Mistral Opens Mistral Large 4 Preview, Promises Open Weights in October
Mistral opened a public preview of Mistral Large 4 on October 6, 2026. The French company says the 1 trillion-parameter model's weights will be released by the end of October.
// ai models
AI Model Release Tracker
A dated, sourced timeline of major AI model releases since January 2025. It covers OpenAI, Google, Anthropic, xAI, Meta, DeepSeek, Alibaba and Mistral, with a short note on what each release was.
Researched and fact-checked by AI, with no human review. How we verify
25 entries from 27 sources. Checked for news daily; last checked
What this tracker follows
This tracker lists major general-purpose AI model releases since January 2025 from eight labs: OpenAI, Google, Anthropic, xAI, Meta, DeepSeek, Alibaba (Qwen) and Mistral. It is current as of October 6, 2026.
The selection is editorial. An entry is a text or multimodal model. It is one that its lab presented as a new generation or family, a new top model or a first release to developers. Entries are dated to the day the model became available or, where access was restricted, the day it was announced. Each entry cites the lab's own announcement or documentation, or an established news outlet. Performance and cost statements are the labs' own claims, not independent findings.
The list is capped at 25 entries, with no more than four per lab. As a result, many interim upgrades are left out. The labs' own logs record those. They include the OpenAI changelog, Gemini API release notes, Claude Platform release notes and DeepSeek change log. Image, video, audio and embedding models, coding tools, apps and most smaller variants are also left out.
Some terms recur. A mixture-of-experts model uses only part of its network at each step. That is why labs quote separate total and active parameter counts. Open weights are downloadable model files. Tokens are the units of text, often word fragments, that models read and write. A context window is how many tokens a model can handle at once. The xAI website now carries the name SpaceXAI, as in its Grok 4.7 announcement.
Timeline
Mistral opens public preview of Mistral Large 4
Mistral opened a public preview API for Mistral Large 4. Mistral describes it as a multimodal model with 1 trillion parameters, 52 billion of them active. The company said it would release the model weights by the end of October 2026.
Google announces Gemini 4 Argon with a restricted first rollout
Google announced Gemini 4 Argon, which it calls its new frontier model for coding, enterprise knowledge work and cyber defense. Access began with a set of trusted cyber defenders in its Fairwind Program. Google gave no date for developers, enterprises or consumers.
Source: Gemini 4 Argon: our next era of frontier intelligence, Google
Anthropic releases Claude Opus 5.5
Anthropic released Claude Opus 5.5, the first model in its Claude 5.5 family. Anthropic says it performs on par with Claude Fable 5.1 on most work. It says its own tests show a 40% lower cost than Opus 5 on typical workloads. List prices are $4 and $20 per million input and output tokens.
SpaceXAI releases Grok 4.7
SpaceXAI released Grok 4.7 for coding and knowledge work in Cursor, Grok Build and the Grok API. The company says the model keeps going longer on difficult tasks and is better at verifying its own work. It says the price is the same as Grok 4.6: from $2 and $6 per million input and output tokens.
DeepSeek releases DeepSeek-V4.1-Flash
DeepSeek put V4.1-Flash on its API with built-in image understanding. It describes a 552 billion-parameter mixture-of-experts model whose benchmark results it says are ahead of V4-Pro. It said V4-Pro requests would be routed to the new model from September 14, 2026.
Source: DeepSeek-V4.1-Flash: Smarter, Faster, More Efficient, DeepSeek
OpenAI releases GPT-6 Astra
OpenAI released GPT-6 Astra in its API. The company described it as its most capable model and recommended it for reasoning, coding, computer use, research and document creation. The same changelog shows GPT-6 Sol and GPT-6 Luna following on September 22 and GPT-6.1 Sol on September 29.
Source: Changelog | OpenAI API, OpenAI
Alibaba lists Qwen3.8-Max, then an open-weight Qwen3.8 model
Alibaba Cloud's Model Studio release table lists qwen3.8-max, described as a 2.4 trillion-parameter flagship, on August 2, 2026. An open-source version, Qwen3.8-2.4T-A95B, is listed on August 12 with about 95 billion active parameters and a 1 million-token context.
Source: Alibaba Cloud Model Studio:Model lifecycle and updates, Alibaba Cloud
Meta releases Muse Spark 1.1 and opens the Meta Model API
Meta released Muse Spark 1.1, which it describes as a significant upgrade to Muse Spark. Meta says it is aimed at agent tasks such as tool use, computer use and coding, and has a 1 million-token context window. Developers got access through the new Meta Model API, which opened in public preview.
Source: Introducing Muse Spark 1.1, Meta
Anthropic launches Claude Fable 5 and Claude Mythos 5
Anthropic launched Fable 5 for general use. It also launched Mythos 5, the same model with some safeguards lifted, for Project Glasswing security partners. Both were priced at $10 and $50 per million input and output tokens. Updates on the page say access was suspended on June 12 and restored on July 1, 2026.
Google releases Gemini 3.5 Flash, first of the Gemini 3.5 family
Google introduced the Gemini 3.5 family and released 3.5 Flash first. It was generally available through the Gemini API and Google's enterprise products. It was also open to all users in the Gemini app and AI Mode in Search. Google says it outperforms Gemini 3.1 Pro on coding and agent benchmarks.
Source: Gemini 3.5: frontier intelligence with action, Google
DeepSeek releases DeepSeek-V4 preview (V4-Pro and V4-Flash)
DeepSeek released a preview of V4 with open weights and same-day API access. The preview covered V4-Pro (1.6 trillion total parameters, 49 billion active) and V4-Flash (284 billion total, 13 billion active). DeepSeek said a 1 million-token context became the default across its services.
