Time machine

AI Atlas as of 12 Sept 2026

Pick any date: the models that existed, their context lengths and status at the time, the provider prices valid that day, the benchmark leaders known by then, the hardware. Every value opens the claim that establishes it, with its validity interval.

ObservedThis date is inside the observation history: values are claims that were current that day.

Observation history starts 11 Sept 2026 20:50 UTC. AI Atlas observation history starts at 2026-09-11T20:50:56.012288+00:00. The state is taken from claims, prices and results as they were known at the end of that UTC day.

Models existing

1,578

Offers valid

887

Benchmark leaders

Hardware

Basis

observed

Benchmark leaders

Leaders known by 12 Sept 2026 23

Benchmark leaders as of 2026-09-12
Benchmark	Leader	Organization	Score	Metric · group	Trust	Models ranked
Aider polyglot	gpt-5	OpenAI	86.7	pass_rate_2	official-benchmark	61
Aider polyglot — well-formed responses	Codestral 25.01	Mistral AI	100	percent_cases_well_formed	official-benchmark	61
Artificial Analysis Intelligence Index	Claude Fable 5.1	Anthropic	53.37	index	independent-evaluator	477
GPQA Diamond	gpt-6-astra	OpenAI	96.26	accuracyaccuracy · variant=GPQA Diamond · evaluator=Artificial Analysis	independent-evaluator	459
Humanity's Last Exam	Claude Fable 5.1	Anthropic	59.13	accuracyaccuracy · evaluator=Artificial Analysis	independent-evaluator	453
IFBench	Grok 4.3	xAI	83.33	accuracyaccuracy · evaluator=Artificial Analysis	independent-evaluator	334
LiveBench	Claude Fable 5.1	Anthropic	83.41	global_average	official-benchmark	57
LiveBench Agentic Coding	DeepSeek-V4.1-Flash	DeepSeek	77.27	average scoreaverage score · variant=Agentic Coding	official-benchmark	57
LiveBench Coding	Claude Fable 5.1	Anthropic	86.38	average scoreaverage score · variant=Coding	official-benchmark	57
LiveBench Data Analysis	gpt-6-astra	OpenAI	82.97	average scoreaverage score · variant=Data Analysis	official-benchmark	57
LiveBench Instruction Following	Gemini 3.8 Flash	Google	81.41	average scoreaverage score · variant=IF	official-benchmark	57
LiveBench Language	Claude Fable 5	Anthropic	90.68	average scoreaverage score · variant=Language	official-benchmark	57
LiveBench Mathematics	Claude Fable 5.1	Anthropic	97.01	average scoreaverage score · variant=Mathematics	official-benchmark	57
LiveBench Reasoning	gpt-6-astra	OpenAI	92.65	average scoreaverage score · variant=Reasoning	official-benchmark	57
MMMU-Pro	gpt-6-astra	OpenAI	86.88	accuracyaccuracy · evaluator=Artificial Analysis	independent-evaluator	157
SWE-bench (full test split)	claude-3-opus	Anthropic	3.79	resolvedresolved · board=Test · system=RAG baseline	official-benchmark	6
SWE-bench Lite	claude-3-opus	Anthropic	4.33	resolvedresolved · board=Lite · system=RAG baseline	official-benchmark	6
SWE-bench Multilingual	gemini-3-flash	Google	72.7	resolvedresolved · board=Multilingual · system=mini-SWE-agent	official-benchmark	13
SWE-bench Multimodal	o3	OpenAI	35.98	resolvedresolved · board=Multimodal · system=GUIRepair	official-benchmark	4
SWE-bench Verified	Claude Opus 4.5	Anthropic	76.8	resolvedresolved · board=Verified · system=mini-SWE-agent	official-benchmark	42
SciCode	Claude Fable 5.1	Anthropic	63.08	accuracyaccuracy · evaluator=Artificial Analysis	independent-evaluator	129
Terminal-Bench	gpt-5.6-sol	OpenAI	65.91	accuracyaccuracy · variant=hard · evaluator=Artificial Analysis	independent-evaluator	317
τ²-bench	Z.ai GLM 5.2	Z.ai (Zhipu AI)	99.12	pass^1pass^1 · evaluator=Artificial Analysis	independent-evaluator	323

One row per benchmark: the best current row of the primary comparability group among results observed by the date. See the benchmark's Frontier tab for the full leader history.

leaders from results observed by the date (evaluation dates are not used: a result is known only once observed)

Shareable: this URL reproduces the view. Per-entity: every entity page has a History tab with the same as-of reconstruction. What changed since? Diff 12 Sept 2026 → today. Methodology →