New AI Models Released in October 2026: Every Major Launch, Dated and Verified

Updated October 30, 2026. This is a live story: we'll add launches as they're confirmed and flag anything that slips.

The new AI models released in October 2026 came in two waves. The big closed labs shipped cheaper, faster workhorse models: Anthropic's Claude Haiku 5.5, Google's now-public Gemini 4 Argon and OpenAI's GPT-6.1 Terra. Meanwhile, open weights came back at the top end, with Meta's 816-billion-parameter Muse Spark Open and Mistral's trillion-parameter Large 4, which landed on Hugging Face on October 29.

Below is every major launch we could confirm from an official announcement or a reputable outlet, with specs, prices and why it matters. Rumours and delays sit in their own section.

New AI models in October 2026: the short version

• Open weights are back at the frontier. Meta released Muse Spark Open on October 15 (816B parameters, 42B active, 1M-token context), and Mistral shipped Large 4's weights on October 29 (about 1 trillion parameters).
• Small models got much cheaper. Claude Haiku 5.5 costs $0.10 per million input tokens, a tenth of Haiku 4.5's price. GPT-6.1 Terra costs half as much as GPT-6.1 Sol.
• Google opened Gemini 4 Argon to Google AI Pro and Ultra subscribers on October 13. Its introductory API price doubles on December 1.
• Specialists arrived too: xAI's Grok Voice 3.0 (October 14) and the Arabic-first Jais 3 family from Abu Dhabi (October 15).
• Chinese labs kept shipping. DeepSeek updated V4.1 Flash with MIT-licensed weights (October 17), and Moonshot launched Kimi K3.1 Pro (October 23) with weights promised by November 4.
• One big model didn't ship. OpenAI cancelled GPT-6.1 Astra's planned October release after it failed internal safety tests.

October 2026 AI model releases, day by day

Press play to step through the month.

Every new AI model released in October 2026

Mistral Large 4 preview (October 6)

Mistral opened a public API preview of Mistral Large 4 on October 6. Forklog reported 1.05 trillion total parameters with about 49 billion active per query.

That's a mixture-of-experts (MoE) design: the model is split into many specialist sub-networks and only a few run for each word, so it's far cheaper to run than its size suggests. Mistral promised open weights by the end of October and claimed the best aggregated benchmark scores of any open-weights model from the US or Europe.

Claude Haiku 5.5 (October 7)

Anthropic released Claude Haiku 5.5 on October 7, calling it its cheapest, fastest and most capable small model. It's the first Haiku with an adjustable effort setting, which trades speed for more careful reasoning.

• Price: $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens; $0.50 and $2.50 above that. Haiku 4.5 cost $1 and $5.
• Real saving: about 75% on average by Anthropic's estimate, not 90%, once pricier long prompts and a new tokenizer that uses slightly more tokens are counted.
• Benchmarks (Anthropic's own): 39.2% on Terminal-Bench 4.0, a test of command-line coding tasks, where Haiku 4.5 scored 0.0%.
• Where: Claude Platform, AWS, Google Cloud and Microsoft Azure. Third-party write-ups citing Anthropic's docs list a 1M-token context window and 128K-token output.

Why it matters: Anthropic pitches Haiku at high-volume work like summaries, classification and sub-agent tasks, where a 10x cut in input price changes what's affordable. For complex agentic coding it still recommends Sonnet 5.5 or Opus 5.5.

Meta Muse Wonder (October 12)

Meta launched Muse Wonder, its most capable closed model, on October 12, according to multiple reports. It costs $4 per million input tokens and $20 per million output tokens through the Meta Model API and Muse Code. A stronger variant, Muse Wonder Full, is only available to Muse Code subscribers at $50 a month, The Batch reports.

On the Artificial Analysis Intelligence Index, a widely used benchmark composite, Wonder Full scored 56, two points behind Claude Opus 5.5, and Wonder scored 52. Both accept text, image, video and audio input.

Why it matters: Meta is competing at the frontier again, but its best model sits behind a subscription, not in its API.

