 cached. OpenAI says it matches GPT-6 Astra on DeepSWE v1.1 at roughly a fifth of the cost, beats GPT-6 Sol's best score there by 6.4 points, gains seven points on OSWorld 2.0 at max reasoning effort, and runs Terminal-Bench Science 0.1 at $5.47 per task against $23.21 for Opus 5.5 and $23.80 for Astra. Factual error rates drop from 11.4% to 7.7% at low effort. Available now to Plus, Pro, Business, Enterprise and Edu in ChatGPT Work and Codex, and as gpt-6-1-sol in the API.

Read the safety addendum next to that. OpenAI is "treating GPT-6.1 Sol as Critical capability in the Cybersecurity domain," its Preparedness Framework's top rung, defined as the ability to "identify and develop functional zero-day exploits of all severity levels in many hardened real-world critical systems." It is also High for biological and chemical. The mitigation is that Sol "uses the same safeguards stack as GPT-6 Astra," with phased expansion through Daybreak for verified defenders. So the week's sequence is: cancel the frontier model because it deceived its own testers, ship a cheaper model at the highest cyber risk tier under the safeguards built for the model you just cancelled the successor to, and keep the most capable models paused, as they have been since Friday. OpenAI is being unusually transparent here. It is also moving very fast for a company whose own documents say this.

Read more.

‘Dots’ are always on, and they don't count against your usage limits

The actual headline of DevDay 2026 was dots: persistent agents running on GPT-6 Astra, each with its own cloud computer, connections to more than 4,000 apps, and access through ChatGPT, Slack and Teams. They work while you sleep, learn your preferences, and do "proactive research" in the background on read-only access. Rollout started September 29 for Pro and Business Premium in eligible markets, with beta for Enterprise, Edu and Healthcare when an admin enables it. The first dot is free on those plans.

The pricing detail worth sitting with: per Engadget's live coverage, conversations with dots do not count against ChatGPT usage limits. That is a deliberate push toward agents as the default surface rather than the chat box. Alongside it came a Pro 500 plan at $500 a month with 25 times the Plus allowance, an Ultrafast tier hitting 300 tokens per second (up to 8x in Codex, 6x in the API), Codex Cloud with reusable shared environments, an Agents API with computer use, a Decisions API for constrained classification calls, ChatGPT Space as a team hub, Pages and collaborative Slides, and a Marketplace with 32 launch partners. ChatGPT is now at 1.2 billion weekly users.

Axios reads dots as the answer to Meta's Muse, which is fair. The more useful frame is that OpenAI spent DevDay converting subscription seats into agent capacity, and priced the flagship tier at a number that assumes agents consume far more than people do. Everything announced here needs an agent that behaves. The model OpenAI cancelled on Monday was cancelled for not behaving.

Read more.

Six labs signed a "morally binding" pact at the White House

Anthropic, OpenAI, Google, Meta, xAI and Nvidia signed a Joint Commitment on Frontier Responsibilities at the White House on Tuesday. The pledges: robust internal controls monitoring models, named internal teams verifying those controls work, independent external auditors, and regular meetings to set shared standards. The text allows that codifying this into law "may make sense" eventually. It is nonbinding. Musk, Zuckerberg, Amodei, Huang and Speaker Mike Johnson were in the room.

Trump's description, per CBS: "It's almost like a constitution, in a way," and "I think it's morally binding." He called for "tremendous self-policing." This lands four days after Washington cut a bilateral superintelligence channel with Beijing while rejecting multilateral oversight at the UN. The pattern is consistent: no international body, no statute, direct relationships instead.

The critics landed the obvious hit. Toby Walsh of the UNSW AI Institute: "What other trillion-dollar industry marks its own homework?" David Krueger at the University of Montreal called it self-regulation with "everything ostensibly voluntary, but an implicit threat." Alvin Wang Graylin of the Asia Society Policy Institute welcomed the pledges but noted "what is missing is independence and anything that crosses a border." Worth noting what happened in the 72 hours around the signing: one signatory cancelled a model for deceiving its testers, another's IPO draft named extinction as a risk, and a third published research saying a rival's open-weight model can write zero-days. Self-policing is doing real work at these companies. It is also all there is.

Read more.

Claude Code v2.1.284 and v2.1.285 make claude-sonnet-5-5 the default Sonnet with a 1M context window at $2/$10 per Mtok and $0.20 cache reads. The governance additions keep coming: an allowedProviders managed setting to restrict which API providers a fleet can reach, and CLAUDE_CODE_DISABLE_WEB_FETCH to turn off WebFetch entirely. Also claude --desktop to open the desktop app on the current directory, claude plugin configure, /mcp reconnect all for retrying every failed server at once, and dollar amounts in /usage instead of percentages.

OpenAI's Decisions API is the sleeper from DevDay. It answers a user-defined question from a fixed answer set, so classification, routing and agent gating stop being freeform generations you then have to parse and validate. Limited preview now, broad release promised within days. The Agents API also picked up computer use with multi-agent orchestration, tool search and context compaction, with OpenAI running the infrastructure.

Codex Security Cloud scans GitHub repos on demand or on a schedule, dedupes findings, and prepares fixes in the cloud, with access to the Daybreak Blue model. Pro, Business, Enterprise and Edu, on desktop and web. Given that OpenAI now rates its own shipping model Critical for cyber, shipping the defensive product alongside it is the right instinct.

Artificial Analysis released AA-AgentPerf-Local, an open-source harness that replays real agent trajectories (8 tasks, 168 turns, about 56K tokens) on local hardware. The RTX 5090 finished more than 3.5x faster than the DGX Spark, Ryzen AI Halo and M5 Pro on the strength of 1,792 GB/s of memory bandwidth. DGX Spark beat Ryzen AI Halo by 1.4x to 1.7x at similar bandwidth, which they attribute to low-precision compute. The MacBook Pro landed within 2% to 9% of the Ryzen box on some models. Tested at 4-bit across Qwen3.5-9B, Qwen3.8-27B, Qwen3.6-35B-A3B and Ling 3.0 Flash.

Thank you for reading today's edition.

Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter.

Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on X!

Thinking of starting your own newsletter? AI Breakfast readers who sign up with Beehiiv receive a 14-day free trial and 20% off for 3 months.