Current Frontier AI Models (September 2026): GPT-6, Claude, Gemini and Grok Compared

As of September 2026, the frontier AI models are OpenAI's GPT-6 Astra, Anthropic's Claude Opus 5.5 and Claude Fable 5.1, Google's Gemini 3.1 Pro and Gemini 3.8 Flash, xAI's Grok 4.7 and Meta's Muse Spark 1.3. The strongest open-weight models come from Chinese labs: Xiaomi's MiMo-V2.6-Pro, Z.ai's GLM-5.3, Moonshot's Kimi K3, Alibaba's Qwen3.8 and DeepSeek's V4.1-Flash. Independent leaderboards split the top spot, with GPT-6 Astra leading the hardest math, science and reasoning tests and Claude Opus 5.5 leading the Artificial Analysis Intelligence Index and Arena's human-vote ranking.

This page replaces our March 2025 comparison of GPT-4o, Claude 3.7 Sonnet and Gemini 2.0. Specs and prices come from vendor docs, checked on September 27, 2026. If you are looking for the 2024 or 2025 lineup, the timeline after the lab-by-lab section lists the frontier releases since GPT-4o.

Current top frontier AI models compared (September 2026)

Prices are for standard-length prompts; OpenAI charges more above 272K input tokens, Google (3.1 Pro) and xAI above 200K. Gemini 3.8 Flash's price is introductory until December 31, 2026 ($1.50 / $7.50 after). DeepSeek's input price is the cache-miss rate, and both DeepSeek ranges are off-peak / peak.

Latest frontier AI models, lab by lab

OpenAI: GPT-6 Astra, Sol and Luna

OpenAI announced GPT-6 Astra on September 3, 2026 as its most capable model, "built for the hardest end-to-end work". It went to a limited set of organizations first and, over the following days, to ChatGPT Plus, Pro, Business and Enterprise users and the API. Astra is OpenAI's first model at the "Critical" level of cybersecurity capability in its Preparedness Framework, so at launch it refuses advanced security tasks such as creating proof-of-concept exploits. GPT-6 Sol and Luna followed on September 22, Sol for coding and agents and Luna for high-volume work at $0.10 / $0.50. All three take text and images, with a 1.05M-token context window and up to 128K tokens of output (model docs, pricing).

Anthropic: Claude Opus 5.5 and Claude Fable 5.1

Anthropic's current lineup is Fable 5.1, Opus 5.5, Sonnet 5 and Haiku 4.5 (models overview). Claude Opus 5.5, released September 22, 2026, is the one Anthropic recommends starting with. Anthropic says it "performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5", and it lists at $4 / $20 (pricing). Claude Fable 5.1, released September 1, is the pricier option, which the models overview positions for "demanding reasoning and long-horizon agentic work". Both have a 1M-token context window at standard pricing, 128K output and a June 2026 knowledge cutoff. Claude Mythos 5.1, the same model with safeguards built for cybersecurity and life-sciences work, is invitation-only.

Google DeepMind: Gemini 3.1 Pro and Gemini 3.8 Flash

Google's newest Pro model is still Gemini 3.1 Pro, in preview since February 19, 2026. Google said in May that Gemini 3.5 Pro would roll out "next month", and in July that it was still "currently testing with partners". The Flash line moved almost monthly instead, reaching Gemini 3.8 Flash on September 2, and Google's changelog now points new projects to it. Gemini 3.1 Pro shares second place on Humanity's Last Exam and 3.8 Flash is level with GPT-6 Astra on GPQA Diamond, but on the broader Artificial Analysis index they score 30 and 41, below the best open-weight models. Gemini models accept text, images, video, audio and PDFs, with a 1M-token window and 64K output (model page).

xAI: Grok 4.7

xAI, which now brands itself SpaceXAI, released Grok 4.7 on September 21, 2026, about six weeks after Grok 4.6. It targets "coding, agentic tasks, and knowledge work", takes text and images, and has a 500K-token context window (docs). It is on the xAI API, in Cursor and on gateways such as OpenRouter; the model card says the consumer Grok apps get it later.

Meta: Muse Spark 1.3

Meta's frontier model is no longer a Llama. Muse Spark launched on April 8, 2026 from Meta Superintelligence Labs without downloadable weights. Version 1.3 shipped on September 2 with better agentic and coding performance, and it is available through the Meta Model API (model docs). At its max reasoning setting it is the highest-scoring model on the Artificial Analysis index outside Anthropic and OpenAI. Meta says an open-weight Muse Spark release is on its roadmap. For now Meta's newest open-weight language model is the much smaller Muse Glimmer (30B, Apache 2.0); the last large Llama was Llama 4 in April 2025.

DeepSeek and the open-weight challengers

DeepSeek released its V4 family as a preview on April 24, 2026, with MIT-licensed weights. Its API now runs DeepSeek V4.1-Flash, released September 10 with open weights, and V4-Pro is being phased out: since September 14 its requests are served by V4.1-Flash until a V4.1-Pro launches.

