Epoch AI · Sep 23, 2026
Epoch AI reports progress on AI furniture-assembly reasoning
Markdown · JSON · Original source
On September 23, 2026, Epoch AI described the Furniture Assembly Benchmark (FAB), which tests whether models can spot assembly mistakes in photos using instructions and image-inspection tools. Epoch AI reports that GPT-6 Astra scored 80%, compared with 28% for Claude Opus 4.5 in November 2025. The benchmark also measures how long models take: Epoch AI says Astra had a median time of three minutes per photo, which it considers too slow for practical real-time feedback. Epoch AI presents photo-based assembly checking as a possible proxy for tasks such as car and appliance repair, while describing reliable real-time guidance as a possibility rather than a demonstrated use. The results cover 60 photos from three furniture builds, and Epoch AI says it has not measured human performance on the benchmark. It also reports no clear relationship between furniture complexity and model accuracy. These limits leave open how well the results extend to other physical tasks or to real-world repair. The benchmark measures performance on its test images; it does not establish that models can provide reliable guidance in practical settings.
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
Epoch AI · Article
Source details
Original source date: Sep 23, 2026 (2026-09-23T00:00:00.000Z).
Date basis: Date supplied by the publisher’s feed.
Added to Escape Meridian: 2026-10-29T15:50:30.157Z. First observed: 2026-10-18T02:43:54.812Z. Retrieved: 2026-10-23T20:45:52.646Z.
Referenced by
Recorded citations and public file attachments. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
Text as retrieved on October 23, 2026; originally published September 23, 2026. The benchmark covers 60 photos from three furniture builds; Epoch AI says it has not measured human performance, and the results' relevance to other physical tasks remains unclear. Epoch AI presents assembly checking as a possible proxy for other tasks, not as a demonstrated real-world repair capability.
Open question
How well would models perform on different furniture, with human comparisons, or in real repair settings?
Source-stated
Epoch AI says the Furniture Assembly Benchmark uses 60 images from three IKEA furniture builds to test whether models identify assembly mistakes and describe them.
The benchmark consists of 60 images spanning three different IKEA furniture builds.
Source-stated
Epoch AI reports GPT-6 Astra scored 80% on FAB, while Claude Opus 4.5 scored 28% in November 2025.
The best score in November 2025 was 28%, set by Anthropic’s Claude Opus 4.5. Just ten months later, OpenAI’s GPT-6 Astra now tops the leaderboard at 80%.
Source-stated
Epoch AI reports Astra's median completion time was three minutes per photo and says this is still too slow for practical real-time feedback.
At a median of three minutes per task, it is still a little too slow for practical real-time feedback, but it is not far off.
Source-stated
Epoch AI says it has not yet measured human performance on FAB and that the results' relevance to other physical tasks remains unclear.
We tested models on only 60 photos from three furniture builds, so it remains unclear how well these results extend to other physical tasks.
Source and contributor context
Epoch AI · Publisher
The publisher is identified here. No individual author has been inferred.
Source media
Media retrieval status: unchecked.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Epoch AI
Epoch AI studies trends in AI capability, training resources and the infrastructure behind them.
Claude
Claude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
Editorial provenance
Version 1. Publication ID: 3e5830b7-f78f-9b7f-d1d3-9e8ca42d6d5b.
Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5.
Source check: passed by anthropic/claude-opus-5.5 on 2026-10-29T15:50:14.765Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: bridge_admitted_by_model, second_opinion_categorical.
Observation: e34cf06fac550ad50c9eadb84ed9a19ceab0a737f3b611ab09ee93190c60c404. Source text hash: 876a2951cabd7b19b68aa5c48941beb1b95fcbf03ef77392a025ece47aa47ab1.
Revision and source-check history
Version 1 · 2026-10-29T15:50:30.157Z
Epoch AI reports progress on AI furniture-assembly reasoning
On September 23, 2026, Epoch AI described the Furniture Assembly Benchmark (FAB), which tests whether models can spot assembly mistakes in photos using instructions and image-inspection tools. Epoch AI reports that GPT-6 Astra scored 80%, compared with 28% for Claude Opus 4.5 in November 2025. The benchmark also measures how long models take: Epoch AI says Astra had a median time of three minutes per photo, which it considers too slow for practical real-time feedback. Epoch AI presents photo-based assembly checking as a possible proxy for tasks such as car and appliance repair, while describing reliable real-time guidance as a possibility rather than a demonstrated use. The results cover 60 photos from three furniture builds, and Epoch AI says it has not measured human performance on the benchmark. It also reports no clear relationship between furniture complexity and model accuracy. These limits leave open how well the results extend to other physical tasks or to real-world repair. The benchmark measures performance on its test images; it does not establish that models can provide reliable guidance in practical settings.
