Methodology

Sources

Every event and every score comes from the lab's own official announcement: its news or blog post, or a system or model card the launch post links. Each event links its source post. When a lab takes a post down, the event keeps the original link and adds an archived copy.

Labs covered

Full coverage: OpenAI, Anthropic and Google. Every event from their official posts, back to each lab's first.
Model releases: Meta, DeepSeek, Mistral, Alibaba, Moonshot AI, MiniMax, Z.ai and NVIDIA. Every language-model release from 2025 on, and their milestones.
Milestones: xAI, Snowflake, Sber, Thinking Machines and Poolside. Their milestones, and their firsts in a capability the timeline traces, such as open weights.

What counts as an event

An event is something a lab shipped or decided that changes what people can do with AI, or how the lab operates: a model, a capability, a product, or a policy or corporate move. Customer stories, tutorials, conferences, small feature or pricing changes and most research papers are left out.

Each event has one of three levels of importance:

Milestone: an event the history of the field is told by. About one per lab per year.
Major: what an informed reader would expect on a one-page history of the period, such as each frontier model launch.
Minor: real and notable, but a follow-on: smaller model tiers, wider availability, funding and partnerships.

Before ChatGPT (30 November 2022), the timeline holds milestones and major events only.

Dates

An event's date is the publication date shown on its source post. A post without a full date (day, month and year) is not used. An announcement and a later release are two events, each dated by its own post. A few older posts show only the date they were last updated; their events carry that date.

Benchmark scores

Each score is one a lab reports for its own model, in the launch post or a card it links. Scores a lab prints for other labs' models are left out, as are values that could only be read off a bar or an axis.

Each score is recorded exactly as printed, with its benchmark version (Terminal-Bench 2.0 and 2.1 are separate benchmarks) and setting (such as with or without tools).
A score's notes give the conditions the source states that affect the score: the harness, effort level, tools, context budget and number of runs averaged.
A model has one main score per benchmark: the one in its own launch post or card. Charts, rankings and event pages show the main score. When the lab reports the same score again later, for example in a newer model's card, that report is listed under the main one.
Scores from parallel test-time compute (best of several attempts, or majority voting) are listed, but left out of charts and rankings.
A score carries a caveat when the lab ran its own version of the benchmark's task set, or didn't name the version it ran.
Charts and rankings compare models only on the same benchmark version and setting.

Verification

Every score is read from its source twice, independently, and published only when the two reads match exactly; a third read settles any mismatch. A person then checks each score against its source. The Human-verified column shows “yes” once they have, and “not yet” until then.

An event's source, date and lab come straight from its post. Its title, description, importance and tags are drafted from the post and checked by a person before publishing.

Corrections

Report a wrong date, title, source, tag or score through the API's feedback endpoint, or on X to @paradite_. A corrected event keeps its page address.

Citing

Cite the event page, with the lab's post alongside it:

Claude 3 family (Anthropic, 2024-03-04). Frontier Log. https://frontierlog.ai/event/2024-03-04-claude-3-family