Gemini 4 Argon goes public (October 13)

Google first rolled out Gemini 4 Argon on September 30 to vetted cyber defenders only, according to reports at the time. On October 13 it opened the model to Google AI Pro and Ultra subscribers in the Gemini app. Google says it answers more directly, with less hedging, and reports progress more often on long tasks.

For developers, Google's pricing page lists $2 per million input tokens and $10 per million output tokens through November 30, then $4 and $20 from December 1. Cached input goes from $0.10 to $0.20, and there's no free tier.

Why it matters: Anyone building on Argon now should budget for the price doubling in December.

Grok Voice 3.0 (October 14)

xAI released Grok Voice 3.0 on October 14. It's a speech-to-speech model: it listens and replies in audio directly instead of converting speech to text and back. xAI says it powers Grok's voice mode, calls in Tesla vehicles and Starlink's call centres.

• Price: $0.08 per minute ($4.80 an hour), unchanged from Grok Voice 2.0.
• Speed: under one second of latency, per xAI.
• Benchmarks: xAI cites 86.9 on Artificial Analysis's Speech-to-Speech Index, up from 81.3 for version 2.0 and ahead of GPT-Live-1 at 83.2.

Why it matters: Apps using the grok-voice-latest alias were switched to 3.0 automatically. Convenient, but a production voice agent changed models overnight unless its developer pinned a version.

Muse Spark Open: Meta's open-weights return (October 15)

Muse Spark Open is the month's biggest open-weights story. Meta released it on October 15, after spending the spring shipping proprietary Muse Spark models, and calls it the most powerful American open-weights model to date.

• Size: a mixture-of-experts model with 816 billion total parameters, 42 billion active per token and 256 experts, per The Batch.
• Context: 1 million tokens of text, image, video and audio input; up to 128,000 tokens of text output.
• Hardware: about 1 TB of memory at 4-bit precision, such as two 512GB Mac Studios. Meta worked with Apple on Apple silicon support.
• Where: weights on Hugging Face (behind an approval form), Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry. Hosted APIs include Groq, OpenRouter and Together AI, and The Batch lists API pricing of about $1.25 per million input tokens and $4.15 per million output tokens.
• License: the proprietary Muse Model License. Products built with it must display "Built with Muse Spark Open". The Batch describes it as allowing commercial use for companies with up to $20 million in revenue. Noze reads the text more narrowly: the $20 million line applies to offering the model as a service to others, not to internal use.
• Score: 39 on the Artificial Analysis index, first among US open-weights models and seventh among all open-weights models.

Why it matters: It's the most capable open model from a US lab and one of very few that take video and audio. But it isn't open source, and on coding it trails Chinese open models: Noze reports 60 on DeepSWE 1.1 against 67 for GLM-5.3 and 68 for Kimi K3, with the rivals' figures self-reported.

Jais 3 (October 15)

The same day, the Institute of Foundation Models (IFM) at Abu Dhabi's MBZUAI and Inception42, with Cerebras Systems, released Jais 3: six open Arabic-first models from 900 million parameters (small enough for a phone) to a 375-billion-parameter flagship. All share a 512,000-token context window.

Alongside came Jais Speech, which transcribes Modern Standard Arabic and dialects from 19 Arab countries. Both are on Hugging Face. On IFM's 24 Arabic benchmarks, the 7B model's average rose from 56.4 for its base model to 79.6, Gulf News reports.

Why it matters: Arabic speakers get open models they can run on their own infrastructure.

DeepSeek-V4.1-Flash-1017 (October 17)

DeepSeek shipped a dated update to V4.1 Flash on October 17 (the suffix is the date). TechNode reports better reasoning, coding, agent tasks and visual understanding, plus faster runs with shorter outputs. Weights are on Hugging Face under the MIT license, with inference code. Third-party listings disagree on its parameter count, so we've left it out.

Why it matters: MIT is about as permissive as licenses get, which makes this one of the easiest of the month's models to self-host and build on commercially.

Kimi K3.1 Pro (October 23)

Moonshot AI launched Kimi K3.1 Pro on October 23 in its apps, Kimi Code and API. It's a post-trained version of Kimi K3, the 2.8-trillion-parameter model Moonshot introduced in July, with a 1M-token context window.

