Our team will be out of office on Friday, May 1, 2026. We’ll be back and ready to assist you starting Monday, May 4th.

Should I Wait for the Next AI Model Before I Start?

Contents

Google lost roughly $225 billion waiting for a model that still has not shipped, Kimi K3 has a teaser video and no model card, and this article answers the question directly: no, and here is the framework for knowing what actually compounds versus what resets with every release.


The Answer Is No, and Here Is the Math

There is always a better model eight days out.

That is not a complaint about the pace of the industry. It is a structural description of it, and it is the single most expensive thing happening to entrepreneurs in 2026, because eight days out is not a date. It is a permanent address.

Here is the direct answer to the question in the headline. No, you should not wait, and the reason is not motivational. It is that model access does not compound and your process does. The next model will make your documented workflow better. It will not create one. So every quarter you spend waiting for a better component, you are not accumulating an advantage, you are accumulating a wish list, while the people who shipped on a worse model are three months into a compounding curve you cannot buy your way onto later.

The last two weeks made this unusually easy to see, because the industry ran the experiment for us in public.


Key Takeaways

  • Google delayed Gemini 3.5 Pro to July 17 after abandoning its architecture for a ground-up rebuild, amid senior researcher departures and roughly $225 billion coming off Alphabet’s market cap.
  • Kimi K3 has a teaser video, a since-deleted promo page dated July 15, and no model card, weights, or benchmarks, yet its leaked specs are being repeated as fact across the industry.
  • Meanwhile GPT-5.6 shipped July 9 in three purchasable tiers, and Claude Sonnet 5 became the free default for every user on July 1, with no countdown from either.
  • Model access does not compound. Documented process, prompt libraries, and team fluency do, and they survive every model change.
  • Inference costs for a given capability level have fallen more than 280-fold in roughly 18 months, which means waiting for cheaper is waiting for a train that already left and is coming back every quarter anyway.

The Problem: Roadmap-Watching Feels Exactly Like Diligence

I want to name this carefully, because the people doing it are not lazy. They are usually the most conscientious owners I talk to.

Here is what it looks like. An entrepreneur has read about every model. They can tell you the difference between the tiers. They know what is rumored to ship this month and what benchmark it is expected to clear. They have opinions about context windows. They have been “evaluating” for two quarters.

And they have deployed nothing.

That is not a knowledge problem. That is the most sophisticated form of procrastination ever invented, because it produces the exact felt sense of progress without any of the risk of failure. Reading about AI is genuinely informative. It is also completely safe. Nothing you read can embarrass you in front of a client. Nothing you read can turn out to have been the wrong choice. Nothing you read requires you to admit, out loud, that your first version was mediocre.

Building does all of those things. Which is why building is what actually moves a business and reading is what actually fills a Tuesday.

I know this pattern intimately because I lived in it. There was a stretch where I could speak fluently about AI capabilities I had never once used in my own business. I was current. I was informed. I was, in terms of anything a client would pay me for, standing completely still.

And the AI release cycle is engineered, whether deliberately or not, to keep you there. Teasers are marketing. Your anticipation is the product being sold. There is always a countdown, and the countdown always resets.

But what if the thing you are waiting for is not actually the thing that determines whether this works for you?


The Evidence: What Happened While Everyone Was Watching

Google spent $225 billion learning that waiting is not free.

Google DeepMind pushed Gemini 3.5 Pro to July 17 after abandoning the existing Gemini 2.5 Pro architecture entirely in favor of a ground-up redesign, targeting improvements in mathematical reasoning, SVG scene generation, and image quality to compete with GPT-5.6 and Anthropic’s Fable 5. The model brings a 2 million token context window and a Deep Think reasoning layer.

The delay landed alongside four senior DeepMind researcher departures in a single week, including Gemini co-lead Noam Shazeer to OpenAI and Nobel laureate John Jumper to Anthropic. Markets responded by wiping roughly $225 billion off Alphabet’s market cap. (BigGo Finance, The Agent Report)

If waiting for perfect costs Google a quarter of a trillion dollars, the idea that it costs you nothing deserves a second look.