Meta releases Muse Spark, first model in its Muse family
Meta released Muse Spark, the first model in the Muse family from Meta Superintelligence Labs. It was released on meta.ai and in the Meta AI app, with a private API preview for selected users. Meta describes it as a multimodal reasoning model that supports tool use.
Source: Introducing Muse Spark: Scaling Towards Personal Superintelligence, Meta
OpenAI releases GPT-5.4 and GPT-5.4 Pro
OpenAI released GPT-5.4 in its API, presenting it as a frontier model for professional work with built-in computer use. It released GPT-5.4 Pro alongside it, for harder problems that need more compute. Smaller GPT-5.4 mini and nano versions followed on March 17.
Source: Changelog | OpenAI API, OpenAI
Alibaba releases Qwen3.5
Alibaba's Qwen team dates the Qwen3.5 release to February 16, 2026, starting with a mixture-of-experts model labeled 397B-A17B. Alibaba Cloud's release table lists the hosted model a day earlier. The team's news log shows more sizes, down to 0.8B, on February 24 and March 2.
Mistral releases Mistral 3 (Mistral Large 3 and Ministral 3)
Mistral released Mistral Large 3, a mixture-of-experts model with 41 billion active and 675 billion total parameters. It also released three smaller Ministral 3 models at 3, 8 and 14 billion parameters. All were released under the Apache 2.0 open-source license.
Google releases Gemini 3 Pro in preview
Google released Gemini 3 Pro in preview in the Gemini app. The preview also came to AI Mode in Search for Google AI Pro and Ultra subscribers. Developers got it in AI Studio and Vertex AI. Google launched Antigravity, an agent-based coding platform, the same day.
OpenAI launches GPT-5
OpenAI launched GPT-5, which TechCrunch described as its first unified model, pairing o-series reasoning with fast GPT-style replies through an automatic router. It became the default model for free ChatGPT users the same day.
xAI releases Grok 4 and Grok 4 Heavy
xAI released Grok 4 with built-in tool use and live search to SuperGrok and Premium+ subscribers and through its API. It also introduced Grok 4 Heavy, which weighs several hypotheses in parallel, on a new SuperGrok Heavy tier.
Source: Grok 4, SpaceXAI (xAI)
Anthropic releases Claude Opus 4 and Claude Sonnet 4
Anthropic released Claude Opus 4 and Claude Sonnet 4. Both can answer quickly or reason at length and can use tools while reasoning. Prices stayed at $15/$75 (Opus) and $3/$15 (Sonnet) per million input/output tokens. Claude Code became generally available.
Source: Introducing Claude 4, Anthropic
Mistral releases Mistral Medium 3
Mistral released Medium 3 on its API and Amazon SageMaker at $0.40 per million input tokens and $2 per million output tokens. The company claims it reaches at least 90% of Claude Sonnet 3.7's benchmark results at a much lower price.
Alibaba releases Qwen3 as open weights
Alibaba's Qwen team released Qwen3 as open weights. The release had two mixture-of-experts models, the larger with 235 billion parameters, and six smaller dense models. The team says the models can switch between step-by-step thinking and fast replies. It says they support 119 languages and dialects.
Source: Qwen3: Think Deeper, Act Faster, Qwen team, Alibaba Group
Meta releases Llama 4 Scout and Llama 4 Maverick
Meta released Llama 4 Scout and Llama 4 Maverick for download on llama.com and Hugging Face. Meta says they are its first models built on a mixture-of-experts design. A larger model, Llama 4 Behemoth, was previewed while still in training.
Source: The Llama 4 herd: The beginning of a new era of natively multimodal AI innovation, Meta
Google introduces Gemini 2.5 Pro Experimental
Google introduced Gemini 2.5 with an experimental version of 2.5 Pro. Google calls it a thinking model because it works through a problem before answering. It was available in Google AI Studio and to Gemini Advanced subscribers in the Gemini app. Vertex AI was to follow.
xAI releases Grok 3
xAI released Grok 3 late on February 17, 2025, in a livestreamed launch. It also released a smaller Grok 3 mini and a research feature called DeepSearch. TechCrunch reported that some features were still in beta and that X Premium+ subscribers got access first.
Source: Elon Musk’s xAI releases its latest flagship model, Grok 3, TechCrunch
DeepSeek releases DeepSeek-R1
DeepSeek released its R1 reasoning model with code and weights under the MIT license. It offered the model in its API as deepseek-reasoner. DeepSeek claimed performance on par with OpenAI's o1. It released six smaller distilled models alongside it.
Source: DeepSeek-R1 Release, DeepSeek
Additional sources
- Release notes | Gemini API | Google AI for Developers, Google
- Claude Platform release notes, Anthropic
- Change Log | DeepSeek API Docs, DeepSeek
Update history
- Updated the Mistral Large 4 entry to 52 billion active parameters, the figure Mistral's announcement now gives. The announcement gave 49 billion when it was published on October 6.
- Rewritten in shorter, plainer sentences. No facts were changed.
Articles on AI Models
ChatGPT, Gemini, Claude and Grok: How the Leading AI Models Differ
ChatGPT, Gemini, Claude and Grok differ on price, access and specialty, and no one model leads every test. Current versions, costs and benchmark claims as of October 2026.
Open-Weight vs Closed AI Models: What the Difference Means for You
Open-weight models can be downloaded and run on your own hardware, while closed models stay on their maker's servers. Open-weight is not the same as open source, and licenses vary.