Epoch AI report finds steep decline in cost of benchmark performance

Image: Primary

Published Wednesday, September 30, 2026 · 1:30 PM EDT

A report from Epoch AI finds that the cost of reaching a specified performance level on AI tests has fallen rapidly, with its headline measure declining by just under 50% per quarter. The measure tracks the cost of answering a question while meeting an 81.25% threshold on the GPQA Diamond test. The finding describes the price of performance on selected benchmarks, not the cost of every AI workload. Data before 2023 were extrapolated, and Epoch found different rates of improvement across the other tests it examined.

Sources

Tom's Hardware techspot.com

Published by Tech & Business, a media brand covering technology and business. This story was sourced from Tom's Hardware, techspot.com and reviewed by the T&B editorial agent team.

Back to Newswire

Keep reading

Full wire

Products AI

Databento raises $100M Series C at $1.1B valuation

Databento raised a $100M Series C at a $1.1B valuation, FinTech Global reported. The company supplies market data through APIs to the financial industry.

Security

Fullerton Health says breach compromised about 140,000 work pass records

Fullerton Health said on November 14 that about 140,000 work pass holder records were compromised in a breach of its Migrant Worker Primary Care Plan programme in Singapore. About 43,000 records belonged to people whose work passes were no longer valid, CNA reported.

Capital

Tendrils raises $30M for CPU architecture with thousands of parallel cores

Cambridge, UK-based Tendrils raised $30M in funding led by Benchmark, Tech.eu reported. The startup is designing a CPU architecture with thousands of cores operating in parallel. Tendrils Compute is among several UK chip startups attracting investor backing.

AI Capital

Luminal raises $25 million Series A led by Standard Capital

Luminal raised a $25 million Series A led by Standard Capital, CEO Joe Fioti told Axios Pro. The company provides an inference cloud and develops compiler software for testing how AI models run on different chips.

Capital AI

WhiteMatter raises $20 million Series A for AI agent oversight

WhiteMatter raised a $20 million Series A, Business Insider reported. Founded in 2024, the startup builds AI agents designed to oversee other AI agents, monitoring their file access, software use, decision making, and actions. It calls these oversight systems "guardian" agents.

Policy Products

Italian privacy authority fines Meta €805 million over gallery access

The Italian Data Protection Authority imposed a fine of €805 million ($887 million) on Meta Platforms Inc. The penalty concerns Meta's access to photos and videos in the galleries of Android users without those users' explicit consent.