Model names and prices move fast. This ledger tracks the current model or product lineup for 11 of 75 tracked tools, plus 3 research watches without public pricing — with the date each lineup was checked against official vendor sources. Research watches are not directory SKUs.
Omni world model: camera-controlled generation, spatial reconstruction, and space-time simulation
Shared spatial context over text, images, video, and 3D
Early access with select partners. No public self-serve plan prices published as of 2026-09-01.
Source: World Labs blog post “Atlas: A World Model for Spatial Intelligence” (https://www.worldlabs.ai/blog/atlas), checked 2026-09-01.
Pretrained from scratch as a multimodal autoregressive diffusion transformer. Camera-controlled images/videos up to 1 minute at 1440p; reconstruction from 1–dozens of images with explicit 3D (point clouds, 3D Gaussian splats); video reframing and Real-to-Sim for robotics; also generates images and 360 panoramas from text.
Will power future versions of Marble and other World Labs products. Not a public self-serve SKU; request access on the official blog.
Verify current names at openai.com — OpenAI iterates point releases frequently.
2026-09-03: OpenAI released GPT-6 Astra. Official Path to Astra post designates it at Critical cybersecurity capability, with advanced cyber access gated to trusted testers / Daybreak Blue. Axios: limited Daybreak Access first, then ChatGPT Plus / Pro / Business / Enterprise and API in the coming days. It is not a new ChatGPT SKU and is tracked as a separate research watch. Plus/Pro plan prices were not confirmed as changed on this check.
New OpenAI frontier model; limited Daybreak Access first, then announced ChatGPT and API roll-out
No public context-window figure on the official Path to Astra post as of this check
Announced onto existing ChatGPT Plus / Pro / Business / Enterprise and the API in the coming days. No Plus/Pro price change confirmed as of 2026-09-03. Advanced cyber access gated separately.
Source: official OpenAI post “Path to Astra: critical capabilities and frontier safeguards” (https://openai.com/index/path-to-astra/), dated 2026-09-01 and used as the primary official page on this 2026-09-03 check. No separate system-card URL was found on openai.com during this check.
Official post: first model designated at Critical cybersecurity capability under the Preparedness Framework. Advanced cybersecurity workflows limited at launch to trusted testers, with Daybreak Blue expansion for defensive use.
Launch-day access timing (limited Daybreak organizations first; ChatGPT Plus / Pro / Business / Enterprise and API developers in the coming days) is from Axios, corroborated by TechCrunch. OpenAI customer stories dated 2026-09-03 use the name GPT-6 Astra. Those stories’ task-specific scores were not copied as site benchmarks.
Research watch only. Not a new ChatGPT SKU page and not a product rating. Plus/Pro plan prices were not changed in sources checked 2026-09-03. AGI remarks and third-party market comparisons were not used as facts.
Local inference engine for Apple silicon + Qwen3.6-35B-A3B (Perplexity Computer hybrid compute)
Specialized for one architecture; official demo uses the Qwen3.6-35B-A3B MLX affine 4-bit checkpoint format
Open-source Apache-2.0 demo in perplexityai/pplx-garden (lily/). Not a Perplexity plan SKU. No public Lily price as of 2026-09-02.
Source: official Perplexity blog “Optimizing On-Device Inference for Apple Silicon” (https://www.perplexity.ai/hub/blog/optimizing-on-device-inference-for-apple-silicon), checked 2026-09-02. Standalone demo named on that post; official pplx-garden README points to github.com/perplexityai/pplx-garden (lily/) and the same blog.
Official blog: Rust runtime; OpenAI-compatible chat-completions API; custom Metal kernels; neither PyTorch nor MLX in the execution path. MLX affine 4-bit in the demo README is the weight-file format, not MLX as the runtime.
Official lily README (checked 2026-09-02): Apple GPU family 10 or later (M5 and newer), macOS 26 or later, greedy decode, POST /v1/chat/completions plus GET /v1/models and GET /health. Streaming, sampling, tools, and multimodal requests are rejected in that demo surface.
Local-engine watch only. Not a directory SKU and not a product rating. Vendor-published throughput figures were not independently reproduced.
Perplexity routes to whichever underlying model you pick; its own Sonar handles most searches.
2026-09-02 official blog: Lily is an open-source local engine for Apple silicon and Qwen3.6-35B-A3B used in Perplexity Computer hybrid compute. It is not a new Pro/Max SKU and is tracked as a separate Local AI research watch, not as a 4.5 product rating.