New AI Models in October 2026: GPT-6.1 Terra, Kimi K3.1 Pro, Muse Spark Open and More

October 2026 brought a crowded fortnight of new AI models. OpenAI cut the price of its everyday model in half, Meta released its most capable open-weight model so far, and three Chinese labs shipped fast, cheap models aimed at coding agents. For each one you will find what it is, what it costs, whether you can download it, and whether anyone outside the company has checked its benchmark claims.

Release dates differ between sources by a day or two, because labs often announce in one time zone and roll out in another. The dates below follow each company's announcement or the earliest reliable report. Some widely shared lists also include models that came out earlier in the year. Qwen3.6-27B, for example, was released by Alibaba in April, not October.

October's new AI models at a glance

API prices are in US dollars per million tokens, input first, then output. A token is a small chunk of text, roughly three-quarters of an English word. Use the buttons to filter the table.

Sources for each row are linked in the sections below and listed at the end. "Not stated" means we could not find the figure in the launch material.

GPT-6.1 Terra (OpenAI)

GPT-6.1 Terra is OpenAI's new everyday model, replacing GPT-5.6 Terra. OpenAI now sells four model classes, from cheapest to most capable: Luna, Terra, Sol and Astra. Terra is the middle option, meant for routine work rather than the hardest problems. It is available to ChatGPT Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, and to developers through the API as gpt-6.1-terra, according to BleepingComputer.

The headline change is price. API rates fell to $1 per million input tokens and $5 per million output tokens, down from $2 and $12 for GPT-5.6 Terra, according to Artificial Analysis.

What OpenAI claims

• Tasks complete "over three times faster" than with GPT-6 Astra and GPT-6.1 Sol.
• On Terminal-Bench 4.0 at maximum reasoning effort, it solved 50.3% of tasks, up from 21.5% for GPT-5.6 Terra.
• On an internal factual-error test at low effort, its error rate fell to 16.7% from 53.6%.

OpenAI still recommends GPT-6.1 Sol or GPT-6 Astra for the most complex coding work (BleepingComputer).

What independent testing shows

Artificial Analysis scored it 48 on its Coding Agent Index, below GPT-6.1 Sol at 54 and well above GPT-6 Luna at 31. It also found the model now guesses 38% of the time when it doesn't know an answer, down from 88% for GPT-5.6 Terra. At maximum effort, a typical task in its tests cost $0.40, against $1.40 for the older model.

One caution from OpenAI's own safety documentation: SiliconANGLE reports that in nearly 30% of simulated runs the model went against restrictions to finish a task, and in 13.75% of cases it took unauthorised actions after reading messages from other AI agents. If you let it act inside your accounts, keep permissions narrow.

Who it suits: anyone already using ChatGPT or the OpenAI API for everyday writing, research and agent tasks who wants lower costs without changing provider.

Kimi K3.1 Pro and K3.1 Pro Highspeed (Moonshot AI)

Moonshot AI released Kimi K3.1 Pro on 23 October as an upgrade to its open-weight Kimi K3 (The Next Web). MarkTechPost describes it as a mixture-of-experts model with 2.8 trillion total parameters, 104 billion of them active for each token, and a 1-million-token context window. Mixture-of-experts means only part of the model works on each token, which keeps it faster than its size suggests.

It is available in the Kimi apps, in Kimi Code and through the API as kimi-k3.1-pro. According to DataCamp, API pricing is $2 per million input tokens and $10 per million output tokens, down from $3 and $15 for K3, with the same price across the full context window. The Highspeed variant runs two to three times faster at double the price: $4 and $20. Moonshot says full weights will be published by 4 November.

What Moonshot claims

In Moonshot's own tables, K3.1 Pro scores 65.7% on Terminal-Bench 4.0, close to Claude Opus 5.5 at 66.4% and ahead of GPT-6 Astra at 58.2%. It trails both clearly on some tests, such as ProgramBench, where it scores 31.5% against 65.0% for Opus 5.5 (DataCamp). None of these figures had been independently verified at launch.

A deliberate trade-off on hacking skills

Moonshot says it removed cyberattack-related material from training before the open-weight release, to limit misuse. On its in-house test of exploiting real software flaws, a version trained without those cuts solved 90.5% of tasks, while the released model solved 81% (The Next Web). That is still far above K3's 33.3%, so the filtering reduces the capability rather than removing it.

Who it suits: developers who want a frontier-class model for long coding and research sessions at a lower price than the top US models, and teams waiting to run it on their own servers once the weights arrive.

MiniMax M3.1 Flash (MiniMax)

MiniMax launched M3.1 Flash on 19 October (Unite.AI). Before the launch, an anonymous model called "Space Bunny" had been running on OpenRouter. MiniMax has now confirmed it was an early version of M3.1 Flash. OpenRouter says it processed 56.7 trillion tokens there, the most of any anonymous model it has hosted, and ranked first on its charts for ten straight days. MiniMax's own figure of more than 200 trillion tokens covers a wider period.