Other labs now sit ahead of DeepSeek among open-weight models on the Artificial Analysis index. Xiaomi's MiMo-V2.6-Pro leads that list at 46 points, against 58 for Claude Opus 5.5, followed by Z.ai's GLM-5.3 at 45 and Moonshot's Kimi K3 (July, 2.8 trillion parameters) at 44. Alibaba's Qwen3.8-Max (August) takes images and video in the API, while its open-weight version, Qwen3.8-2.4T-A95B, is text-only and scores level with DeepSeek V4.1-Flash. MiMo and DeepSeek weights are MIT-licensed; Kimi K3, GLM-5.3 and Qwen3.8 use custom licenses that add conditions for model-hosting businesses above a revenue threshold.

Frontier AI models in 2024 and 2025: what changed

Reasoning moved into the flagship. In 2024 OpenAI's o1 was a separate "reasoning model"; GPT-6 Astra and Grok 4.7 now expose a reasoning-effort setting on the main model, and thinking is on by default for Claude 5 models (Anthropic docs).

A 1M-token window became normal. Gemini 1.5 Pro launched in February 2024 with a standard 128K window and 1M only in private preview; every model in the table except Grok 4.7 now takes about a million tokens.

What are frontier models? Frontier AI vs generative AI

"Frontier" is a moving label for the most capable general-purpose models of the moment. The Frontier Model Forum says frontier AI "refers broadly to those general purpose AI models that constitute the state of the art, a collection which will shift over time as the field progresses." The UK government's version, written for the 2023 AI Safety Summit, is "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" (DSIT discussion paper).

Lawmakers count training compute instead. California's SB 53, in force since January 1, 2026, defines a frontier model as a foundation model trained with more than 10²⁶ operations. The EU AI Act presumes that a general-purpose model trained with more than 10²⁵ FLOPs has "high impact capabilities", the trigger for its systemic-risk rules (Article 51).

Generative AI is the wider category: any model that produces text, images, audio, video or code. Every frontier language model is generative AI, but a small open model running on a laptop is generative too and nowhere near the frontier. "Frontier" says where a model sits on the capability curve; "generative" says what it does.

Frontier AI benchmarks: who leads FrontierMath, GPQA, HLE and ARC-AGI

Labs run benchmarks with their own settings, so the numbers here come only from leaderboards the benchmark owners publish (checked September 27, 2026). Claude Opus 5.5 came out on September 22 and is still missing from most of them.

FrontierMath (Epoch AI, not Google DeepMind)

FrontierMath is a set of very hard math problems: Tiers 1-3 run from undergraduate to advanced graduate level, and Tier 4 is research-level mathematics. Epoch AI runs it with funding from OpenAI, and Epoch discloses that OpenAI commissioned the core problems and has access to much of the set, apart from a holdout.

On Epoch's benchmarking hub, GPT-6 Astra leads Tiers 1-3 at 93.7%, with Claude Fable 5.1 second at 90.2%. Astra also leads Tier 4 at 97.6%. That is higher than its Tiers 1-3 score because the private Tier 4 set has only 41 problems, and Astra solved 40 of them. Epoch's best Gemini result on Tiers 1-3 is 3.7 Flash at 71.6%.

Google DeepMind doesn't run FrontierMath, but it has commissioned and published results on it. In October 2025 Google commissioned Epoch to test Gemini 2.5 Deep Think, which set records at the time with 29% on Tiers 1-3 and 10% on Tier 4. In May 2026 a DeepMind paper reported 48% on Tier 4 for its AI co-mathematician, an agent built on Gemini 3.1 Pro that Epoch tested blind. Google's Gemini model pages don't list FrontierMath scores.

GPQA Diamond

GPQA Diamond is 198 multiple-choice questions in biology, physics and chemistry, written by domain experts to be "Google-proof". On Epoch's runs GPT-6 Astra (95.8%), Gemini 3.8 Flash (95.4%) and Gemini 3.7 Flash (94.8%) top the list, but about ten models sit within the leader's margin of error, so the test now mostly confirms a model is top-tier.

SWE-bench

SWE-bench gives a model a codebase and a real GitHub issue to fix. The official SWE-bench Verified leaderboard (500 human-checked tasks) hasn't added a model since February 2026, so its bash-only leader, Claude Opus 4.5 at 76.8%, is several generations old. Scale AI's SWE-Bench Pro board stops at July 2026 and lists none of the September models. For the newest coding models you only have vendor numbers, each run on the vendor's own agent setup.

Humanity's Last Exam

HLE, from the Center for AI Safety and Scale AI, has 2,500 questions from nearly 1,000 subject experts. On Scale's leaderboard, GPT-6 Astra is first at 54.8%. Claude Fable 5.1 (46.5%), Gemini 3.1 Pro (46.4%), Gemini 3.8 Flash (44.5%) and GPT-5.4 Pro (44.3%) share second place, and Scale marks every entry with a potential-contamination warning.

