GPT-6 Astra dominates easy robot task but ties Claude on harder one, tests show

·1 views

Independent evaluator Robocurve tested OpenAI's GPT-6 Astra and Anthropic's Claude Fable 5.1 on physical robot arms in September 2026, publishing results on consecutive days. Astra completed a simple block-picking task 19 out of 20 times, while Fable 5.1 managed only 8 out of 20. However, both models performed identically on the harder puzzle-placement task, each succeeding just 2 out of 20 trials. Astra also proved faster and cheaper per run on the block task, averaging 2.5 minutes at $0.94 compared to Fable 5.1's 6.8 minutes at $2.12. Robocurve, a Y Combinator-backed Public Benefit Corporation, used an open-source evaluation framework with standardised hardware and human graders, allowing results to be independently reproduced and verified.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Websheet: Open-Source Front-End Library for Rendering Excel Files in Browser

Websheet is a front-end spreadsheet display control built on HTML5 Canvas and JavaScript that enables Excel (XLSX) file rendering directly in the browser. The library supports a wide range of features including file import and export, cell formatting, merged cells, formula parsing, and row or column management. Developers can interact with it programmatically through scripting support for sheet operations, view freezing, table management, and print functionality. It also exposes a set of events such as cell value changes and sheet switches, and allows customization of functions, display controls, and right-click menus. The project is available for non-commercial use, with commercial licensing requiring prior permission from the developer.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

iOS Build Tools Explained: xcodebuild, Fastlane, CocoaPods, XcodeGen, and KXApp

iOS app development relies on a range of build tools, each serving distinct roles in compilation, packaging, signing, and deployment. xcodebuild is Apple's command-line tool bundled with Xcode, widely used in CI/CD pipelines via platforms like Jenkins and GitHub Actions, though it requires Xcode on a Mac and involves complex parameter configuration. Fastlane, built on top of xcodebuild, automates workflows for building, testing, signing, and releasing apps, while also managing certificates through its 'match' feature. CocoaPods and Swift Package Manager handle third-party dependency integration, and XcodeGen replaces manual project file editing with YAML-based configuration to reduce merge conflicts in team settings. KXApp takes a different approach by bundling its own iOS compilation toolchain, enabling Swift and Objective-C builds without a separate Xcode installation, though it may still fall short for complex projects with heavy CocoaPods dependencies.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How Repetitive KYB Compliance Burdens Are Stalling Early-Stage Startups

Early-stage startups in regulated sectors like fintech and Web3 face repeated Know Your Business (KYB) verification demands each time they onboard a new partner or investor, consuming time that could otherwise go toward product development. This cycle, often called the 'repetition tax,' does not improve security but slows down operations significantly. Centralised compliance databases that store sensitive company and director data present additional risk, as a single breach can expose thousands of businesses simultaneously. Founders also report 'compliance fatigue,' an emotional and creative toll stemming from constant administrative overhead despite following all required procedures. Critics argue the current compliance framework was designed for large institutions and is poorly suited to small, fast-moving teams that operate on sprint timelines rather than fiscal quarters.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Craigslist JSON-LD lacks ID fields, forcing title-based joins with 89% success rate

Developers building a Craigslist multi-city listings scraper discovered that the site's JSON-LD blocks contain no shared identifiers such as product IDs, URLs, or SKUs to match against static HTML listings. The only common field available for joining the two data sources is the listing title, despite titles not being unique across posts. Using a per-title FIFO queue approach, engineers were able to successfully match 290 of 325 listings — an 89% recovery rate — on a captured 298-item search page. Listings that could not be fully matched still returned partial data including title, price, and location, with fields like coordinates, currency, and images left null rather than dropping the row entirely. Separately, Craigslist was found to have migrated from numeric post IDs to alphanumeric strings in a new URL scheme, meaning any scraper relying on legacy numeric-only regex patterns will silently fail on current pages.

0 comments Read more at DEV Community