AI Model Launch Tracker

What we captured

AutoTrust AI released JEV-27B, an Apache-2.0 open-weights model for structured decisions in self-hosted AI agent workflows.

• • Released September 29, 2026.
• • Apache-2.0 open-weights release includes weights, decision adapter, training and serving code, vLLM support, and evaluation reports.
• • Built on a frozen Qwen3.8-27B backbone with a 108.9-million-parameter decision block.
• • AutoTrust reported a 84.07% mean across six benchmark groups and cautioned that its comparison was not independent validation.

Source evidence

Confidence 100% / Completeness 91%

What we captured

Anthropic released Claude Sonnet 5.5 as a generally available, closed-weights model, with availability on its Claude Platform and through AWS, Google Cloud, and Microsoft Azure.

• • Released September 28, 2026.
• • Anthropic reports 70.6% on Terminal-Bench 4.0.
• • List pricing is $2 per million input tokens and $10 per million output tokens, unchanged from Sonnet 5.
• • The model has a 1-million-token context window and 128,000-token maximum output.

Source evidence

What we captured

ElevenLabs launched its v4 and v4 Turbo speech models, with 90-language support, voice cloning from 10 seconds of audio, lower latency for voice agents, and context-aware expression control.

• • Launch announced September 28, 2026.
• • Supports 90 languages and voice cloning from 10 seconds of audio.
• • Offers lower latency for voice agents and context-aware expression control.

Source evidence

Confidence 98% / Completeness 87%

What we captured

NaiveAI released an MIT-licensed open-weight coding and AI research model with a 1-million-token context window.

• • Released September 27, 2026.
• • 309 billion total parameters; 15.5 billion active per token.
• • Native 1-million-token context window.
• • Weights and inference code released under the MIT license.

Source evidence

Confidence 99% / Completeness 100%

What we captured

Supersonic Labs released its 144.3-million-parameter CPU-oriented classifier and router with weights and code under Apache 2.0.

• • Released September 26, 2026.
• • The model has 144.3 million parameters and is designed for CPU inference.
• • Weights and code are available under Apache 2.0.
• • The API was in development; the stated token prices were planned pricing.

Source evidence

Confidence 100% / Completeness 96%

What we captured

Respan's Span-01 behavior-scoring model was listed on OpenRouter, with the listing dated September 26, 2026.

• • OpenRouter's listing dates the release to September 26, 2026.
• • Respan says its model family has processed more than 4.5 billion tokens since launch; this usage claim is not independently confirmed.
• • Span-01 is priced at $0.02 per million input tokens, with no charge for output tokens.

Source evidence

Confidence 98% / Completeness 92%

What we captured

Anthropic released Claude Opus 5.5, a proprietary model available through Claude.ai, API, and cloud providers.

• • The article says Claude Opus 5.5 was released September 25, 2026.
• • It accepts up to 1 million input tokens and has API pricing of $4 per million input tokens and $20 per million output tokens.
• • The article reports first-place rankings on Artificial Analysis’ Intelligence Index v4.3 and Vals AI’s Index.

Source evidence

What we captured

Stanford and Nvidia released CLM-8B weights under Apache 2.0 with open-source code and related developer tools.

• • CLM-8B is designed to rank or select among known candidate actions using state and action embeddings.
• • The released model uses a frozen Qwen3-8B backbone and separate state and action projection heads.
• • The team reports up to 9x faster inference than Jev in its tests.

Source evidence

Confidence 99% / Completeness 91%

What we captured

Google unveiled Gemini 3.8 Live with selectable and customizable avatars for Gemini Enterprise customers.

• • Gemini 3.8 Live includes preset and custom avatars.
• • Avatar output combines real-time speech with low-latency video generation.
• • Google says avatar audio and video includes SynthID watermarking.

Source evidence

What we captured

Sarvam launched a document-reading model that handles forms, tables, and handwritten text in English and 22 Indian languages.

• • Launched September 24, 2026.
• • Reads documents in English and 22 Indian languages, including forms, tables, and handwriting.
• • Sarvam reported scores of 87.3 on olmOCR-Bench and 87.39 on its Indic benchmark.

Source evidence

Confidence 99% / Completeness 87%