ARC-AGI, Artificial Analysis and Arena

On the ARC Prize leaderboard, where ARC Prize tests models on its own semi-private set, GPT-6 Astra leads ARC-AGI-2 at 95.0%, and Claude Opus 5.5 reaches 93.3% at $0.41 per task against Astra's $1.12. ARC-AGI-2 is close to saturated (the human panel scores 100%). On the newer ARC-AGI-3, Astra leads at 62.7% on the standard harness, and the best Anthropic entry, Claude Opus 5, scores 30.2%.

Two broader rankings put Anthropic first. The Artificial Analysis Intelligence Index (v4.3) scores Opus 5.5 at 58, ahead of Fable 5.1 and GPT-6 Astra at 53. Arena's text leaderboard (formerly LMArena), built from human votes, ranks Opus 5.5 first, within the margin of error of several other Claude models; GPT-6 Astra is 26th.

How to pick a frontier model for your task

• Hard math, science or research questions: GPT-6 Astra leads FrontierMath, HLE and both ARC-AGI boards, and shares the top price with Fable 5.1 ($10 / $50).

• Long coding and agent jobs: Claude Opus 5.5 tops the Artificial Analysis index at 40% of Astra's per-token price, and Anthropic says it suits "long and sprawling jobs like codebase-wide migrations and audits".

• Video, audio or large PDFs: Gemini 3.8 Flash reads them directly and costs $0.75 / $3.75 until the end of 2026.

• Cheap, high-volume work: GPT-6 Luna or DeepSeek V4.1-Flash. Structured extraction, like the receipt parsing in our n8n invoice workflow, rarely needs a frontier model.

• Running the weights yourself: MiMo-V2.6-Pro or DeepSeek V4.1-Flash (both MIT), or GLM-5.3, Kimi K3 and Qwen3.8 if their licenses fit your company.

Because the leaderboards disagree at the top, test two or three candidates on a real task of yours before committing.

Trying several frontier models side by side in BrowseWiz

BrowseWiz is an AI side panel for Chrome and Edge that chats with a model about the page, PDF or selection in front of you. It offers three ways to reach models:

• BrowseWiz Chat, the built-in model, needs no setup and is available on the free plan with limited usage.

• Your own OpenAI API key (Settings > Models) adds the OpenAI models your key can access to the model menu.

• Custom models work with any endpoint that follows the OpenAI chat completions format. You enter a model name, a base URL and an API key.

Most labs in this post document such an endpoint:

Anthropic describes its compatibility layer as intended "to test and compare model capabilities", which is this exact job. Our guide to OpenAI-compatible custom models walks through the setup, OpenRouter included.

To compare, attach the page or file you're working on and ask your question. Then pick another model in the model menu and click Try again on the answer. BrowseWiz regenerates it with the newly selected model and the original context, so every model sees the same input. Each answer is kept, and arrows on the message (1/2, 2/2) switch between them. Do this on the latest answer, because Try again removes any messages that came after it.

FAQ

What are the current frontier AI models?

As of September 2026: GPT-6 Astra (OpenAI), Claude Opus 5.5 and Fable 5.1 (Anthropic), Gemini 3.1 Pro and 3.8 Flash (Google), Grok 4.7 (xAI) and Muse Spark 1.3 (Meta). MiMo-V2.6-Pro, GLM-5.3, Kimi K3, Qwen3.8 and DeepSeek V4.1-Flash lead the open-weight models.

Which AI model is the best in 2026?

No single model wins everywhere. GPT-6 Astra leads FrontierMath, Humanity's Last Exam and ARC-AGI. Claude Opus 5.5 leads the Artificial Analysis Intelligence Index, ranks first on Arena's text leaderboard and costs 40% of Astra's per-token price.

What is the difference between frontier AI and generative AI?

Generative AI is any model that creates text, images, audio, video or code. Frontier AI is the small set of general-purpose models at the state of the art, a set the Frontier Model Forum says "will shift over time". Frontier language models are generative; most generative models aren't frontier.

Is FrontierMath a Google DeepMind benchmark?

No. Epoch AI runs FrontierMath, with funding from OpenAI. Google DeepMind has reported results on it, most recently 48% on Tier 4 for its AI co-mathematician agent in May 2026. GPT-6 Astra now leads both Tiers 1-3 (93.7%) and Tier 4 (97.6%) on Epoch's leaderboard.

What were the frontier AI models in 2024 and 2025?

In 2024: GPT-4o, Claude 3.5 Sonnet, Gemini 1.5 Pro and Llama 3.1 405B, with OpenAI's o1 arriving as a preview in September. At the start of 2025: GPT-4o and o1, Claude 3.5 Sonnet, Gemini 2.0 Flash and DeepSeek-V3, joined by DeepSeek-R1 in January. By December 2025: GPT-5.2, Claude Opus 4.5, Gemini 3 Pro, Grok 4.1 and DeepSeek V3.2.