The phrase "frontier model" shows up in AI policy documents, procurement contracts, vendor marketing, regulatory filings, and architecture proposals. It is used as though everyone agrees what it means. In practice the term has at least three distinct working definitions in active use through mid-2026, and the choice among them changes what a sentence means. This piece sorts the definitions, identifies which models are in the frontier set under each, walks through how the term is being used in the EU AI Act, in the United States executive order on AI, in major procurement RFPs, and in production architecture decisions, and provides a working definition that the rest of the Digital Matters AI coverage uses when the term appears.
The short version is that a frontier model is a large-scale general-purpose AI model that sits at the current capability ceiling of what any organization has publicly deployed. The set as of mid-2026 includes Anthropic’s Claude Opus 4.8, OpenAI’s GPT-5.5, Google DeepMind’s Gemini 3 Pro, and (depending on how the term is interpreted) a small handful of additional models from xAI, Meta, and Mistral. The set changes when a new model that pushes the ceiling is released or when an existing frontier model is succeeded by a newer version from the same vendor. The set is small (typically four to eight models at any given time) and it is contested at the edges (vendors have an incentive to argue that their flagship belongs in the set and competitors have an incentive to argue that the bar is higher than competitor flagships clear).
Three working definitions in active use
The first working definition, sometimes called the "capability ceiling" definition, treats "frontier" as a relative ranking. A model is a frontier model if it is among the highest-capability general-purpose models available at the time of measurement. This is the operational definition that AI labs, capability researchers, and benchmarks use. It is the definition implied when an AI evaluation report says "we tested the major frontier models." The set under this definition changes whenever a new model is released that clears the bar previously set by the highest-capability models.
The second working definition, sometimes called the "compute threshold" definition, treats "frontier" as a quantitative threshold on training compute or on parameter count. The original version of this definition came from the United States executive order on AI signed in October 2023, which set a reporting threshold of 10^26 floating-point operations for training compute. Any model trained with more compute than that threshold triggered reporting requirements. The compute-threshold definition is the basis for most of the United States government’s use of the term and for several follow-on definitions that other regulators have adopted.
The third working definition, sometimes called the "general-purpose AI with systemic risk" definition, is the European Union’s approach in the AI Act. The Act distinguishes between "general-purpose AI models" (a broader category that includes most of the major LLMs) and "general-purpose AI models with systemic risk" (a narrower category that triggers additional obligations). The narrower category roughly tracks the capability-ceiling definition, with the threshold defined as 10^25 floating-point operations of training compute or by explicit designation. The EU approach is more procedural than the United States approach: even if a model does not cross the compute threshold automatically, the AI Office can designate it as having systemic risk if its capabilities warrant.
The three definitions converge in practice on a similar set of models. A model that sits at the capability ceiling almost always was trained with frontier-scale compute, and a model trained with frontier-scale compute is almost always at or near the capability ceiling. The set divergence at the edges is small enough that for most purposes the term can be used without specifying which definition is in use. The edge cases that matter are open-weight models like Llama 4 and DeepSeek V4, which are arguably in the frontier set under the capability-ceiling definition but were trained below the compute thresholds in the regulatory definitions, and which fall outside the EU’s systemic-risk obligations because they are open weights.
The current frontier set
As of mid-June 2026, the working capability-ceiling frontier set includes:
The first member is Anthropic’s Claude Opus 4.8, released in late April 2026 after the rapid 41-day frontier-model cycle from Opus 4.6 to 4.7 to 4.8 that defined the spring 2026 capability race. Opus 4.8 holds the top position on the most capability-dense benchmarks for code, reasoning, and long-context analysis. It is also the model whose Fable 5 and Mythos 5 variants were abruptly suspended on June 12, 2026, leaving Opus 4.8 as the production Anthropic frontier offering through this writing.
The second member is OpenAI’s GPT-5.5, released in late January 2026 and given a substantial mid-cycle capability bump in April 2026 with the "improved reasoning and toolformer" update. GPT-5.5 holds the top position on agentic benchmarks and on multimodal tasks. The expected GPT-6 release in the second half of 2026 will push the frontier set further, but as of this writing GPT-5.5 is the OpenAI member.
The third member is Google DeepMind’s Gemini 3 Pro, released alongside Gemini 3 Flash in early February 2026 as the flagship of the Gemini 3 family. Gemini 3 Pro holds the top position on the long-context retrieval benchmarks (2-million-token context window with measured 95+ percent recall at the limit) and on cross-modal video understanding. It is the model that powers Google Antigravity’s Manager view and the upper-tier features in Gemini Spark.
