Models & Multimodal Intelligence

Evidence statusPrimary source · Product release

OpenAI moved the GPT-5.6 family—Sol, Terra and Luna—from limited preview to general availability. The release emphasizes higher capability per token, programmatic tool calling and a multi-agent ultra mode for difficult knowledge-work, coding, cyber and science tasks.

Evidence boundaryCapability, cost and benchmark comparisons are reported by OpenAI and should be read with the published evaluation methods and system card; performance can vary by harness, effort setting and workload.

Evidence statusPrimary source · Product release

Anthropic positioned Claude Sonnet 5 as its most agentic Sonnet release, able to plan, use browser and terminal tools and run more autonomously at a lower cost than its largest model tier. The release narrowed the capability gap between Sonnet and Opus for coding and knowledge work.

Evidence boundaryAutonomy still depends on the surrounding harness, permissions, tools and human review; vendor benchmark comparisons are not guarantees for every workload.

Evidence statusPrimary source · Product release

Claude Opus 4.8 improved coding, tool use, computer interaction and long-running professional workflows. Anthropic launched it with adjustable effort, lower-priced fast inference and a research-preview dynamic-workflows feature for large parallel-agent tasks.

Evidence boundaryDynamic workflows were a research preview rather than a generally available baseline feature; quoted customer results are not uniform third-party benchmarks.

Evidence statusPrimary source · Product release

GPT-5.5 expanded OpenAI's frontier model line for agentic coding, professional knowledge work and scientific research, with a later API availability update on 24 April. The release also foregrounded inference efficiency and stronger cyber safeguards.

Evidence boundaryThe launch page combines vendor benchmarks and selected external evaluations; production results depend on task design, tool access and prompting.

Evidence statusPrimary source · Technical preview

Meta introduced Muse Spark as a natively multimodal reasoning model supporting tool use, visual chains of thought and multi-agent orchestration. It was available in Meta's consumer surfaces while API access began as a private preview.

Evidence boundaryAPI availability was limited at announcement, and capability claims are based primarily on Meta's own evaluations.

Evidence statusPrimary source · Research

SAM 3.1 updated Meta's video segmentation and tracking stack with object multiplexing and global reasoning. Meta reported real-time multi-object tracking at higher throughput on a single H100, pointing toward more practical perception systems for video and embodied applications.

Evidence boundaryThroughput figures are vendor-reported and are sensitive to object count, hardware, resolution and implementation details.

Evidence statusPrimary source · Technical preview

Gemini Embedding 2 introduced one shared representation space for text, images, video, audio and documents. The public preview targeted multimodal retrieval, classification and search pipelines that previously required separate encoders.

Evidence boundaryThis record refers to the public-preview announcement; Google separately announced general availability on 22 April 2026.

Evidence statusPrimary source · Product release

Google updated Gemini 3 Deep Think as a specialized reasoning mode for modern science, engineering and research problems. Consumer access began through Google AI Ultra, while API access was initially offered through an expression-of-interest process.

Evidence boundaryAt launch, API access was not generally available; benchmark results should not be interpreted as proof of dependable autonomous scientific discovery.