Kimi K3 is a teaser video that the internet is treating as a product.

Moonshot AI posted a 36-second Kimi K3 teaser on July 15 with hidden “3” motifs. Beta selectors appeared in the app, desktop, CLI, and LMSYS Arena. A top-up promotion on Kimi’s Open Platform was dated to start July 15 and was subsequently deleted. As of today, there is no model card, no weights, and no benchmarks. The widely repeated specification, a 2.5T MoE with a 1M-token context window, is leak-sourced. Moonshot has published nothing. Community expectation has slipped to July 17-20. (Kimi K3 status tracker)

Watch how that specification is traveling. A leak gets a blog post. The blog post gets cited. The citation becomes “reported.” And within a week, a number nobody confirmed is being repeated in strategy conversations as though it were on a spec sheet.

Meanwhile, two frontier models shipped quietly and became the default for millions.

On July 9, OpenAI released GPT-5.6 publicly in three tiers: Sol, Terra, and Luna. Real tiers, purchasable today. (OpenAI, Axios)

On July 1, Anthropic made Claude Sonnet 5 the default for every Free and Pro user, describing it as performing close to flagship Opus 4.8 on many tasks. Not a paid tier. The default. For everyone. (Anthropic)

Neither of those needed a countdown. Both are in your hands right now.

And the cost argument for waiting collapsed a long time ago.

Stanford’s AI Index found that querying a model at GPT-3.5-equivalent capability fell from $20 per million tokens in November 2022 to $0.07 per million tokens by October 2024, a more than 280-fold reduction in roughly 18 months. (Stanford HAI)

Read that as a pattern, not a data point. Prices for a fixed capability level fall continuously. Which means “I will wait until it is cheaper” is not a strategy, it is a permanent condition. It will always be cheaper next quarter. You will always have been better off starting last quarter.


The Solution: Separate What Compounds From What Resets

Here is the reframe that got me out of the loop, and it is the whole article.

Some of your AI investment compounds. Some of it resets with every release. Almost everyone spends their time on the part that resets.

What resets when a new model ships:

  • Which model you picked
  • Model-specific tricks and workarounds
  • Benchmark knowledge
  • Your read of who is currently winning

What compounds regardless of what ships:

  • Your documented process, written down in a form someone else could execute
  • Your library of briefs that actually worked
  • Your team’s fluency, meaning the reflex to reach for AI at the right moment
  • Your definitions of done for the work you do repeatedly
  • Your judgment about which problems are worth pointing AI at

Look at those two lists and then look at where your last twenty AI hours went. If you are like most owners, and I was, you spent them almost entirely on column one. Column one is the fun one. It is also the one with a shelf life measured in weeks.

The strategic implication is straightforward. Build so the model is a component you can swap, not a foundation you poured. Document the process at a level above the tool. Write down the definition of done. Save the brief that worked. Then when Gemini 3.5 ships on the 17th, or Kimi K3 ships whenever it ships, you are not rebuilding. You are upgrading a part, and the upgrade takes an afternoon because everything around it was written down.

This is exactly what I do now, and the difference is not subtle. When Claude Sonnet 5 became the default on July 1, it did not disrupt anything I had built. It just made it better, for free, overnight. That is what it feels like to be positioned correctly for a release instead of held hostage by one.


Six Steps to Stop Waiting

1. Audit your consume-to-build ratio for one week. Track the hours you spend reading about AI versus the hours you spend building or shipping with it. Do not estimate, track. The number will be worse than your estimate. That gap is the diagnosis.

2. Count models read about versus models deployed. In the last six months, how many models have you read about? How many have you put into actual production use in your business? If that ratio is worse than 5:1, you do not have an information gap. You have an action gap, and more information will make it worse.