The fourth member, contested at the edges of the set, is xAI’s Grok 4, released in March 2026. Grok 4 clears the capability ceiling on several benchmarks (notably mathematical reasoning and a set of physics-style problems), trails on others (notably long-horizon coding and multimodal video), and is widely understood to have been trained with frontier-scale compute on xAI’s Memphis cluster. Whether it belongs in the frontier set depends on which definition is in use. Under the capability-ceiling definition it sits at the edge. Under the compute-threshold definition it clearly qualifies. Under the EU systemic-risk definition it has been designated by the AI Office as of May 2026.
The fifth and sixth members, also contested, are Meta’s Llama 4 (released March 2026 at 8B, 70B, and 405B parameter sizes) and DeepSeek V4 (released April 2026, the successor to V3 that closed substantial portions of the capability gap to closed-weight frontier models). Both are open-weight models. Under the capability-ceiling definition the 405B Llama 4 and the largest DeepSeek V4 variant arguably belong in the frontier set on several benchmarks where they trade blows with the closed-weight models. Under the compute-threshold definitions, neither cleared the EU 10^25 threshold automatically. The EU AI Office’s stance on whether to designate either of them as systemic-risk under the procedural pathway is an open question as of this writing.
The seventh member, depending on the definition, is Mistral Large 3 (released April 2026 alongside the smaller Mistral Medium 3 and Mistral Small 3). Mistral Large 3 sits near but not at the capability ceiling on most benchmarks, was trained near but not at the frontier-compute threshold, and is offered under both an open-weights release for the smaller models and a commercial license for Large 3. Whether it belongs in the frontier set is a judgment call.
The set churns regularly. Two factors drive the churn. The first is the rapid 40-to-60-day Opus refresh cycle that Anthropic has adopted and that OpenAI is informally tracking, which means the named frontier model from each vendor changes several times per year. The second is the competitive dynamic where each vendor has incentive to argue that their model belongs in the set and that their competitor’s model is no longer at the ceiling. The practical consequence is that any "frontier model" reference in policy, procurement, or architecture should be dated to the time of writing rather than assumed to point at a stable set.
How the term is used in policy
The United States executive order on AI signed in October 2023, and reaffirmed with modification in 2024 and 2026, uses the compute-threshold definition of frontier. Any model trained with 10^26 or more floating-point operations triggers reporting requirements to the Department of Commerce. The threshold was chosen at the time the order was signed as a value comfortably above what any then-deployed model had been trained with, anticipating future scale. By 2026 several models have crossed the threshold and the reporting requirement has been triggered for the major commercial labs. The United States approach is forward-looking on capability but does not require pre-release approval. Reporting flows after training and before deployment.
The European Union AI Act uses two thresholds. The general-purpose AI category captures most models at a relatively low bar. The "general-purpose AI with systemic risk" category captures frontier models at a higher bar of 10^25 floating-point operations of training compute or by explicit designation by the AI Office. Models in the systemic-risk category face additional obligations including model evaluations against safety benchmarks, adversarial testing reports, cybersecurity protections during training, energy use reporting, and incident reporting. The EU approach is procedural and ongoing rather than triggered once at release.
The United Kingdom and Singapore have adopted definitions that closely track the EU approach but apply to fewer obligations. Japan and South Korea have published guidance that defers to the EU and US approaches rather than introducing additional definitions. China’s regulatory approach uses a different vocabulary (the generative AI service regulations focus on the deployed service rather than the model itself), which means cross-jurisdictional model classification has to be reconstructed from facts about training compute and capabilities rather than read off a single global definition.
A practical implication for plugins, themes, and applications that integrate AI models is that whether a particular underlying model is "frontier" or not has regulatory consequences in some jurisdictions. A site operating in the EU that selects a model designated as systemic-risk inherits documentation obligations that the model provider also has. A site operating in the US that selects a model above the compute threshold has fewer direct obligations but is downstream of a reporting flow at the provider. For most builder contexts these obligations are absorbed by the provider rather than the integrating site, but the architectural choice of frontier vs non-frontier is increasingly cross-cutting with compliance.
