AI Model Tiers Explained: Flash, Mini, Nano, and Where

Nano. Flash. Mini. Flash-Lite. Sol, Terra, Luna. Every AI company ships a ladder of models, and the names are not marketing noise; they encode a real engineering decision. Here are AI model tiers explained in one sentence: each tier trades capability for cost, speed, and where it can physically run, and the tier a product uses determines what your kid actually experiences. The model in the news is almost never the model in your kid’s homework app. Knowing the difference is the cheapest way to understand why an app feels smart in one moment and dumb in the next.

Key Takeaways

• Tiers exist because cost scales with model size. Published prices per million output tokens in 2026 ranged from $6 (GPT-5.6 Luna) to $50 (Claude Fable 5.1).
• Nano-class models run on the device. Gemini Nano runs inside Android’s AICore service on the Pixel 11’s Tensor G6, with no network required for documented tasks.
• Flash and Mini class models run in the cloud but are optimized for price and speed. Google’s Gemini 3.7 Flash launched at “half the original 3.6 Flash cost per million tokens.”
• Flagship models are for hard reasoning. Claude Fable 5.1 scored 52.6 percent on Terminal-Bench-Science 0.1 versus 22.4 percent for GPT-5.6 Sol, at $10 input and $50 output per million tokens.
• The compression techniques that create smaller tiers (distillation, quantization, pruning) are why small models punch above their weight, and why they are different models rather than shrunken copies.

The 2026 ladder, with real numbers

Start with the map, because the vocabulary is genuinely confusing and no company uses the same words.

On-device (Nano class). Gemini Nano is the smallest member of Google’s family and runs inside Android’s AICore system service. Google’s developer documentation says it lets you deliver generative AI “without needing a network connection or sending data to the cloud,” and lists concrete capabilities: summarize an article or conversation into bullets, proofread a short message, rewrite in a different tone, describe an image, transcribe speech, answer a custom prompt. The Pixel 11’s Tensor G6, announced August 12, 2026, has 50 percent more TPU compute than the G5 and runs on-device AI tasks “up to 3.5 times faster while using up to 3.5 times less energy.”

Cheap cloud (Flash, Mini, Lite, Luna class). These run in data centers but are built for volume. Gemini 3.7 Flash launched at half the per-million-token cost of Gemini 3.6 Flash, three weeks after it. GPT-5.6 Luna launched at $1 input and $6 output per million tokens. This is the tier that powers most consumer apps, because at scale the price difference decides whether a product can exist.

Mid tier (Terra class, Pro class). GPT-5.6 Terra at $2.50 input and $15 output. Described as offering performance comparable to a previous flagship at half the cost, which is exactly what a mid tier is for.

Flagship (Sol, Fable, Opus class). GPT-5.6 Sol at $5 input and $30 output. Claude Fable 5.1 at $10 and $50, with a one-million-token context window and 128K max output. Claude Opus 5, released July 24, 2026, at $5 and $25, positioned as near-flagship performance at half the price.

Restricted tier. A newer category worth knowing: Claude Mythos 5.1 is not available to the public at all, only to vetted US organizations through cyber and life-sciences verification programs. GPT-5.6-Cyber is similar, available only through Daybreak Red. Capability is now gated by who you are, not just what you pay.

The gap between tiers on hard tasks is large. On Terminal-Bench-Science 0.1, Fable 5.1 scored 52.6 percent while GPT-5.6 Sol scored 22.4 percent. On easy tasks the gap nearly vanishes, which is precisely why cheap tiers exist.

AI model tiers explained: how you make a small model good

A small model is not a big model with pieces removed. It is usually a different model trained specifically to imitate a bigger one. Three techniques do the work, and each one is teachable.

Distillation. Hinton, Vinyals, and Dean introduced this in 2015. Train a large “teacher” model, then train a small “student” on the teacher’s full probability distributions rather than just the correct answers. The soft targets carry more information than a hard label: they tell the student not only that the answer was “cat” but that “lynx” was a near miss and “car” was not. Students trained this way outperform students trained on labels alone.

