AI Model Launch Tracker

What we captured

Google announced Gemini 4 Carbon, beginning limited availability for a subset of cyber defenders and planning a gradual expansion to trusted partners.

• • Announced October 28, 2026.
• • Initial availability is limited to a subset of cyber defenders.
• • Google reports scores of 88.2% on SEC-bench Pro, 54.8% on ExploitGym, and 71.0% on CWE-bench.
• • Introductory prices are $2 per million input tokens and $10 per million output tokens.

Source evidence

Confidence 100% / Completeness 92%

What we captured

Gnani AI launched its speech-to-text model Prisma v3.0, available through the company's developer API.

• • Trained on Indian telephone audio and heavily code-mixed speech.
• • Supports nine Indian languages and Indian English.
• • Gnani AI reports a 16.9% macro word error rate across nine Vistaar datasets in its evaluation.

Source evidence

Confidence 99% / Completeness 87%

What we captured

NVIDIA released open weights for Nemotron 3.5 Transcribe, a speech model supporting 40 languages and audio up to one hour.

• • The article states NVIDIA released the model on October 28, 2026.
• • The model supports 40 languages and audio up to one hour.
• • The article reports about 34 billion total checkpoint parameters and about 4.3 billion active parameters.
• • The stated license is OpenMDW-1.1, which allows commercial use.

Source evidence

What we captured

Liquid AI announced six LFM3 models for language, vision, speech, multimodal tasks, and structured decisions, with device, private-server, and cloud deployment options.

• • The announced family includes six models.
• • LFM3-1.7B has a documented 256,000-token downloadable context window and a 128,000-token hosted API limit.
• • The hosted LFM3-1.7B rate card lists $0.08 per million input tokens and $0.20 per million output tokens.
• • Reported benchmark results are Liquid AI's own and were not independently tested by VentureBeat.

Source evidence

What we captured

AssemblyAI launched its first text-to-speech model for voice-agent developers.

• • AssemblyAI described TTS-1 as its first text-to-speech model.
• • The stated price is $15 per million characters.

Source evidence

Confidence 98% / Completeness 87%

What we captured

Calix announced a publicly available, open-weight communications language model on Hugging Face.

• • Calix said it released the model on October 27, 2026.
• • The model was publicly available on Hugging Face.
• • Calix described its Communications Language Models as open-weight.

Source evidence

Confidence 97% / Completeness 87%

What we captured

OpenAI introduced GPT-6.1 Terra, making it available through ChatGPT subscriptions and an API.

• • The article dates the introduction to October 26, 2026.
• • Listed API prices are $1 per million input tokens and $5 per million output tokens.
• • The article reports availability through ChatGPT Work, Codex, and an API.

Source evidence

What we captured

Block introduced Dessa, a locally runnable voice model, and is inviting safety researchers to test it. It is not yet publicly available.

• • Block introduced Dessa, a voice model from its research lab.
• • The model can speak and listen simultaneously.
• • Block said it was trained entirely on human speech using a single NVIDIA GPU rack.
• • Block is inviting safety researchers to test the model; it is not yet publicly available.

Source evidence

What we captured

Canva launched Craft, its first image-generation foundation model, with worldwide availability in Dream Lab and access through its public API.

• • The article states that Canva launched Craft and made it available worldwide through Dream Lab and its public API.
• • Craft produces images up to 2K resolution and supports generated text in 55 languages.

Source evidence

What we captured

NVIDIA released open-weight checkpoints of Nemotron 3.5 Bolt E4B 4GB, a small conversational-agent model with multi-turn tool-use capabilities and a 64K context window.

• • Published October 23, 2026.
• • GGUF and BF16 checkpoints are listed.
• • NVIDIA reports a 64K context window and support for 35 languages.
• • The model builds on Gemma 4 E4B with NVIDIA Nemotron post-training.

Source evidence

Confidence 99% / Completeness 91%