AI Model Versions Explained: What a Number Hides

Here are AI model versions explained with one real example. In September 2026 OpenAI launched GPT-6. A few weeks later, OpenAI’s own write-up of its Navier-Stokes work said the mathematics had been done not by GPT-6 Astra but by “an internal model that is significantly more capable than GPT-6 Astra.” On September 30, Google released Gemini 4 Argon, which TechCrunch reported Google compared against “GPT-6 Astra” and Anthropic’s “Fable and Opus.”

Count what is in that paragraph: a number, a sub-name, a codename, and a model with no public name at all that is better than the numbered one. The number is the least informative thing on the list.

Key Takeaways

• In software, version numbers mean something specific. Semantic Versioning 2.0.0, authored by Tom Preston-Werner, defines MAJOR for “incompatible API changes,” MINOR for backward-compatible additions and PATCH for backward-compatible bug fixes. AI model names do not follow this.
• An AI model name is three things fused together: a family, a release order, and a marketing tier. It is not a compatibility contract and it is not a quality score.
• The publicly numbered model is often not the best one a lab has. OpenAI stated plainly that the model used for its September 8, 2026 Navier-Stokes work was internal and more capable than the released GPT-6 Astra.
• What actually determines your kid’s experience is the variant served to them, which can depend on plan tier, region and server load. The product name stays the same while the model underneath changes.
• The fix already exists on paper. Mitchell and colleagues proposed model cards at FAT* in 2019: documentation of intended use, evaluation procedure and performance characteristics. Ask for one before trusting a number.

What a Version Number Means in Real Software

Software has a convention for this, and it is worth knowing because it shows what AI naming is not.

Semantic Versioning 2.0.0 sets three rules. Increment the MAJOR version “when you make incompatible API changes.” Increment MINOR “when you add functionality in a backward compatible manner.” Increment PATCH “when you make backward compatible bug fixes.” So 2.4.1 going to 3.0.0 is a promise that something you depended on will break. Going to 2.5.0 is a promise that it will not.

Notice what that system measures: compatibility, not quality. Version 3.0.0 is not better than 2.9.9. It is incompatible with it. A developer reading a version number learns what work they have to do, which is a precise and limited piece of information.

Now compare “Gemini 4 Argon.” The 4 signals release order inside Google’s Gemini family. “Argon” is a variant name. Nothing in the string tells you whether code written against Gemini 3 will still work, what the context window is, or what the training data cutoff was. It is a product name wearing a number.

That is not a scandal. Cars do the same thing, and nobody expects a Civic Type R to follow semantic versioning. The mistake is reading an AI model name as though it carried a technical guarantee, because the number looks like the numbers engineers use.

AI Model Versions Explained: The Four Things a Name Carries

Break any current model name into parts and you get at most four signals.

The family. GPT, Gemini, Claude, Llama. This tells you the lab and, loosely, the architecture lineage. It is the most reliable part of the name.

The generation number. GPT-6, Gemini 4. This is ordering within the family, and it is comparable inside one company and meaningless across companies. GPT-6 is not “two better” than Gemini 4. The numbers were never on the same scale.

The tier or variant. Astra, Argon, Opus, mini, flash, pro. This usually encodes size, speed and price rather than capability in any single direction. A smaller variant is often better for a kid’s homework because it answers faster and costs less, and the difference on a careful task may be small.

The release date, implied. This is the part that genuinely predicts capability most of the time, and it is the part the name leaves out.

Here is what the name hides, and each item changes what your kid experiences:

The training data cutoff. A model that stopped learning in March will not know about a June event, and it will often answer anyway.

The context window. How much text it can hold at once. This matters enormously for a student pasting in a long reading.

Silent updates. The same product name can be served by a quietly revised model. Behaviour shifts and nothing in the interface tells you.

Routing. Which variant you get can depend on your plan, your region and current demand. Two kids in the same class can type the same question into the same app and be answered by different models.

The unnamed frontier. OpenAI’s September 2026 statement is the cleanest public example. The work was done by an internal model, more capable than the released one, coordinating around 10,000 concurrent agents that exchanged 2.7 million messages and used roughly 130 billion output tokens. None of that capability had a version number the public could buy.

What the Number Tells You, and What It Does Not

Question a parent actually has	Does the model name answer it?	Where the answer lives
Is this newer than what we used last term?	Usually yes, within one company	Release notes
Is it better than a rival’s model?	No	Independent benchmarks, with protocol stated
Will my kid’s old project still work?	No	API deprecation notices
How recent is its knowledge?	No	Model card or documentation
How much text can it read at once?	No	Documentation
Which exact model am I being served?	No	Account settings, or an API response field
Is there a better model this lab has not released?	No, and often yes there is	Company research posts

Six of seven rows say no. That is the article in one table.

How to Teach Your Kid About AI Model Versions

Ages 5–8: Draw the same house three times

Materials: three sheets of paper, crayons.

Your child draws the same house three times, labels them 1, 2 and 3, and writes or dictates one sentence under each about what changed. Chimney added. Windows fixed. Tree moved.

Then ask the question that does the work: “is number 3 the best one?” Let them argue. Often it is not, because they got bored by the third. The point lands when they realise the number records order, not quality. Pin all three on the fridge in order. A six-year-old who can say “3 is the newest, not the best” has understood something most headlines get wrong.

Ages 9–12: The LEGO instruction test

Materials: ten to fifteen LEGO bricks, paper, pencil.

Your child builds a small model, then writes instructions another person could follow. Now make two changes, one at a time.

Change one: swap a red brick for a blue brick of the same shape. The instructions still work. That is a MINOR change.

Change two: remove the base plate the whole thing was built on. The instructions no longer work. That is a MAJOR change.