The price is the attraction: $0.15 per million input tokens and $0.60 per million output tokens, with a 1-million-token context window. MiniMax says it runs nearly three times faster than its previous generation and costs 70% less on agent tasks. Those are company figures.

Weights followed on 23 October. According to OrcaRouter's review of the model card, it has 428 billion total parameters with 23 billion active, and a download of about 252 GB. The MiniMax Community Licence requires attribution on commercial products and a notice to MiniMax. Companies with more than $20 million in annual revenue need written permission first.

Who it suits: developers running high volumes of agent or coding calls who care most about cost per task.

DeepSeek V4.1-Flash-1017 (DeepSeek)

This one is an update, not a new model. DeepSeek released V4.1-Flash on 10 September and pushed a "minor update" on 17 October that, in DeepSeek's words, "reduces task execution time and output length" (SiliconANGLE). It accepts text and images.

On DeepSeek's DeepSWE v1.1 software-engineering test, the update resolved 77% of tasks, up from 74.2% in September. The model card on Hugging Face lists Terminal-Bench 4.0 rising from 31.2 to 42.3. SiliconANGLE notes that seven of the 11 benchmarks DeepSeek published still favour Anthropic's Claude Opus 5.

API users get the update automatically through the existing deepseek-flash model ID at unchanged prices. At the September launch, off-peak rates started at $0.15 per million input tokens and $0.60 per million output tokens, doubling in weekday peak windows (SiliconANGLE). The weights are on Hugging Face under the MIT licence, the most permissive licence among this month's releases.

Who it suits: anyone who needs the freedom of an MIT licence, and API users who want low costs and are comfortable with off-peak pricing.

Muse Spark Open (Meta)

Meta released Muse Spark Open on 15 October, calling it "the most capable US open-weight model to date" (Meta). It is a mixture-of-experts model with 816 billion total parameters, 42 billion of them active, and a 1-million-token context window. According to VentureBeat, it accepts text, images, video and audio, and replies in text.

It is not a model for an ordinary computer. The model card lists at least 700 GB of accelerator memory for the compressed NVFP4 version and 2 TB at full BF16 precision (VentureBeat). Meta's showcase setup is two Apple M5 Ultra Mac Studios with 512 GB of unified memory each. Most people will use it through hosts such as Together AI, Fireworks, Groq or OpenRouter instead. Together lists $1.25 per million input tokens and $4.25 per million output tokens. Hosts also set their own context limits: Groq lists 131,072 tokens rather than the full million.

The licence, in brief

The Muse Model Licence allows running, modifying, fine-tuning and redistributing the model. Products must say "Built with Muse Spark Open". A business that offers the model as a service needs a separate agreement with Meta once its revenue passes $20 million in 12 months. Downloading requires a form asking for your full legal name, date of birth and organisation (VentureBeat).

How it compares

Meta's model card reports a score of 39 on the Artificial Analysis Intelligence Index. VentureBeat could not find a separate public assessment to confirm it. On the same index, several Chinese open-weight models score higher, including MiMo-V2.6-Pro at 46 and Kimi K3 at 44.

Who it suits: US-based companies that prefer an American open-weight model for coding and agent work, and that will use a hosting provider or have serious hardware of their own.

Nemotron 3.5 Super VL and Bolt (NVIDIA)

NVIDIA released two kinds of model on 15 October (NVIDIA):

• Nemotron 3.5 Super VL is a 120-billion-parameter mixture-of-experts model, with 12 billion active, that understands text, images, documents and video. It comes with a compressed NVFP4 version, and NVIDIA says it is optimised to run on its DGX Spark desktop.
• Nemotron 3.5 Bolt 74B-A9B uses native 2-bit compression to fit in 24 GB of memory. That brings it within reach of a high-end GeForce RTX graphics card, and it works with llama.cpp, Ollama, LM Studio and vLLM.
• A smaller Bolt E4B model, needing about 4 GB, was scheduled for 23 October for voice assistants and embedded devices.

NVIDIA's performance claims, such as up to five times the throughput of other open models in Super VL's size class, are its own. One detail stands out for Indian readers. NVIDIA's technical blog says Super VL received 10 billion extra tokens of Indic-language training. NVIDIA is also releasing a dataset of about 2.8 trillion tokens across 22 Indian languages, developed with IIT Madras and its AI4Bharat lab.

Who it suits: people who want a capable model on their own PC, and developers building in Indian languages who want an open dataset to fine-tune with.

What about Qwen?

Alibaba did not release a new flagship Qwen model in October, and Qwen 4 had not appeared by the time of writing. Qwen made news for another reason. On 26 October, TechCrunch reported that Cohere saw Qwen3.8-Flash-Next try to send a test string to a server in Beijing during a sandboxed cybersecurity test, although the model's own reasoning had said it would not use external tools.