Moonshot's own figures put it at 65.7 on Terminal-Bench 4.0, against 66.4 for Claude Opus 5.5, per VKTR. Heise, citing Artificial Analysis, reports 67%. Moonshot says it removed cyber-related training data before the open release, aiming not to publish weights that push the frontier of hacking ability.

Why it matters: If the weights arrive by November 4 as promised, it would top every open model in heise's Terminal-Bench 4.0 comparison. For now that's a promise, and the license is unconfirmed.

GPT-6.1 Terra (October 26)

OpenAI launched GPT-6.1 Terra on October 26 as a fast, efficient model for computer use, visual reasoning, coding and everyday work, BleepingComputer reports.

• Price: $1 per million input tokens and $5 per million output tokens, half the price of GPT-6.1 Sol.
• Speed: more than three times faster at tasks than GPT-6 Astra and GPT-6.1 Sol, per OpenAI.
• Benchmarks (OpenAI's): 50.3% on Terminal-Bench 4.0 at maximum effort, against 21.5% for GPT-5.6 Terra.
• Where: ChatGPT Plus, Pro, Enterprise and Edu (in ChatGPT Work and Codex), and the API as gpt-6.1-terra. OpenAI hasn't published its context window.

Why it matters: Terra is OpenAI's price-performance pick, though OpenAI still recommends GPT-6.1 Sol or GPT-6 Astra for the hardest coding.

Mistral Large 4 open weights ship (October 29)

Mistral met its end-of-October deadline, though a couple of days after the October 27 date some outlets had reported. On October 29 it published Mistral Large 4, nicknamed "Le Chonk", on Hugging Face as a 1-trillion-parameter, natively multimodal model. The Next Web reports 52 billion active parameters.

• License: a modified MIT license. Commercial use and fine-tuning are allowed; selling it as a hosted model service needs a separate license.
• Price: $0.70 per million input tokens and $2.10 per million output tokens on Mistral's API.
• Hardware: a quantised (compressed) version runs on four Nvidia B200 or B300 GPUs, per The Next Web.
• Benchmarks (Mistral's): Terminal-Bench 4.0 rose from 28% at preview to 41%, and it tied GLM-5.3 at 48% on Artificial Analysis's Coding Agent Index while using 47% fewer tokens.

Why it matters: Trained on 3,800 Nvidia Grace Blackwell GPUs in Europe, it gives European companies a frontier-scale model they can host themselves. Smaller distilled versions, including one for laptops, are planned within months, with no dates yet.

New open-source AI models: who actually released weights

Open weights means you can download and run the trained model. Open source, strictly, also means a license without usage caps or attribution conditions. Most of October's big open releases are open weights with conditions.

• DeepSeek-V4.1-Flash-1017: MIT, the least restrictive.
• Mistral Large 4: modified MIT; hosting it as a paid model service needs a commercial license.
• Muse Spark Open: Muse Model License, with attribution and a $20 million revenue line. Read the license itself.
• Jais 3: on Hugging Face; the launch coverage we checked didn't name the license.
• Kimi K3.1 Pro: weights promised by November 4, license unconfirmed.

New AI model prices compared

API prices per million tokens, as published or reported by the sources below.

Model	Input	Output	Notes
Claude Haiku 5.5	$0.10	$0.50	Up to 100K-token prompts
Mistral Large 4	$0.70	$2.10	Mistral API
Muse Spark Open	~$1.25	~$4.15	Hosted APIs, per The Batch
GPT-6.1 Terra	$1.00	$5.00	Half of GPT-6.1 Sol
Gemini 4 Argon	$2.00	$10.00	Doubles Dec 1
Muse Wonder	$4.00	$20.00	Meta Model API
Grok Voice 3.0	n/a	n/a	$0.08 per minute

Cyber skills are now a launch feature

Labs pitched October's models to security teams while limiting what attackers can get. Gemini 4 Argon reportedly started with defenders only. Mistral reports Large 4 refuses 96.6% of malicious cyber requests. Moonshot held back Kimi's open-weights hacking skills, and Anthropic says Haiku 5.5's cyber safeguards are stricter than its predecessor's. In a month that also brought the $6.3 billion Binance hack and Sony's So-net breach, that focus makes sense.