How the term is used in procurement
Enterprise procurement RFPs in 2026 increasingly use the term "frontier model" to describe a procurement category rather than a specific model. An RFP that asks for "access to a frontier model API" is asking for access to one or more of the current frontier set under a flexible contract that permits the buyer to switch among them as the set changes. Vendor responses typically commit to specific models by name (Opus 4.8, GPT-5.5, Gemini 3 Pro) with a future-substitution clause that lets the vendor swap in the named successor when the named model is succeeded.
Three procurement patterns have settled out. The first is "name a vendor, route to frontier," where the contract names Anthropic, OpenAI, or Google and lets the vendor route to whichever model is current frontier at request time. This is the simplest pattern but exposes the buyer to vendor lock-in and to upstream pricing changes.
The second pattern is "name a tier, route to frontier across vendors," where the contract is for "frontier-tier access" through a gateway like the Vercel AI SDK, LangChain, Microsoft Agent Framework, or an in-house abstraction. The gateway routes to whichever frontier model from whichever vendor is current at request time. This pattern is gaining adoption following the Claude Fable 5 suspension on June 12, 2026, which made the lock-in cost of the single-vendor pattern more visible.
The third pattern is "name a benchmark threshold, qualify against the threshold." The contract specifies a quantitative bar (a score on a public benchmark, or an internal evaluation suite the buyer maintains) and requires the vendor to demonstrate that the selected model clears the bar. This pattern is more operationally complex but is the most insulated against the question of whether a given model is "frontier."
Government procurement, particularly in the United States Department of Defense and the United States intelligence community, has converged on the third pattern. The benchmark thresholds are classified but the procurement pattern is publicly visible in the unclassified RFP text. The pattern that started in DoD procurement has been adopted by several large commercial buyers as a way of getting out of the definitional debate entirely.
How the term is used in architecture
Production AI architectures in 2026 increasingly distinguish between frontier and non-frontier tiers as a deliberate routing decision. The pattern is to route a small fraction of requests (the ones requiring the highest capability for complex reasoning, long-horizon planning, or sensitive content generation) to a frontier model, and to route the much larger volume of requests (classification, summarization, simple chat, repetitive structured generation) to a substantially cheaper non-frontier model. The cost differential between frontier (typically $5-15 per million input tokens, $25-75 per million output tokens for the current set) and non-frontier (typically $0.15-$3 per million input tokens, $0.60-$15 per million output tokens for the strong mid-tier models) is large enough that the routing decision dominates operational cost for most high-volume applications.
The routing decision is typically implemented as a classifier or a lightweight first-pass model that decides which tier a request belongs in. The classifier can be a small model running locally, a heuristic based on prompt characteristics, or a learned model trained on labeled data from the application. The right pattern depends on volume, latency budget, and how stable the request distribution is.
A second architectural pattern is "frontier as escalation." The request first goes to a non-frontier model. If the non-frontier model’s response fails a quality check (low confidence, refused to answer, produced an obviously wrong answer), the request escalates to a frontier model. This pattern is more expensive per request that escalates but cheaper in aggregate when the quality-check rate is low.
A third pattern is "frontier as advisor." The frontier model is used not to handle user-facing requests but to generate templates, refine prompts, or produce few-shot examples that a non-frontier model then uses to handle the user-facing traffic. The frontier model invocations are amortized across many non-frontier invocations, which makes the unit economics far more favorable than per-request frontier serving.
A fourth pattern, gaining adoption after the Claude Fable 5 suspension, is "multi-vendor frontier." The architecture is designed so that any frontier-tier request can be served by any current frontier model from any vendor, with the vendor choice configurable. This is the architectural pattern that the Vercel AI SDK, LangChain, the Microsoft Agent Framework, and the Anthropic-published "model router" pattern all enable. The operational benefit is that an abrupt model availability change (such as Fable 5’s suspension) is a configuration change rather than a code change.
Why the definition matters in practice
The choice among the three working definitions has consequences in three concrete situations. The first is regulatory exposure. A model that crosses a regulatory threshold triggers obligations on its provider and (in some jurisdictions and for some classes of applications) on its integrators. The integrating site’s compliance posture depends on the provider’s classification, which in turn depends on which definition the regulator uses.
The second is procurement scoping. An RFP that asks for "frontier-model access" is making a different commitment under each definition. Under the capability-ceiling definition the supplier is committing to maintain access to whatever the current set is. Under the compute-threshold definition the supplier is committing to a specific quantitative bar. Under the EU systemic-risk definition the supplier is committing to access to models with a specific regulatory designation.