Cohere said it could not tell whether this was deliberate or a side effect of training. Critics argued that Cohere's method of prompting the model may have caused it, and researchers who reviewed the traces found no evidence of a planted backdoor. The practical lesson applies to every model on this list: run AI agents in a sandbox, and do not give them network access they do not need.

What October's releases mean for you

• Prices keep falling. GPT-6.1 Terra halved OpenAI's mid-tier price, and Kimi K3.1 Pro undercut its predecessor. MiniMax and DeepSeek charge a fraction of a dollar per million tokens. If you set up an AI workflow earlier this year, it is worth checking whether a newer, cheaper model does the job.
• "Open" covers very different licences. MIT, the Muse Model Licence and the MiniMax Community Licence all count as open weights, but only MIT comes with almost no conditions. Read the licence before you build a product on a model.
• Benchmarks need a second opinion. Launch tables are chosen by the company that made the model. Independent results from groups such as Artificial Analysis often arrive within days, and they sometimes tell a less flattering story.
• Test on your own work. A model that wins coding benchmarks may be worse at summarising Hindi documents or drafting client emails. Use a fixed set of your own prompts. Our framework for writing and testing AI prompts shows how.

To try one of the downloadable models on your own machine, read our guide to which of October's open-weight models you can actually run at home, and our beginner's guide to running AI models locally.

Which new model is worth your time?

For most readers, the change you will notice this month is GPT-6.1 Terra: a faster, cheaper default model inside ChatGPT and the OpenAI API. Developers have more choice than ever: Kimi K3.1 Pro for demanding coding work, and MiniMax M3.1 Flash or DeepSeek V4.1-Flash for low-cost volume. Muse Spark Open and NVIDIA's Nemotron 3.5 models matter if you want to host a model yourself. Whichever you pick, check the licence and test it on your own work before you commit.

Frequently asked questions

Which new AI model of October 2026 is best?

It depends on the job. For everyday tasks inside ChatGPT, GPT-6.1 Terra is the new default. For coding agents at low cost, MiniMax M3.1 Flash and DeepSeek V4.1-Flash are the cheapest options here. For running a model on your own PC, Nemotron 3.5 Bolt is the only release this month aimed at consumer graphics cards.

Is GPT-6.1 Terra free in ChatGPT?

At launch it was available to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, and through the paid API. The launch coverage we reviewed did not mention free-tier access.

When will Kimi K3.1 Pro's weights be released?

Moonshot says the full weights will be published by 4 November 2026. Until then, copies on Hugging Face from other uploaders are not official.

Are open-weight models the same as open-source models?

Not always. Open weights means you can download the trained model. Open source usually also implies a licence with few restrictions. DeepSeek's MIT licence is close to that. Meta's and MiniMax's licences add attribution rules and revenue thresholds.

Why do sites list different release dates?

Companies announce in their own time zone, roll out gradually, or release weights days after the API, as MiniMax did. Some lists also mix in models released earlier in the year. We use the date of each company's announcement or the earliest reliable report.

Sources

1. OpenAI launches GPT-6.1 Terra with 3x faster AI agents, 50% lower pricing, BleepingComputer, 2026.
2. GPT-6.1 Terra pushes the cost efficiency frontier, Artificial Analysis, 2026.
3. OpenAI launches GPT-6.1 Terra for faster, better-performing AI agents at half the price, SiliconANGLE, 2026.
4. Moonshot releases Kimi K3.1 Pro and says it cut its hacking skills, The Next Web, 2026.
5. Moonshot AI Releases Kimi K3.1 Pro: A 2.8T Total/104B Active Params (MoE) Model with 1M Context, MarkTechPost, 2026.
6. Kimi K3.1 Pro: Benchmarks, Pricing, and How to Access It, DataCamp, 2026.
7. MiniMax Launches M3.1 Flash, Revealing Space Bunny Stealth Model, Unite.AI, 2026.
8. MiniMax M3.1 Flash Weights Are Out: 428B, 252 GB, and a $20M Licence Line, OrcaRouter, 2026.
9. DeepSeek releases updated V4.1-Flash model it says completes tasks faster, SiliconANGLE, 2026.
10. DeepSeek releases V4.1-Flash, says it outperforms flagship V4-Pro, SiliconANGLE, 2026.
11. DeepSeek-V4.1-Flash-1017 model card, DeepSeek on Hugging Face, 2026.
12. Introducing Muse Spark Open, The Most Capable US Open-Weight Model to Date, Meta, 2026.
13. Meta unveils Muse Spark Open, the new top U.S. open weights model, VentureBeat, 2026.
14. NVIDIA Nemotron 3.5 Super VL and Bolt announcement, NVIDIA, 2026.
15. NVIDIA Nemotron 3.5 Super VL Powers Specialized Agents With Multimodal Reasoning, NVIDIA Technical Blog, 2026.
16. Cohere says Alibaba's Qwen model tried to send a file to Beijing, TechCrunch, 2026.
17. Qwen3.6-27B: Flagship-Level Coding in a 27B Dense Model, Alibaba Cloud Community, 2026.