THE FRONTIER SIGNALSWhich advanced AI systems can the public try or inspect?
Public APIs, model weights, research code and evaluation datasets reveal different parts of AI capability; none provides a complete view of undisclosed internal systems.
Open interactive Research access hub →
What can I access in Claude Opus 5.5?
Product behavior, published evaluations and system-card documentation.
Platform account and applicable plan/API charges. Some biology and cybersecurity work uses additional verification or fallback models.
The linked release does not provide downloadable model weights or a complete training recipe.
Access & system card ↗What can I access in DeepSeek V4 releases?
Official weight collection and a technical report linked from the preview announcement.
Check the specific checkpoint license and hardware requirements. Large models may require substantial infrastructure.
Public weights do not imply disclosure of all training data or exact equivalence with the current hosted API.
Release & report ↗ · Weight collection ↗What can I access in AlphaGenome?
Research implementation, evaluation examples and weights under non-commercial model terms.
API key for programmatic access. Weights require acceptance of model terms; the repository recommends at least an H100 for local inference.
Code licensing and model-weight terms differ. Predictions still need scientific interpretation and validation.
Atlas & overview ↗ · API examples ↗ · Code & weights ↗What can I access in Genie 3 / Project Genie?
Interactive product behavior, example worlds and prompting guidance.
Follow the official Project Genie entry point for current account, region and plan eligibility; access is not guaranteed for every visitor.
A playable world is not access to the training system, model weights or a validated physical simulator.
Try & prompt guide ↗What can I access in METR time-horizon data?
Model-level results, uncertainty intervals and evaluation definitions.
Read the methodology and dataset version. A spreadsheet is enough to begin inspecting the estimates.
Downloading estimates is not access to the evaluated models or proof of reliability outside the test distribution.
Methods & results ↗ · Download dataset ↗What can I access in Claude-assisted enzyme research?
The company’s description of model-assisted hypotheses and laboratory experiments.
Read as company-reported research. Specialized laboratory access and expertise are separate from reading the report.
The report alone does not provide access to the entire internal workflow or establish independent replication.
Read the report ↗What can I access in Unit 42 Continuous Frontier AI Defense?
Provider documentation describes use of gated cybersecurity models in a managed service.
Commercial subscription and provider onboarding; model options vary. Restrict testing to authorized assets.
No downloadable weights or direct general-purpose model access is established by this announcement.
Service announcement ↗What else do readers ask?
Can I see AI capabilities beyond chatbots?
Yes. Scientific prediction tools, world models, model repositories and evaluation datasets provide public entry points into other domains.
Is a hosted API the same as downloadable weights?
No. An API exposes service behavior; weights allow running a model under the applicable license and hardware requirements.
Does open code mean open training data?
No. Code, weights, datasets and training recipes can each have different availability and terms.
Can I run AlphaGenome locally?
Its research repository provides code and gated weights under non-commercial model terms, with substantial hardware requirements. The API is another route.
Does a public demo reveal the most advanced internal system?
That cannot be established from a demo. Undisclosed internal systems remain unknown.
How should I choose something worth trying?
Start with a representative task, a baseline and a fixed budget. Compare accepted results, supervision, latency and total cost.