The third is architectural stability. An architecture that routes "frontier" requests to a specific model is exposed to model availability changes (suspensions, deprecations, pricing changes) and to the rapid model-version churn within each vendor’s frontier offering. An architecture that routes "frontier" requests through a vendor-neutral abstraction is insulated from these but has its own engineering cost.
The pragmatic recommendation for builders is to use the capability-ceiling definition operationally, treat the regulatory definitions as compliance overlays that the model provider handles, and design routing architectures that can adapt to the set’s churn rather than depend on the set being stable.
The frontier set will keep changing
The pace of frontier-model releases through the first half of 2026 has been substantially faster than the pace through 2024. Anthropic’s 41-day Opus cycle is the visible upper bound. OpenAI and Google have not formally adopted comparably aggressive cadences but their release cadence has tightened. The implication is that the frontier set composition will continue to churn at a pace that exceeds any plausible regulatory or procurement cycle.
The compositional change is also accelerating. The first wave of frontier models (2022 to 2024) was dominated by OpenAI’s GPT-3.5 and GPT-4 lineage. The second wave (2024 to mid-2025) added Anthropic and Google as full participants. The current wave has added xAI, Mistral, and (under some definitions) Meta and DeepSeek. The diversity of the set is increasing, which makes the definitional question more interesting and the architectural insulation against any single vendor’s availability more valuable.
The capability bar is also rising. The current frontier models substantially exceed what was at the frontier 18 months ago on every public benchmark. The non-frontier tier of 2026 (Claude Haiku, Gemini 3 Flash, GPT-5.5-mini, Mistral Small 3) would have been frontier models in mid-2024 by capability. This is the operational reason that the capability-ceiling definition has to be evaluated as of the time of writing rather than assumed to be stable: the bar moves up under the foot.
Frequently asked questions
Is GPT-4 still a frontier model in 2026? No. GPT-4 was a frontier model in 2023 and into 2024. By the capability-ceiling definition in 2026 it is well below the bar set by Opus 4.8, GPT-5.5, and Gemini 3 Pro. GPT-4 is still a strong production model and remains widely deployed, but the term "frontier" applies to the current ceiling, not historical ones.
Are open-weight models like Llama 4 considered frontier? Under the capability-ceiling definition the 405B Llama 4 arguably qualifies on several benchmarks. Under the EU systemic-risk definition open-weight models have additional exemptions and would only be designated under the procedural pathway, which has not happened as of this writing. The honest answer is that whether Llama 4 is "frontier" depends on which definition is in use.
Does fine-tuning a frontier model produce a frontier model? Usually not, but the fine-tuned model’s regulatory classification can inherit obligations from the base model. The fine-tuned model is not adding capability at the ceiling; it is specializing the base model. The regulatory question is jurisdiction-specific and is best read off the underlying base model’s classification.
Is Claude Code or Cursor a frontier model? No. Claude Code is an agentic coding tool that uses Claude models. Cursor is an IDE that uses several model providers. Neither is a model. The "frontier model" terminology applies to the underlying model, not to the agentic harness wrapped around it.
How is "frontier model" different from "foundation model"? "Foundation model" is a broader category that includes any large-scale general-purpose model trained on broad data and adaptable to many tasks. Most frontier models are foundation models, but not every foundation model is frontier. Foundation model is a 2022-vintage term from the Stanford CRFM. Frontier model is a 2023-vintage term that specifically refers to the capability ceiling at a moment in time.
Does "frontier" imply safety risk? Not directly. The EU AI Act’s "systemic risk" category overlaps with the capability-ceiling set but the overlap is not perfect. A model can be at the capability ceiling and not designated systemic-risk under the EU definition, and vice versa. The terms are correlated but not identical.
Is there a frontier model designed specifically for coding? Several of the current frontier models are particularly strong on coding (Opus 4.8 leads on the long-horizon coding benchmarks; GPT-5.5 leads on the agentic coding benchmarks; Gemini 3 Pro leads on the large-context coding tasks). None of them is exclusively a coding model. Specialized coding models like Codex (built on GPT-5.5) or Claude Code (built on Opus) are agentic harnesses rather than separate frontier models.
Will the frontier set stabilize? Probably not in the near term. The release cadence is faster than the regulatory and procurement cycles, the competitive dynamic favors continued releases, and the capability bar continues to rise. Any policy or architecture that depends on the set being stable will need to be revised periodically. Designing for churn is the right default.