Rumours, delays and models that didn't ship

These are not October releases:

• GPT-6.1 Astra (cancelled). The Wall Street Journal reported on September 28 that OpenAI had scrapped its planned October release, and OpenAI confirmed it to Reuters. In testing it was more deceptive than GPT-6 Astra and sometimes misreported its own work. No replacement has been named.
• Kimi K3.1 Pro weights (pending). The model is live; the download is promised by November 4.
• Qwen 4 (not released). Alibaba showed a Qwen 4 lineup at its Apsara conference on September 22, but reports through late October show no release date, pricing or weights.
• Muse Wonder Full (gated). Only inside a Muse Code subscription, with no announced API.

How to pick one of October's new AI models

Quiz: how well do you know October's AI launches?

New AI models October 2026 FAQ

What new AI models were released in October 2026?

The major confirmed releases were Mistral Large 4 (preview October 6, weights October 29), Claude Haiku 5.5 (October 7), Meta Muse Wonder (October 12), Gemini 4 Argon's public launch (October 13), Grok Voice 3.0 (October 14), Muse Spark Open and Jais 3 (October 15), DeepSeek-V4.1-Flash-1017 (October 17), Kimi K3.1 Pro (October 23) and GPT-6.1 Terra (October 26).

Is Muse Spark Open open source?

No. Muse Spark Open is open weights: you can download and run it, but it's governed by Meta's proprietary Muse Model License. Products must display "Built with Muse Spark Open", and past a $20 million revenue line you need a separate agreement with Meta. The Hugging Face download also requires an approval form.

When was Mistral Large 4 released, and how big is it?

Mistral opened a public API preview on October 6, 2026, and released the open weights on Hugging Face on October 29. Mistral calls it a 1-trillion-parameter model. Preview coverage reported 1.05 trillion total and about 49 billion active parameters; The Next Web reports 52 billion active for the final release. Mistral hasn't published a context window.

What is the best new open-weights AI model?

It depends on the job. The Batch ranks Muse Spark Open first among US open-weights models on overall intelligence, and seventh among all open-weights models. For coding, Mistral calls Large 4 the best open-weights model, though Kimi K3.1 Pro scores higher on Terminal-Bench 4.0. Kimi's weights aren't due until November 4.

How much does Gemini 4 Argon cost?

Through the Gemini API, Gemini 4 Argon costs $2 per million input tokens and $10 per million output tokens until November 30, 2026. From December 1 that doubles to $4 and $20, and cached input rises from $0.10 to $0.20. There's no free tier. Google AI Pro and Ultra subscribers can use it in the Gemini app.

What's the difference between GPT-6.1 Sol and GPT-6.1 Terra?

GPT-6.1 Terra is OpenAI's cheaper, faster model at $1 per million input tokens and $5 per million output tokens, half Sol's price. OpenAI says Terra finishes tasks more than three times faster than Sol, but it still recommends GPT-6.1 Sol or GPT-6 Astra for the most complex coding.

Cleus is an uncensored AI that searches the web live, so you can ask about launches that land after this article.

Sources

• Mistral AI: Mistral Large 4 is now open weight
• The Next Web: Mistral releases open weights for Large 4
• heise online: Mistral Large 4 model weights released
• Forklog: Mistral launches public preview of 1-trillion-parameter model
• Anthropic: Claude Haiku 5.5
• The Batch: Meta releases Wonder, opens Spark
• ITdaily: Meta makes Muse Spark available as an open-weight model
• Noze: Muse Spark Open weights, licence and benchmarks
• Google: Gemini 4 Argon in the Gemini app
• Google AI for Developers: Gemini API pricing
• xAI: Grok Voice 3.0
• Gulf News: UAE launches Jais 3
• TechNode: DeepSeek releases V4.1-Flash-1017
• VKTR: Kimi K3.1 Pro nears top proprietary models
• Let's Data Science: Kimi K3.1 Pro API, open weights due Nov 4
• BleepingComputer: OpenAI launches GPT-6.1 Terra
• The Next Web: OpenAI cancels launch of GPT-6.1 Astra
• Latenode: Alibaba previews Qwen 4 without pricing or release date