Have them write v1.0.0, v1.1.0 and v2.0.0 on three index cards next to the three states of the model. That is semantic versioning, taught correctly, with bricks. Finish with the real-world hook: ask them to guess whether “GPT-6” means something broke compared to GPT-5. It does not, and now they know why that question has no answer from the name alone.

Ages 13+: Find out what you are actually using

The assignment is an audit of one AI tool they use. Five facts to find, written down: the exact model name and variant, the stated training data cutoff, the context window, whether the tool documents which model serves which plan, and whether a model card exists.

Then the fun part. Ask the assistant itself which model and version it is. Compare that answer with the documentation. The two frequently disagree, because the model is predicting a plausible answer about itself rather than reading a configuration file. That single experiment teaches more about how these systems work than a month of explanation.

The question to ask: “If the company replaced the model behind this app tonight and kept the name, how would you find out?”

What to Do at Home

Write down the model, not the app

When your kid uses an AI tool for something that matters, note the model name and date in the same place as the work. One line. Months later, when behaviour has changed and they insist “it used to do this,” you will have the evidence. This is ordinary engineering hygiene and it costs nothing.

Check the cutoff before trusting a current fact

The single most common failure in a student’s AI-assisted work is a confidently stated fact from before the model’s cutoff, presented as current. Teaching a kid to ask “when did your training data end?” first, and to treat anything after that as unverified, prevents most of it. This is a one-sentence habit with a large payoff.

Treat tier names as price information

Astra, Argon, mini, flash, pro. These mostly tell you about cost and speed. For homework, the cheaper faster variant is frequently the right call, and a family that defaults to “the biggest model” is often paying for latency it does not need. Try the small one first on real tasks and compare.

Ask for the model card at school

If a school or district adopts an AI tool, the useful request is not “is it safe.” It is “can we see the model card and the evaluation documentation.” Mitchell and colleagues set out in 2019 exactly what that document should contain, including intended use and performance across groups. A vendor who cannot produce one has told you something.

What not to do

Do not let a kid use the version number as a status marker. “I’m on 6, you’re on 5” is a social dynamic, not a technical one, and with the top models converging to within a 0.7% gap on benchmarks according to Stanford’s AI Index, it does not even track capability. The habit to build is checking what you have, not announcing it.

What to Watch For Over the Next 3 Months

• Week 4: Find out, for one tool your family uses, which exact model serves your account. If the product does not tell you, that is the finding, and it is worth knowing before you rely on the tool for school work.
• Month 2 red flags: A model’s behaviour changes noticeably with no announcement. A deprecation notice arrives for a model a kid’s project depends on. A school tool switching its underlying model mid-term without telling parents. All three are routine and all three are worth noticing.
• Month 3 self-check: Ask your teenager to explain the difference between a version number that promises compatibility and one that signals marketing position. If they can explain it with the LEGO example, it has stuck. If they say “bigger number is better,” run the activity again.

Frequently Asked Questions

Is GPT-6 better than Gemini 4?

The numbers cannot answer that, because they count generations inside different companies and were never on a shared scale. The only honest comparison is a named benchmark run under a stated protocol, and Stanford’s AI Index reported the top ten models sitting within a 5.4-point band, which means the answer varies by task.

Why do labs give models codenames like Astra and Argon?

Variants need distinguishing, and a word is easier to remember than a string of digits. The names usually encode a tier, which relates to size, speed and price. They rarely encode a capability guarantee, so treat them as product labels.

Does a model get updated without the name changing?

Yes, this happens, which is why writing down the date alongside the model name is useful. Behaviour can shift under a stable product name. Checking documentation and release notes is the only reliable way to notice.

Is the best model always available to buy?

Not necessarily. OpenAI’s own September 2026 Navier-Stokes write-up described using an internal model “significantly more capable” than the publicly released GPT-6 Astra. Labs routinely run internal systems ahead of what they ship, for cost and safety reasons.

Should I pay for the highest tier for my kid’s homework?

Usually no, at least not first. Smaller variants answer faster and cost less, and for typical homework the quality gap is often small. Test both on three real assignments and decide from the comparison rather than from the tier name.

What single question exposes a vague model claim?

“Which exact model and variant, as of what date?” If a school, a vendor or a review cannot answer that, nothing else they say about performance can be checked. It is the same question engineers ask, and it works just as well at a parent-teacher meeting.

---

About the author

Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.

---

Sources

1. Preston-Werner, T. “Semantic Versioning 2.0.0.” https://semver.org/
2. Mitchell, M., Wu, S., Zaldivar, A., Barnes, P., Vasserman, L., Hutchinson, B., Spitzer, E., Raji, I. D., & Gebru, T. (2019). “Model Cards for Model Reporting.” FAT ‘19: Conference on Fairness, Accountability, and Transparency*. https://arxiv.org/abs/1810.03993
3. OpenAI. (2026, September 8). “Navier–Stokes solution.” https://openai.com/index/navier-stokes-solution/
4. Ropek, L. (2026, September 30). “Google releases Gemini 4 Argon, called its most powerful model yet.” TechCrunch. https://techcrunch.com/2026/09/30/google-releases-gemini-4-argon-called-its-most-powerful-model-yet/
5. Stanford Institute for Human-Centered AI. (2025). AI Index Report 2025, Chapter 2: Technical Performance. https://hai.stanford.edu/ai-index/2025-ai-index-report
6. Wikipedia contributors. (2026). “2026 in science.” Wikipedia. https://en.wikipedia.org/wiki/2026_in_science
7. Wikipedia contributors. (2026). “2026 in artificial intelligence.” Wikipedia. https://en.wikipedia.org/wiki/2026_in_artificial_intelligence

Related reading on HiWave Makers: what GPT-6 changed for families, three frontier models in five weeks, and how large language models work.