ChatGPT Drops Free-Tier Chat Limits, Google Prices Its Always-On Agent Down to Match Claude Cowork, and Washington Still Won't Say What Its Frontier-Model Review Checks For

Three moves landed within days of each other that only make sense read together: the free chatbot got more free, the $20-a-month agent tier just became a two-company price-match, and the government finished building a gate that sits in front of every frontier model's launch day without saying what triggers it. None of this is about a smarter model. It's about where the money and the scrutiny are actually concentrating as chat itself stops being the product.

1. ChatGPT's Free Tier Loses Its Rate Limit

OpenAI announced on August 6 that GPT-5.6 Luna becomes the default model for Free and Go users, with the previous rate limits — as tight as 10 messages every few hours on flagship models — lifting entirely for text chats the week of August 10. Free users also get a Think button that gives Luna extra time to reason through harder questions on a per-message basis. The limits that remain apply to file uploads, images, and other tools, and OpenAI says the unlimited tier is still subject to abuse guardrails. Paid Plus and Pro users get their own upgrade in the same release: a retuned GPT-5.6 Sol plus a slider for how much thinking effort the model spends per answer, across a product OpenAI says a billion people now use weekly.

// Before: free-tier text chat was rate-limited like a paid resource
if (user.tier === 'free' && messagesThisWindow >= 10) {
  return respondWithUpgradePrompt();
}

// After: text chat is unlimited on the free tier; the meter
// moved to a different layer entirely
if (user.tier === 'free' && wantsFileUpload) {
  return respondWithUpgradePrompt(); // limits now live on tools, not chat
}

Pulling rate limits off free text chat right as DeepSeek's V4-Flash undercuts frontier coding pricing by roughly 99% isn't a coincidence — it's a signal that plain conversational chat has stopped being the thing worth metering. If your product still gates a basic chat loop behind a message quota, this is the week that assumption got more expensive to defend.

2. Gemini Spark Gets Cheaper, Right as Claude Cowork Gets Everywhere

Google's always-on agent, Gemini Spark — the one that keeps working on a Google Cloud VM after your phone is locked and your laptop is closed — launched in May exclusively on the $99.99/month AI Ultra tier. On July 24 it expanded down to AI Pro subscribers in the US, a $19.99/month plan, putting Google's autonomous-agent price point within a couple of dollars of Anthropic's Claude Cowork, which requires nothing more than the $20/month Claude Pro plan and expanded from desktop-only to mobile and web on July 7. Two companies that shipped fundamentally different agent architectures — Spark's persistent cloud VM versus Cowork's desktop-first folder access — have converged on the same $20-ish monthly price band to get an autonomous agent into a user's hands, while reserving the $99.99 Ultra tier for higher usage caps and earlier access rather than a categorically different product.

That convergence matters more than either launch on its own. When two competitors independently land on the same price for the same job — "autonomous agent that acts on your files and calendar without supervision" — it stops looking like pricing strategy and starts looking like the market's actual clearing price for that capability. Anyone building an agent product priced meaningfully above $20/month now has to justify the gap.

3. The Frontier-Model Review Framework Shipped On Time — and Stayed Secret

The 60-day clock set by President Trump's June executive order on frontier-model national security review hit its August 1 deadline, and the administration delivered: a classified benchmarking process for deciding which models count as "covered," plus a roughly 30-day window for government pre-release access to those models. The White House met with OpenAI, Anthropic, Google, Microsoft, Nvidia, and others on August 4 to walk through it, then confirmed it won't publicly release the framework — not the capability thresholds, not the review criteria, not which companies have agreed to participate. Developers are told the process is voluntary and can't become a mandatory licensing regime, but it lands a few weeks after Commerce used export-control authority to pull Anthropic's Claude Fable 5 and Mythos 5 offline for 19 days and gate OpenAI's GPT-5.6 Sol behind government-vetted access for 12 — both ad hoc, both before this framework formally existed.

What Connects the Three

Read together, these are the same market pulling apart from both ends at once. Plain conversational chat is being given away with the meter removed, because it no longer differentiates anyone. The real subscription money is consolidating around a roughly $20/month autonomous-agent tier that two unrelated companies landed on independently. And above all of it sits a newly finished, deliberately opaque government checkpoint that every frontier model now has to clear before it ships — one whose rules the companies subject to it aren't allowed to show you. If you're pricing an AI product this quarter, free chat isn't your moat anymore and $20 isn't a floor, it's the going rate for the thing that actually is one.

— Maya