Original source
Epoch AI · Article
Source details
Original source date: Sep 23, 2026 (2026-09-23T00:00:00.000Z).
Date basis: Date supplied by the publisher’s feed.
Added to Escape Meridian: 2026-10-29T15:50:30.157Z. First observed: 2026-10-18T02:43:54.812Z. Retrieved: 2026-10-23T20:45:52.646Z.
Limits
Text as retrieved on October 23, 2026; originally published September 23, 2026. The benchmark covers 60 photos from three furniture builds; Epoch AI says it has not measured human performance, and the results' relevance to other physical tasks remains unclear. Epoch AI presents assembly checking as a possible proxy for other tasks, not as a demonstrated real-world repair capability.
Open question
How well would models perform on different furniture, with human comparisons, or in real repair settings?
Source-stated
Epoch AI says the Furniture Assembly Benchmark uses 60 images from three IKEA furniture builds to test whether models identify assembly mistakes and describe them.
The benchmark consists of 60 images spanning three different IKEA furniture builds.
Source-stated
Epoch AI reports GPT-6 Astra scored 80% on FAB, while Claude Opus 4.5 scored 28% in November 2025.
The best score in November 2025 was 28%, set by Anthropic’s Claude Opus 4.5. Just ten months later, OpenAI’s GPT-6 Astra now tops the leaderboard at 80%.
Source-stated
Epoch AI reports Astra's median completion time was three minutes per photo and says this is still too slow for practical real-time feedback.
At a median of three minutes per task, it is still a little too slow for practical real-time feedback, but it is not far off.
Source-stated
Epoch AI says it has not yet measured human performance on FAB and that the results' relevance to other physical tasks remains unclear.
We tested models on only 60 photos from three furniture builds, so it remains unclear how well these results extend to other physical tasks.
Source and contributor context
Epoch AI · Publisher
The publisher is identified here. No individual author has been inferred.
Source media
Media retrieval status: unchecked.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Epoch AI
Epoch AI studies trends in AI capability, training resources and the infrastructure behind them.
Claude
Claude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
Editorial provenance
Version 1. Publication ID: 3e5830b7-f78f-9b7f-d1d3-9e8ca42d6d5b.
Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5.
Source check: passed by anthropic/claude-opus-5.5 on 2026-10-29T15:50:14.765Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: bridge_admitted_by_model, second_opinion_categorical.
Observation: e34cf06fac550ad50c9eadb84ed9a19ceab0a737f3b611ab09ee93190c60c404. Source text hash: 876a2951cabd7b19b68aa5c48941beb1b95fcbf03ef77392a025ece47aa47ab1.
passed · 2026-10-29T15:50:14.765Z · anthropic/claude-opus-5.5
Publication: 3e5830b7-f78f-9b7f-d1d3-9e8ca42d6d5b
attribution — pass: The benchmark, scores, timings, proxy argument and limitations are all attributed to Epoch AI, the publisher and author of the piece. Model makers are named as the source names them.
We created the Furniture Assembly Benchmark (FAB) to help fill this gap.
summary — pass: The summary is understandable without specialist background. It explains what FAB tests, reports the score jump, and states the proxy argument and the limits. It says Epoch considers three minutes too slow; the source says 'a little too slow' and 'not far off'. That is a slight softening of tone, not a misstatement. The title and question follow from the evidence.
At a median of three minutes per task, it is still a little too slow for practical real-time feedback, but it is not far off.
date — pass: September 23, 2026 matches the UTC calendar day of evidence.originalPublished. The source's own reference to September 2026 is consistent with that date. Nothing is presented as happening after publication. The evidence is a publisher page, so no archive disclosure is required.
Ten months later, in September 2026, Astra scored 80%.
support — pass: Each claim is source-stated: the 60 images, the 80% versus 28% scores, the median of three minutes, the unmeasured human performance, the unclear complexity relationship and the uncertain extension to other tasks. Real-time guidance is correctly framed as Epoch's stated possibility, not a demonstrated benefit.
Although we have not yet measured human performance on the benchmark, we expect many of these mistakes to be challenging to find for those unfamiliar with the build.
media — pass: No media was selected and no verified media candidates were supplied, so there is no identity or relevance issue.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.