Alpha Evals · Public evidence edition

Alpha Frontier Lab Benchmark.

Compare nine frontier AI labs across model capability, public safety practice, long-task autonomy, and transparency. Each source measure remains separate; missing coverage is reported as not rated.

Review the lab benchmark Download the data

Labs compared: 9
Scored sources: 4 independent datasets
Snapshot: 2026-08-16 UTC
Alpha rating status: Not an Alpha rating
Official model results: None published

Current comparison

Frontier capability and public safety practice.

The chart uses the highest covered Epoch ECI model for each lab and the lab's FLI Summer 2026 safety score. METR autonomy and Stanford transparency are reported in the table without being folded into the axes.

Public evidence editionSource measures are shown separately. This comparison is not an Alpha rating or safety certification.

CapabilityOpenAI 161.65Anthropic follows at 161.53 on Epoch ECI.

Safety practiceAnthropic 2.66 · C+Highest FLI score in this nine-lab cohort.

Autonomy coverage3 of 9 labsMETR p80 data; all other labs remain NR.

Transparency8 of 9 labsStanford FMTI; Z.ai was not rated.

Horizontal position is the highest scored model for each lab in Epoch’s downloaded ECI data. Vertical position is the lab’s FLI Summer 2026 overall safety score. ECI is an arbitrary linear scale, not a percentage. The measures are not averaged.

Alpha Frontier Lab Benchmark source measures as of 2026-08-16
Lab	Epoch ECI	Highest covered model	FLI safety	Stanford transparency	METR p80 horizon
AnthropicUnited States	161.53	Claude Fable 5 (high)Released 2026-06-09 · Confident	2.66 · C+	46 / 100	186 hclaude-mythos-preview-early
OpenAIUnited States	161.65	GPT-5.6 Sol (medium)Released 2026-07-09 · Confident	2.28 · C	35 / 100	66 hgpt-5.2-2025-12-11_high
Google DeepMindUnited States	154.75	Gemini 3.1 Pro PreviewReleased 2026-02-19 · Likely	2.01 · C	41 / 100	90 hgemini-3.1-pro-preview
MetaUnited States	154.75	Muse Spark 1.1 (high)Released 2026-07-09 · Confident	1.32 · D+	31 / 100	NR
Z.aiChina	152.21	GLM-5.2Released 2026-06-16 · Confident	0.88 · D-	NR	NR
Alibaba CloudChina	156.05	Qwen3.8 Max (xhigh)Released 2026-08-02 · Confident	0.87 · D-	26 / 100	NR
xAIUnited States	154.02	Grok 4.5 (high)Released 2026-07-08 · Likely	0.65 · F	14 / 100	NR
DeepSeekChina	152.53	DeepSeek V4 Flash 0731 (max)Released 2026-07-31 · Likely	0.47 · F	32 / 100	NR
MistralFrance	142.64	Mistral Medium 3.5Released 2026-04-28 · Confident	0.33 · F	18 / 100	NR

NR means the named source did not publish a score for that lab in the cited edition. It does not mean zero.

Source ledger

What is used, what is missing, and why.

Public scores retain their original scale and date. Alpha does not convert missing coverage to zero or average unlike measures into a single rank.

Epoch AI · Epoch Capabilities IndexHighest scored model in the downloaded ECI data for each covered lab. ECI is an arbitrary linear scale, not a percentage. Retrieved 2026-08-16.9 of 9 labs Future of Life Institute · AI Safety IndexSummer 2026 overall score and grade. This independent assessment supplies the vertical axis. Retrieved 2026-08-16.9 of 9 labs METR · Time Horizon 1.1Longest published 80% success time horizon in the downloaded data for each covered lab. Unmeasured labs are NR. Retrieved 2026-08-16.3 of 9 labs Stanford CRFM · Foundation Model Transparency IndexDecember 2025 score out of 100. Z.ai was not included and is NR. Retrieved 2026-08-16.8 of 9 labs Artificial Analysis · Intelligence and Agentic IndicesThe supplied chart snapshot informed source selection. Model scores are not republished until Alpha has a licensed, authenticated server-side data feed. Retrieved 2026-08-16.Method reviewed; scores withheld

First-party benchmark under development

Boardroom effectiveness and governance assurance.

This second chart shows the proposed Alpha evaluation method using synthetic configurations. It is kept separate from the named-lab public-evidence benchmark above.

Read the boardroom method

Illustrative methodology viewSynthetic configurations A–F. No lab, model, vendor, or enterprise is rated or ranked here.

Authority is non-compensating: strong capability cannot erase a critical governance failure. Bands describe the maximum decision authority a configuration may be considered for after human review; they are not safety certifications.

Decision question

Can the system improve a consequential decision without exceeding its authority?

Representative tasks

Packet synthesis
Materiality
Management challenge
Decision framing

Assurance emphasis

Evidence fidelity, abstention, contradiction detection, and explicit human control

Accessible data for the synthetic chart
Configuration	Boardroom effectiveness	Governance assurance	Illustrative authority band	Critical gate state	Human review	Gate effect
Configuration A	82	71	A2	clear	pending	no_synthetic_cap
Configuration B	74	49	A1	failed_G3	pending	capped_at_A1
Configuration C	63	78	A2	clear	pending	no_synthetic_cap
Configuration D	55	43	A1	failed_G4	pending	capped_at_A1
Configuration E	38	67	A1	clear	pending	capped_at_A1
Configuration F	29	31	A0	failed_G5	pending	capped_at_A0

Measurement lineage

How evidence can become an Alpha rating.

External benchmarks are inputs, not evals. A published Alpha rating requires separate evidence review, governance analysis, and approval.

Stage 1Evidence

Stage 2Claims

Stage 3Signals

Stage 4Metrics

Stage 5Evals

Stage 6Indices

Stage 7Evals

Stage 8Outlook / Watch

Publication boundary

Current publication states.

Lab comparison: Public-evidence benchmark; not an Alpha ratingAvailable
Artificial Analysis scores: Withheld pending an authenticated, licensed data feedLimited availability
Boardroom task bank: Candidate first-party method; official ranking set remains privateLimited availability
Official Alpha results: Published only after reproducible evaluation and human reviewLimited availability