Mistral Large 4 entered public preview on October 6, 2026 as a roughly 1T total, 52B active MoE. Mistral has promised open weights by the end of October 2026 but has not named a license, so for now it is an API model only, on sale at $0.68/$2.09 against a $1.36/$4.18 list price.

DeepSeek describes V4.1 Flash, released September 10, 2026, as the smallest model in its new architecture family, which points to larger siblings. No names or dates have been announced.

Mythos 5.1 shipped in September 2026 at the same $10/$50 list price as Claude Fable 5.1, but Anthropic lists it as limited availability through Project Glasswing. Most developers can buy Fable 5.1 today and not Mythos 5.1.

Microsoft said on October 15, 2026 that MAI-Base-2, a 2.8T total, 80B active model, has finished pre-training. It has not been released, and no price, date, or license has been published.

Alibaba announced Qwen 4 in Max, Flash, Plus, and 27B versions as in training at its Apsara conference on September 22, 2026. Nothing has been released yet.

Anthropic

Anthropic has leaned hard into agentic coding as the wedge. Claude has become the default model for serious developer tools, which creates a revenue base that funds frontier training. The company pairs this with the most aggressive public stance on safety of any major lab. Responsible scaling commitments, detailed model cards, and constitutional AI research are all part of a single story: if you believe the most capable models are coming soon, your commercial strategy should be inseparable from your safety strategy. Claude Opus 5.5, released September 22, 2026 at $4/$20, is now the default model in Claude Code and first on the Artificial Analysis Intelligence Index and Vals Terminal-Bench 4.0, and Anthropic shipped it with the same class of safeguards as Fable 5.1. Sonnet 5.5 followed on September 28 at $2/$10 and sits second on both Vals boards, and Haiku 5.5 arrived October 7 at $0.10/$0.50 for prompts up to 100K tokens.

OpenAI

OpenAI still owns the largest consumer surface area in AI. ChatGPT is the default chatbot for a huge fraction of the market, which creates data, revenue, and distribution. The o-series reasoning models were the first public bet on test time compute as a primary capability lever, and the results shifted how every other lab thinks about reasoning. Expect continued emphasis on multimodal, voice, and vertical integration through Codex and ChatGPT Work, with GPT-6 Astra now available on paid ChatGPT plans. On September 22, 2026 OpenAI filled out the GPT-6 generation with Sol at $2/$10 and Luna at $0.10/$0.50, and said Astra remains its best model across the board. GPT-6.1 Sol followed on September 29 at the same $2/$10, and on October 14 OpenAI set a January 14, 2027 shutdown for the GPT-4o and GPT-4.1 snapshots.

Google DeepMind

Google has structural advantages no one else has. Custom TPU infrastructure, a search index, YouTube, and decades of research depth at DeepMind. Gemini has closed most of the capability gap and owns the long context and multimodal categories. The real leverage is distribution: Gemini is shipping into every Google surface, from Search to Workspace to Android. Every Google user becomes a Gemini user by default. Gemini 4 Argon became generally available on October 13, 2026 and leads the Vals Index; its introductory $2/$10 rate runs through November 30, then rises to $4/$20.

xAI

xAI has compressed an enormous amount of capability into a short timeline. The Memphis training cluster came online with unusual speed, Grok 3 closed most of the gap to the frontier, and the Grok 4 line has stayed close since: Grok 4.6 posts 94.9 on GPQA Diamond in Artificial Analysis testing. Grok 4.7 moved to a new, larger base model on September 21, 2026 at the same $2/$6 rate and scores 46 on the Artificial Analysis Intelligence Index, twelve points behind the leader. Distribution through X gives xAI a feedback loop that other labs do not have. The open question is how long the pace can be sustained.

Meta

Meta did more than anyone to put open weight models near the frontier. The Llama family forced every other lab to compete on value, not just capability, and Mark Zuckerberg framed open weights as a strategic asset, not a charity move. The strategy has since split: Meta sells Muse Spark 1.3 and the new Muse Wonder ($4/$20, October 12) through its paid API, while releasing open weights on its own terms. Muse Glimmer 30B ships under Apache 2.0, but Muse Spark Open, the 816B total, 42B active model released October 15, uses a custom Muse Model License with an access form and extra terms for model-as-a-service businesses above $20 million in revenue.

Which AI model is the best in 2026?

There is no single best model. On the independent aggregate indexes, Gemini 4 Argon leads the Vals Index (68.90 percent on the October 18, 2026 board, ahead of Claude Sonnet 5.5 at 67.04 and Claude Opus 5.5 at 66.97), while Claude Opus 5.5 leads the Artificial Analysis Intelligence Index (58) and Vals Terminal-Bench 4.0 (65.15). In our benchmark data, Claude Opus 5 (97.0) and GPT-5.6 Sol (96.2) top SWE-bench Verified, Claude Opus 5 leads OSWorld 2.0 computer use (70.6), GPT-6 Astra leads GPQA Diamond (96.0) and BrowseComp (91.5), and Claude Opus 5.5 leads Humanity's Last Exam with tools (67.7) and Cognition's FrontierCode (54.6). On price, Gemini 3.8 Flash posts 95.3 on GPQA Diamond at $0.75/$3.75 per 1M tokens. The right answer depends on the workload. See our benchmark leaderboard for category winners.

Who is winning the frontier AI race?

As of October 2026, nobody holds every top spot. Gemini 4 Argon, generally available since October 13, leads the Vals Index. Claude Opus 5.5, released September 22 at $4/$20, leads the Artificial Analysis Intelligence Index and Vals Terminal-Bench 4.0, and Claude Sonnet 5.5 sits second on both Vals boards at half the Opus 5.5 price. GPT-6 Astra remains OpenAI's top model, with GPT-6 Sol, GPT-6.1 Sol, and Luna filling out the cheaper tiers. Meta is now a frontier contender with Muse Wonder, whose subscriber-only Full version ties Sonnet 5.5 for second on the Artificial Analysis index at 56, and xAI keeps pace with Grok 4.7. The open weight race is still led mostly by Chinese labs, with Moonshot's Kimi K3, DeepSeek V4.1 Flash, Alibaba's Qwen3.8 family, and Xiaomi's MiMo-V2.6, whose Pro model posted the top open-weights score on the Artificial Analysis index (46) as of late September. October added large Western open releases too: Meta's Muse Spark Open under a custom license and Reflection's Beam under Apache 2.0.

How often do new frontier models ship?

Frontier releases now land every four to eight weeks on average. Anthropic, OpenAI, and Google each ship a major update every quarter, interspersed with point releases. Open weight launches from Meta, Mistral, and DeepSeek add to the cadence.

What is a frontier model?

A frontier model is one of the most capable AI systems publicly available at a given moment, typically defined by benchmark performance, compute used during training, and agentic capability. The frontier moves constantly as new releases ship.

Which frontier model is cheapest?

Among the big three labs, GPT-6 Luna ($0.10/$0.50 per 1M tokens), Gemini 3.8 Flash ($0.75/$3.75 through December 31, 2026), and Claude Haiku 5.5 ($0.10/$0.50 for prompts up to 100K tokens) offer the lowest prices per token. DeepSeek V4.1 Flash costs $0.30/$1.20 at peak on its API, Xiaomi's MIT-licensed MiMo-V2.6-Flash is $0.14/$0.28, and open weights like DeepSeek V4.1 Flash and Qwen3.8-27B are close to free when you self host.