In November 2026, 27 AI models shipped across 25 vendors, and 12 of them carried open-weight licences. Moonshot AI's Kimi K3.1 Pro, Mistral AI's Mistral Large 4, Anthropic's Claude Fable 5.2 and Google's Gemini Omni 1.2 Flash stood out among the batch. The month also brought small edge-ready models such as ZenteiQ's Dhir 2B and Dhir 5B for offline use. More details about the November release wave are available on the AI category page: AI category.
The speed of these releases matters now because developers and enterprises are pushing AI agents into production. The open-weight models cut licensing fees and let teams run models on their own hardware. That shift could change how much it costs to deploy AI before 2027.
The November 2026 Surge of Open Models
The open-weight wave started earlier in the month. On November 3, Saket Labs released Eka-1, a 26.3 billion parameter Indigenous model trained on 7.5 trillion tokens across 22 Indian languages. On November 5, Moonshot AI and Recursive Intelligence also published open weights for frontier-scale systems.
Mistral AI shipped Mistral Large 4 with 1 trillion total parameters and 49 billion active ones, trained on European computing infrastructure. The company released the weights under an MIT-modified licence. That move put a high-performance European model in the hands of any researcher willing to download the files.
Meta released Muse Ember, an efficient-tier model that matches Claude Haiku 5.5 on an intelligence index at roughly one-third the cost per task. The model scored 36 on the AA Intelligence Index and 68.1 percent on AutomationBench-AA. Meta did not say when or if it would open-source the weights.
After the Kimi release, OpenAI launched GPT-Image-2.6, a developer-focused image model that adds native GIF generation and SVG vector export. The model sits behind the same API as GPT-6 Cyber, but it targets creative tooling rather than security. Pricing for GPT-Image-2.6 was not disclosed in the tracker.
The same month saw the release of Ling-3.2 Flash from Ant Group, a 560 billion parameter multimodal model that runs on API and proprietary access. The first two weeks are free. The model handles text, image and video inputs for medical and technical workflows.
Kimi K3.1 Pro Leads the Open-Weight Frontier
Kimi K3.1 Pro arrived on November 5, and it quickly became the top open model on the Agent Arena leaderboard. Moonshot AI trained the mixture-of-experts system on 2.8 trillion total parameters, with 104 billion active parameters per token. The open weights carry a modified MIT licence and support a 1 million token context.
The model costs $2 per million input tokens and $10 per million output tokens through the API. Moonshot AI claims a 5.1 percent net improvement over the prior K2 model on the Agent Arena benchmark. The company also reported a 7.9 percent rise in Bash reliability and a 6.5 percent gain in controllability.
The benchmark gap between open models and closed ones has narrowed sharply in 2026. FrontierCode scores for Kimi K3.1 Pro reached 69.9 percent under high budget, while Gemini 4 Carbon scored 59 on the Intelligence Index V5. Anthropic's Claude Fable 5.2 posted a 69.9 percent FrontierSWE score with a high budget, and a 49.1 percent FrontierCode score.
Why Low-Cost Agents Matter for Developers
The release of Ripple-65B Preview on November 6 pushed the cost lower. The open-weight model uses a dense transformer trained on 12 trillion tokens, and it costs $0.20 per million input tokens and $0.60 per million output tokens. It ranked top on the Terminal-Bench 2.1 and SWE-bench Pro v2.0 Hard.
Developers building computer-use agents now have options that run on a single GPU. Tencent's Hy4 open-weight system uses 770 billion parameters and supports native visual reasoning. It scored 3.04 across 579 financial tasks in blind evaluation, according to the tracker.
On the closed side, OpenAI launched GPT-6 Cyber on November 5, a specialized cybersecurity model that received a Critical rating under the Preparedness Framework. The model charges $20 per million input tokens and $100 per million output tokens. Its ExploitGym score was 48.5 percent.
Anthropic followed with Claude Fable 5.2 on November 9, which focuses on complex reasoning and autonomous software engineering. The API price sits at $8 per million input tokens and $40 per million output tokens, with a fast mode at $16 per million input tokens and $80 per million output tokens. Reuters reported that.
Google unveiled Gemini Omni 1.2 Flash on November 5, a video generation model that expands to 15 seconds with synchronized audio. It costs $0.07 per second at 720p resolution. Google also released Gemini 4 Carbon under its Fairwind program for cybersecurity defenders.
The mix of closed and open releases shows a market that is splitting into specialist and general-purpose tiers. Small models like Dhir 2B and Dhir 5B fit edge devices and offline settings. Large open models like Kimi K3.1 Pro and Mistral Large 4 target research and enterprise workloads.
The Pricing Shift in Open Models
Open-weight inference costs now sit below one dollar per million tokens for many frontier models. Kimi K3.1 Pro prices at $2 per million input tokens, while Ripple-65B Preview drops to $0.20 per million input tokens. That pricing floor lets startups run agent loops that burn through tokens quickly.
The move is a big bet for open source, and it's likely to reshape how labs share code.
The drop in inference cost matters for data centers. A 2.8 trillion parameter open model can be run on a cluster of GPUs instead of paying per token. That changes the unit economics of AI agents that run for minutes or hours.
Regulation and the Next Wave
The government has already moved to classify advanced AI systems. GPT-6 Cyber became the first Critical label under the Preparedness Framework. Critics say the label may slow open-weight releases, while vendors argue that open code reduces hidden risk.
Cost per task now drops sharply as open-weight releases replace API fees. Ripple-65B Preview and Kimi K3.1 Pro both price below one dollar per million tokens for inference. That range enables high-volume agent loops that would have been out of reach a year ago.
The next test is long-horizon reliability. Kimi K3.1 Pro improved Bash reliability, but agents still fail after long task stretches. Anthropic's Claude Fable 5.2 targets autonomous software engineering, with scores on FrontierSWE at 69.9 percent.
Regulation may tighten around open weights. The US government labeled GPT-6 Cyber as the first Critical AI system under the Preparedness Framework. The move could shape how open models are released in 2027.
For developers, the practical change is simple. A team can now download a 2.8 trillion parameter model, run it on its own cluster, and tune it without paying per token. The open weights carry a modified MIT licence, which removes commercial restrictions.
For enterprises, the cheaper agents mean they can experiment with autonomous workflows without a large budget. Google, Meta, Mistral, Moonshot and Recursive all shipped models that scale from edge to data center. The diversity of vendors widens the set of possible architectures.
For researchers, the open releases provide reproducible training data and weights. The Eka-1 model, trained on 22 Indian languages, adds to the case that open models can serve regional needs. The Soket Labs release did not include pricing details.
Google plans to ship Gemini Omni 1.2 Flash to its Vertex customers in December. The company said the video model will support multi-turn conversational editing at launch. That release will test whether open-weight agents can keep pace with closed flagship models.