Quantization. Store each weight with fewer bits: 8-bit or 4-bit instead of 16. Memory drops proportionally, memory bandwidth drops, and on the right hardware speed improves. Accuracy loss is small if done carefully. This is the single biggest reason a capable model fits on a phone.

Pruning. Remove connections that contribute little to the output. Networks are trained with more capacity than they strictly need, and a meaningful fraction can be cut with modest loss.

Now the part that decides which tier a product uses, which is not capability at all. It is unit economics. An app with a million daily users making ten requests each is making ten million requests a day. At $50 per million output tokens with 500-token answers, that is $250,000 a day. At $6 per million, it is $30,000. No product team picks the flagship for a high-volume feature unless the feature fails without it.

That arithmetic explains three things parents notice. Why a free tier feels less capable than the paid one: it is often a different, smaller model, not a throttled version of the same one. Why the same app is smart at one task and dumb at another: routing, where easy requests go to a cheap tier and hard ones escalate. And why your kid’s school district tool may be surprisingly limited: districts buy at volume, and volume means cheap tiers.

One more honest note. Small does not mean bad, and the ladder is moving upward fast. Google’s Gemma family passed a billion downloads and runs, per Google, in environments “from phones and edge infrastructure to space.” A Flash-class model in 2026 outperforms a flagship from 2024 on many tasks. The tier your kid uses today is better than the best available when they started school.

How to Teach Your Kid About Model Tiers

Ages 5–8: The three helpers

Three imaginary helpers: one lives in your pocket and answers instantly but only knows simple things; one lives across town, knows more, and takes a minute; one lives far away, knows the most, and charges a lot. Which one do you ask what time it is? Which one do you ask to help with a hard project? Kids sort this correctly and immediately, and they have just learned the entire concept of tiering.

Ages 9–12: Find the tier in your house

Go hunting. Which features on your phone work in airplane mode (on-device tier)? Which need internet but answer instantly (cheap cloud)? Which take several seconds and produce long careful answers (flagship)? Make a three-column list. Then have your kid guess which tier their homework app uses and explain their reasoning. There is no answer key, and the reasoning is the point.

Ages 13+: The spreadsheet

Give your teen the real numbers: $6, $15, $30, $50 per million output tokens. Have them build a spreadsheet for a hypothetical app with 100,000 daily users, 5 requests each, 400-token answers. Compute daily cost at each tier. Then ask the business question: which tier would you ship, and what would you do about the users whose questions the cheap tier fails? That is a real product decision, and a 15-year-old can reason about it well.

The question to ask: “If this app used a more expensive model, what would get better, and who would pay for it?”

Tier to device: what runs where

Tier	Examples	Where it runs	Cost per M output tokens	Best at	Fails at
On-device / Nano	Gemini Nano on Tensor G6	The phone itself, offline capable	No marginal cost; battery-limited	Summarize, proofread, rewrite, transcribe, describe images	Long reasoning, current information
Cheap cloud	GPT-5.6 Luna, Gemini Flash-Lite	Data center, optimized for volume	About $6	High-volume chat, classification, simple Q&A	Multi-step problems, subtle judgment
Fast cloud	Gemini 3.7 Flash, GPT-5.6 Terra	Data center	About $15	Most consumer app features	Research-grade tasks
Flagship	GPT-5.6 Sol, Claude Opus 5	Data center	$25 to $30	Coding, agents, hard reasoning	Cost at scale
Frontier	Claude Fable 5.1	Data center, 1M context	$50	Long-context agents, scientific workflows	Being affordable for high volume
Restricted	Claude Mythos 5.1, GPT-5.6-Cyber	Data center, vetted access only	Not publicly priced	Cyber and life-sciences work	Availability to anyone unvetted

Print this and the model-name confusion mostly dissolves. When a new name appears, place it on the ladder by its price and its context window, not by its adjectives.

What to actually do at home

Ask which model, and which version

When a school or an app says “powered by AI,” the useful question is which model and which tier. A district running a Flash-class model for 50,000 students is making a reasonable budget decision; a district that cannot answer the question has not evaluated the tool. Utah’s statewide Gemini partnership covers roughly 680,000 students and 28,000 educators at no cost for 2026-27, and families there should know which tier that means.

