Frontier AI: What It Means, Who Builds It, and How to Judge It

This guide gives you the official definitions, the compute thresholds lawmakers now use, the flagship models and their API prices as checked on 2026-10-25, and a practical way to judge them yourself. Every number links to the page we read it from. We have not run hands-on tests for this article; where our own benchmark comes in, we say so plainly.

What is frontier AI? The definitions that actually get used

There is no single owner of the term. Three definitions do most of the work, and they disagree in useful ways.

The UK government set out a working definition for the 2023 AI Safety Summit. Its discussion paper defines frontier AI as "highly capable general-purpose AI models that can perform a wide variety of tasks and match or exceed the capabilities present in today's most advanced models." Note the anchor: today's models. The bar resets every time a new flagship ships.

The Frontier Model Forum, an industry body established in 2023, uses a stricter, relative test for membership. A frontier model is "a general-purpose model that outperforms, based on a range of conventional performance benchmarks or high-risk capability assessments, all other models that have been widely deployed for at least 12 months."

One distinction helps a lot. A frontier model is the underlying system, such as a model you call through an API. "Frontier AI" is broader and includes the apps built on top of it. When people talk about the AI frontier in news coverage, they usually mean the handful of labs pushing that line forward.

How regulators draw the line: compute thresholds

Capability benchmarks are slippery, so lawmakers who regulate frontier AI reached for something countable: the amount of computation used to train a model.

The EU AI Act's Article 51 says a general-purpose model is presumed to have "high impact capabilities" when its cumulative training compute is greater than 10^25 floating point operations. That presumption puts it in the "systemic risk" tier with extra obligations. The same article lets the European Commission amend the threshold as hardware and algorithms improve.

California went further up the scale. According to the Future of Privacy Forum's explainer, SB 53, signed by Governor Newsom on September 29, 2025, covers "foundation models trained with more than 10^26 computational operations," counting fine-tuning as well as the initial run. The heavier duties fall on "large frontier developers" with annual gross revenues above $500 million.

How crowded is that zone? Epoch AI counted over 30 publicly announced models from 12 developers above 10^25 FLOP as of June 2025, and saw roughly two new ones per month during 2024. So the EU line already captures dozens of models, not a tiny elite.

Here is our honest take: compute is a good legal trigger and a poor buying guide. It is easy to audit, which is why regulators like it. But it tells you nothing about whether a model writes a clean contract summary in Indonesian, or whether a cheaper model trained with less compute does your job just as well. For decisions, you still need task-level evidence. If you want the infrastructure side of this story, our piece on AI data centers covers where that compute physically lives.

Who builds frontier AI in 2026: flagship models and prices

The Frontier Model Forum lists six members: Amazon, Anthropic, Google, Meta, Microsoft and OpenAI. That list is not the whole frontier. The 2026 AI Index from Stanford HAI says the US and China performance gap "has effectively closed," with the top US model leading the top Chinese model by only 2.7% as of March 2026. It also says industry produced over 90% of notable frontier models in 2025.

Below are the flagship API models from the major labs, with list prices read from each official page on 2026-10-25. Prices are in USD per one million tokens (a token is roughly a word fragment). We call these "flagship" because they are the top general models our leaderboard tracks for each lab, not because each one is confirmed above a legal threshold. Some labs sell an even pricier tier, as the Anthropic note below shows.

Three details in that table are easy to miss.

Prices move on fixed dates

Google's page states that Gemini 4 Argon rises to $4 input and $20 output starting December 1, 2026. A cost estimate you make today can be wrong in five weeks.

Time of day can halve the bill

DeepSeek charges peak rates during set UTC windows on weekdays and half price off-peak, which works out to $0.66 input and $1.98 output for V4 Pro. Batch jobs scheduled overnight in Asia can land in the cheap window.

The top tier is not always open to everyone

Anthropic's pricing page lists Claude Fable 5.1 at $10 input and $50 output, above Opus 5.5. It also lists Claude Mythos 5.1 at the same price, marked "limited availability": it is available only to organizations verified through Anthropic's verification programs. The true edge of the frontier is sometimes gated, which is one more reason to judge models you can actually use.

Frontier models vs the AI tools you actually use

Most people never touch a raw frontier model. They use frontier AI through an app, and the app decides which model answers.

OpenAI is unusually clear about this. Its chat-latest model page says the alias "points to the latest Instant model currently used in ChatGPT," and the same page recommends GPT-6 Astra for production API usage. In other words, the default ChatGPT experience and OpenAI's flagship are not the same model. The price pages above show the same pattern of several tiers sold side by side at other labs too.