3. Set a research budget with a hard wall. Unbounded research expands to fill every available hour. Give it a number, say two hours a week, name the sources, and treat everything outside it as entertainment rather than work. Entertainment is fine. Just do not bill it to your strategy.

4. Force a verdict on everything you read. For each piece of AI news you consume, answer one question: does this change a decision in my business, and which one? If it does not change a decision, it was entertainment. Delete it and stop paying for it with your mornings.

5. Build the use case you have read about three times. If a use case has crossed your desk three separate times, it is not news anymore. It is a backlog item you have been avoiding. Pick the one with the best value-to-effort ratio and build it this week, on a model that exists today.

6. Change your scoreboard from informed to shipped. Nobody has ever been paid for being current. Design a five-minute monthly review that measures what you shipped, not what you know. Three to five metrics. If it takes longer than five minutes, you will not do it, and if you do not do it, the old scoreboard wins by default.


Frequently Asked Questions

Should I wait for Gemini 3.5 Pro before starting with AI?
No. Gemini 3.5 Pro was pushed to July 17 after a full architectural rebuild, and delays of that kind have a way of repeating. Meanwhile GPT-5.6 shipped July 9 and Claude Sonnet 5 became the free default July 1. Build on what exists, and treat whatever ships next as an upgrade to a working process rather than a prerequisite for starting one.

What if I pick the wrong model and have to redo everything?
That fear is what the compounding framework is designed to dissolve. If you document the process above the tool level and save your briefs, swapping models is a component change, not a rebuild. The people who get trapped are the ones whose entire workflow lived inside one vendor’s specific quirks and nowhere else.

Are the leaked Kimi K3 specs reliable?
Treat them as unconfirmed. Moonshot has published no model card, no weights, and no benchmarks. The 2.5T parameter and 1M-token context figures circulating widely came from a leak, not from the company. Do not make a business decision on a specification the manufacturer has never stated.

Will AI get cheaper if I wait?
Yes, continuously and forever, which is exactly why waiting for cheaper is not a plan. Stanford’s AI Index documented a more than 280-fold cost reduction for a fixed capability level in about 18 months. Prices will be lower next quarter and lower again the quarter after. There is no moment where the waiting ends.

What should I actually do this week?
Pick one task where a wrong answer is cheap to detect and reverse, write a real brief for it with a definition of done, run it on a model that is generally available today, and save the brief that worked. That is a complete first cycle, and it is worth more than another quarter of reading.


The Quarter You Cannot Buy Back

Here is what I want you to sit with.

Somewhere out there is a business roughly your size, in roughly your market, with roughly your resources, that started six months ago on a model that was worse than what you can access for free today. Their first version was mediocre. Their second was better. By now they have a documented process, a library of briefs that work, and a team that reaches for AI without being told to.

They did not get there because they picked the right model. The model they started on is obsolete. They got there because they started, and then kept going, and every release since then has made their existing system better while you have been evaluating.

That is the part that does not show up in any benchmark. Compounding does not care how smart your model is. It cares how long you have been running.

And the cruelest thing about it is that you cannot buy the six months back. You cannot catch up by picking a better model, because they will get that same model, and they will get it on top of six months of process you do not have. The gap does not close. It widens, quietly, one quarter at a time, while the reading feels like progress.

Gemini 3.5 Pro might ship tomorrow. Kimi K3 might ship this weekend. Something bigger is always about to land, and it always will be, forever, for the rest of your career.

Your competitive advantage was never going to be access to the frontier. It was always going to be the six months of compounding you got while everyone else refreshed a teaser video.


Jonathan Mast is the founder of White Beard Strategies, where he helps entrepreneurs implement AI in businesses without IT departments, AI engineers, or six-figure budgets. He is the creator of the Perfect Prompt Framework and a speaker on practical AI implementation. He spent longer than he would like to admit being extremely well-informed about tools he had never once used in his own business.

About the Author