AI Model Launch Tracker

What we captured

Sarvam AI released Sarvam Prism, an open-weight decision model based on Sarvam 105B, with weights available under Apache 2.0.

• • Released November 3, 2026.
• • Based on Sarvam 105B and adapted with LoRA fine-tuning.
• • Scores predefined options in a single forward pass.
• • Supports up to 131,072 tokens per question.
• • Sarvam reports 44 ms median response time on one NVIDIA B200 GPU.

Source evidence

Confidence 99% / Completeness 95%

What we captured

Cogent Security released VR-1.1, a security reasoning model tuned for repair, and reported results from its own LiveCVEBench run.

• • Cogent said it released VR-1.1 on Nov. 3, 2026.
• • Cogent reported 57.9% repaired across 190 LiveCVEBench tasks in its own run.
• • Cogent reported 45.3% for Claude Code and Codex in the comparison.

Source evidence

Confidence 96% / Completeness 79%

What we captured

ElevenLabs announced its Eleven Real-time voice-to-speech model, with v1 Mini and v1 Max versions, available through Eleven Agent and a real-time API.

• • Announced November 3, 2026.
• • Two versions: v1 Mini and v1 Max.
• • Supports more than 90 languages.
• • Available through Eleven Agent and the real-time API.
• • OpenAI real-time API compatibility was reported; WebRTC was not supported at release.

Source evidence

Confidence 99% / Completeness 87%

What we captured

Deep Principle released Hydrogen-1.0, its first open-weight scientific research model, built by post-training GLM-5.3 on the company's internal scientific data and records.

• • Deep Principle introduced Hydrogen-1.0 on November 2, 2026.
• • The company described it as its first open-weight model for scientific research.
• • The model is a post-training of Z.ai's GLM-5.3 using Deep Principle's internal scientific data, laboratory records, and discovery trajectories.
• • Chenru Duan claimed a 38.3% FrontierScience-Research score for Hydrogen-1.0 and 33.3% for GPT-6-Astra.

Source evidence

Confidence 97% / Completeness 87%

What we captured

Meituan released its native multimodal, long-horizon agent model with a 1-million-token API context window.

• • Released November 2, 2026.
• • 1.6 trillion total parameters; approximately 48 billion activated per token.
• • Native multimodal model for long-horizon agentic work, including coding.
• • 1-million-token API context window.

Source evidence

Confidence 98% / Completeness 91%

What we captured

LongCat lists LongCat-2.5 as available, describing multimodal visual understanding, agentic coding, and long-horizon engineering capabilities.

• • The changelog dates the availability notice to 2026-11-02.
• • The model is described as combining visual understanding and agentic coding.
• • The changelog lists compatibility with Claude Code, Codex, and Hermes.

Source evidence

Confidence 98% / Completeness 87%

What we captured

Canva launched Canva Brush, a vision-language model for prompt-based editing of existing designs. Supported edits are initially available to Canva Pro subscribers.

• • Canva describes Brush as an in-house vision-language model for editing existing designs using natural-language prompts.
• • Canva claims Brush is up to 69% faster and 46% less costly per edit task than its previous design agent.
• • Canva says the performance comparison is based on its own benchmarking and testing.

Source evidence

Confidence 98% / Completeness 87%

What we captured

Arrowhead AI launched Fletch, a text-to-speech model for multilingual customer calls, and said it is already deployed across more than one million calls daily.

• • Launched October 30 for multilingual customer conversations in India and Southeast Asia.
• • Arrowhead AI says Fletch supports 14 languages and is deployed on more than one million calls per day.
• • The company reports 80-millisecond latency and approximately 30% better pronunciation accuracy on selected terms.

Source evidence

Confidence 98% / Completeness 87%

What we captured

TBC launched Mischief, based on Wan 2.2, through fal.ai's image-to-video endpoint.

• • The model is based on Wan 2.2.
• • TBC announced a price of $0.02 per second of generated video.
• • TBC's speed and cost comparisons are company-reported and not independently benchmarked in the announcement.

Source evidence

Confidence 98% / Completeness 87%

What we captured

Mistral released the weights for Large 4 on Hugging Face under a modified MIT license.

• • The weights were released on Hugging Face.
• • The model has about one trillion parameters, with 52 billion active per step.
• • Artificial Analysis rated it at 26 on its Intelligence Index.
• • API pricing is $0.70 per million input tokens and $2.10 per million output tokens.

Source evidence