The price gap between tiers inside one vendor is often larger than the gap between vendors. On OpenAI's own pricing page, GPT-6 Astra output costs $50 per million tokens and GPT-6 Luna costs $0.50, a 100x difference. On Anthropic's page, Claude Opus 5.5 output is $20 and Claude Haiku 5.5 is $0.50 for prompts up to 100,000 tokens, a 40x gap.

That leads to a contrarian point we hold firmly: for most daily work, the frontier model is the wrong default. Drafting emails, summarizing a meeting or cleaning a spreadsheet rarely needs the top tier. Start with a mid or low-cost tier, and move up only when you can see a quality difference on your own tasks. Paying flagship prices for every request is the quickest way to inflate an AI bill without improving the output.

If you are choosing a consumer app rather than an API, compare plans side by side on our ChatGPT vs Gemini comparison, browse ChatGPT alternatives, or start with the free AI tools list. Plan prices change often, so each tool page in our directory shows the current price with its source and check date.

Why benchmarks do not settle who leads

By definition, frontier AI is supposed to beat everything else on benchmarks. In practice, the scores are getting less informative at the top.

Saturation

The 2026 AI Index reports that performance on SWE-bench Verified, a widely cited coding benchmark, "rose from 60% to near 100% in a single year." When everyone scores near the ceiling, the benchmark stops separating the leaders. For coding specifically, our AI for coding page looks at tools rather than raw scores.

The jagged frontier

The same report notes that "Gemini Deep Think earned a gold medal at IMO, yet the top model reads analog clocks correctly just 50.1% of the time." Capability is uneven. A model can solve olympiad mathematics and still fail a task a child handles. Averages hide that.

Language

Most headline benchmarks are written in English. A model's English score says little about how it handles Indonesian formality levels, regional idioms or local legal terms. Regional efforts help: AI Singapore's SEA-HELM leaderboard evaluates models on Southeast Asian languages including Indonesian, covering chat, instruction-following and linguistic tasks.

This is the gap Newly Stack is built to cover. Our Bahasa Indonesia benchmark, NS-ID, gives 20 AI models from 11 vendors the same 60 Indonesian tasks, with the selection rule, scoring rubric and costs published. The first run (25 October 2026) put 12 models within 3 points of the top score of 99.3, so at the top the suite separates weaker models better than it ranks the leaders. Writing tasks are graded by AI judges and labeled model-graded. Paid placements do not change selection or scores.

Risks and safety rules around frontier AI

The policy energy around frontier AI exists because capability cuts both ways. The NCSC says these systems "can perform complex tasks which can generate benefits and risks to cyber security." The 2026 AI Index counts documented AI incidents rising to 362, up from 233 in 2024.

Regulation is now specific. Under California SB 53, as summarized by the Future of Privacy Forum:

• Large frontier developers must publish an annual framework describing how they identify and mitigate catastrophic risks.
• All frontier developers must publish a transparency report before deploying a frontier model.
• Critical safety incidents go to California's Office of Emergency Services, generally within 15 days, or within 24 hours when there is imminent danger of death or serious injury.
• The Attorney General can seek civil penalties of up to $1 million per violation.

The industry side runs in parallel. The Frontier Model Forum describes itself as a non-profit "focused on addressing significant risks to public safety and national security," working on best practices, research and information sharing.

What do frontier AI rules mean for an ordinary user or a small business? Mostly that the big labs now publish more about how they test. Read those safety and transparency documents when they exist; they often state what a model should not be used for. And keep the basics: do not paste secrets into any chatbot, and verify any factual claim before you publish it.

How to choose and use frontier AI without overpaying

Here is the short process we recommend.

1. Write down three real tasks. Use your own documents, in the language you work in. Generic prompts reward generic models.
2. Test one model per tier. Try a flagship, a mid tier and a low-cost tier from at least two vendors. The table above gives you a starting list.
3. Score blind if you can. Hide which model wrote which answer, then rank them, so the brand name does not sway your judgment.
4. Price the winner at your volume. Multiply your expected monthly tokens by the official rate, and check for date-based changes like Google's December 1 increase.
5. Re-check every quarter. The frontier moves. A model that wins today may lose to a cheaper one in three months.

For Indonesian-language work, add one step: look for evidence in Indonesian before you trust an English leaderboard. Regional boards such as SEA-HELM, and our NS-ID results, are a better first filter than global rankings. If you are curious about the labs from China in the table, our overview of Chinese AI goes deeper on DeepSeek, Kimi and Qwen.

Frontier AI will keep shifting faster than any single article. The useful habit is not memorizing today's leader; it is knowing where to find dated, sourced numbers and how to test them on your own work. Our leaderboard updates prices with check dates, and you can join the newsletter for launch winners, new test results and deals once weekly editions start.