Match the tier to the task at home

Quick lookup or transcription: on-device is fine and more private. Learning a new math method: you want a model that shows steps, which usually means a bigger tier and a study mode. Teaching this matching explicitly turns your kid into someone who chooses tools rather than accepting defaults.

Use the free-versus-paid difference as a teaching moment

When your kid complains the free version is worse, that is often literally true and it is often a different model. Explaining the economics is more respectful than saying “it’s free, what do you expect,” and it teaches them something durable about how software businesses work. See our pieces on why AI models come in sizes and Gemini 3.7 Flash and cheap AI.

Do not chase the flagship

Frontier models are optimized for agentic coding and long-context research, not for teaching a 12-year-old fractions. A well-designed study mode on a mid-tier model beats a flagship with no pedagogy. Capability is not the binding constraint in your house.

What not to do

Do not assume a bigger tier is safer. Safety comes from safeguards, age settings, and product design, not from parameter count. A frontier model with no study mode and adult defaults is worse for a 13-year-old than a Flash-class model inside a well-built teen product.

What to Watch For Over the Next 3 Months

• Week 4: Your kid can name one on-device feature, one cheap-cloud feature, and explain why they are different.
• Month 2 red flags: Your kid assumes the newest model name means better for their homework, or cannot tell when an app has silently switched tiers.
• Month 3 self-check: When a new model launches, does your family ask about price and context window before adjectives? That is the reflex worth having.

Frequently Asked Questions

Which tier is in my kid’s phone?

If it is a recent Pixel, Gemini Nano runs on-device through AICore for documented tasks, and cloud Gemini handles the rest. Recent iPhones run Apple’s on-device models plus Private Cloud Compute. Most Android flagships from roughly 2023 onward have an NPU and some on-device capability. Test with airplane mode: what still works is local.

Is a smaller model less safe?

Not by size. Safety depends on the safeguards wrapped around the model and the product’s defaults. A small model inside a well-designed teen product is safer than a frontier model with adult defaults. Judge the product, not the parameter count.

Why do companies use so many confusing names?

Partly branding, partly genuine differentiation across price and capability, partly because the ladder changes every few months. The durable way to read a new name is by its published price per million tokens and its context window. Those two numbers place it accurately.

Does the tier affect hallucinations?

Larger models generally hallucinate less on hard factual questions, and no tier is immune. Grounding matters more than size: a small model that retrieves from a real document often beats a big model guessing from memory. Always verify, at every tier.

Can I choose the tier my kid’s app uses?

Usually not directly. Some consumer apps let paid users pick a model. School tools are chosen by the district. The lever you actually have is which app, and asking the school which model it runs.

Will small models eventually be as good as today’s flagships?

On many tasks, they already match flagships from two years ago, and Gemma-class models now run everywhere from phones to space deployments. Expect the ladder to keep shifting upward while the gap on the hardest reasoning tasks persists, because the flagships keep moving too.

---

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. Google. (2026, September 1). “Google AI updates: August 2026.” The Keyword. https://blog.google/innovation-and-ai/technology/google-ai-updates-august-2026/
2. Android Developers. “Gemini Nano.” https://developer.android.com/ai/gemini-nano
3. TechCrunch. (2026, July 9). “OpenAI launches its new family of models with GPT-5.6.” https://techcrunch.com/2026/07/09/openai-launches-its-new-family-of-models-with-gpt-5-6/
4. Anthropic. (2026, September 1). “Introducing Claude Fable 5.1 and Claude Mythos 5.1.” https://www.anthropic.com/claude-fable-and-mythos-5-1
5. Hinton, G., Vinyals, O., & Dean, J. (2015). “Distilling the Knowledge in a Neural Network.” arXiv 1503.02531. https://arxiv.org/abs/1503.02531
6. Axios. (2026, July 24). “Anthropic releases new model Opus 5.” https://www.axios.com/2026/07/24/anthropic-releases-new-model-opus-5
7. Terminal-Bench. “Terminal-Bench-Science 0.1.” https://www.tbench.ai/news/terminal-bench-science-0-1