Back to News
InnovationAI Understanding briefing

GPT-6 Astra tops furniture assembly benchmark with 80% accuracy

OpenAI’s GPT-6 Astra achieved an 80% success rate on the new Furniture Assembly Benchmark, out‑performing rival models and highlighting rapid advances in visual‑spatial reasoning for physical‑task assistance.

4 min readRead the linked source
Source-provided image accompanying GPT-6 Astra tops furniture assembly benchmark with 80% accuracy
Source referenceSource recorded
Publisher
epoch.ai
Source link
epoch.aihttps://epoch.ai/publications/furniture-assembly
Source type
Linked source — primary-source status has not been established.
ContextUnderstand this in 60 seconds

Start here

Key terms

Benchmark
A standardized test or dataset used to measure and compare model performance.
API (Application Programming Interface)
A structured way for one software system to send requests to and receive responses from another system.
Generalization
How well a model performs on new, unseen data outside the training set.
Test yourselfWhat is AI? Quiz

What happened

OpenAI evaluated its GPT‑6 Astra model on the Furniture Assembly (FAB), a 60‑image test suite of IKEA builds that includes intentional assembly mistakes. Astra identified mistakes in 80% of the cases, beating the next‑best models—Anthropic’s Claude Fable 5.1 (70%) and Claude Opus 5 (61%). The test gave each model the assembly manual, a zoom tool, and a Python interpreter, with an 80‑step budget per image. Astra completed the tasks in a median of three minutes, far faster than competitors, though still too slow for real‑time feedback.

The researchers at Epoch AI created the Furniture Assembly (FAB) to evaluate visual‑spatial reasoning in AI. FAB consists of 60 photographs from three IKEA furniture builds, each paired with the official assembly manual and a set of tools for image inspection.

During evaluation, each model received the manual PDF, a zoom capability, and a Python interpreter, and could take up to 80 reasoning steps per image. If the model reached the step limit, it received a warning and a final answer prompt.

GPT‑6 Astra, OpenAI’s latest vision‑enabled model, correctly identified mistakes in 48 of the 60 images, yielding an 80% score. The next‑best performer, Anthropic’s Claude Fable 5.1, scored 70%, while Claude Opus 5 achieved 61%. Other evaluated models from Google, Moonshot, and Alibaba lagged further behind.

Astra’s median processing time was three minutes per image, substantially faster than competitors, though still above the sub‑second latency required for interactive assistance.

Source details: epoch.ai ↗

Why it matters

The result shows that frontier language models can now combine visual perception with textual instructions to spot physical errors, a capability that underpins future real‑time assistance for complex assembly, repair, and maintenance tasks. By demonstrating a clear leap in both accuracy and efficiency, GPT‑6 Astra narrows the gap between AI reasoning and practical, on‑site problem solving. However, the three‑minute response time indicates that further speed improvements are needed before the technology can be deployed as an interactive guide for consumers or technicians. The also reveals differing failure modes across model families, informing future research on robustness and error detection.

The ability to link visual inputs with textual instructions and pinpoint errors demonstrates a level of embodied reasoning that could be applied to automotive repair, appliance troubleshooting, and other hands‑on domains.

Higher accuracy and lower token usage suggest that newer models are becoming more efficient, reducing computational costs for developers who might integrate such capabilities into consumer products.

The highlights that model size or complexity alone does not predict performance; factors like mistake subtlety and viewing angle play significant roles, indicating areas for targeted improvement.

The three‑minute latency, while a record for current vision‑language models, remains a barrier for real‑time user interaction, underscoring the need for hardware acceleration or algorithmic optimizations.

Interactive Mechanism

Interactive Mechanism: How It Actually Works

Explore the underlying technology behind this development interactively.

Thinking Budget (Test-Time Tokens):1,024 tokens
Complex Accuracy79%Math & Code Logic
Latency3.2sTime to first full output
Inference Cost$0.0092Per query estimated
Reasoning StyleStep VerificationInternal chain depth
Active Thinking Trace:
1Deconstruct user problem into formal constraints
2Propose candidate hypotheses & step-by-step calculation
3Self-correction: Backtrack and refute subtle edge cases
4Exhaustive consistency check & final output synthesis
Core takeaway: Test-time compute fundamentally changes AI economics. Instead of only scaling during pre-training, giving reasoning models more tokens at inference time allows them to systematically solve PhD-level STEM problems.
Interactive Concept Check+10 Points
What is AI? Quiz

Which description best fits "narrow AI", the kind of AI in use today?

What to watch next

Future work will focus on reducing latency to achieve true real‑time guidance, expanding the to more diverse physical tasks, and tracking how competing Chinese models close the performance gap. Adoption will depend on integration into consumer devices or service platforms that can supply images and manuals on the fly. Monitoring OpenAI’s roadmap for vision‑enabled agents and any announced API pricing will be essential for developers planning to build assembly‑assist applications.

OpenAI’s roadmap for reducing inference time on vision‑language models, including potential hardware partnerships or model compression techniques.

Expansion of FAB or similar benchmarks to cover a broader range of physical tasks, which will help validate beyond IKEA furniture.

Competitive progress from Chinese providers such as Kimi K3 and upcoming vision‑enabled models, which could shift the performance landscape.

Announcements of API pricing, licensing, or integration tools that would enable developers to embed GPT‑6 Astra’s capabilities into consumer‑facing applications.

Related guides & quizzes

What is AI?AI Models ExplainedTransformersFuture of AITest what you know — try a free AI quizLook up an AI term in our glossaryFollow the AI model release tracker
Found this useful?