Coding Agent Pass Rates Hold Across English and Russian Tasks, but Codex Skips Web Search

·5 views

A developer tested whether coding agents perform differently when tasks are written in Russian versus English, running 1050 trials across seven agent setups. Pass rates without a guide were statistically similar, at 63% for English and 56% for Russian, and guides lifted results by comparable margins in both languages. Translating the guide into Russian made no difference compared with the English original. The notable exception was Codex, which searched the web far less often on Russian tasks and failed a forum-only fact task until a guide supplied the missing information. The author concludes a guide closes the same gap regardless of language, while noting the agent may quietly skip a research step for non-English prompts.

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)

Tiny world model runs game simulations on a Raspberry Pi Pico 2 W

A developer has released an open-source project called TinyWorldFormer, which runs a small world model on a Raspberry Pi Pico 2 W microcontroller. The project's code is hosted on GitHub and was shared on Hacker News, where it gained modest attention. The source material provides only a link and basic engagement figures, so details about the model's architecture and performance are not available here. The effort demonstrates how compact generative models can be deployed on very low-cost, low-power hardware.

0 comments Read more at Hacker News

ProgrammingDEV Community ·

Base44 or your own code? What to weigh before building on an AI app builder

Halfcycle, an AI-agent software team that works inside Claude Code, argues that the key question for AI app builders like Base44 is not how fast the first version appears but what you keep when building stops. Base44 bundles hosting, a database, logins and analytics with no terminal required, which suits internal tools but ties apps to the platform. Its documentation says that exported code still relies on Base44 to run, self-hosting is unsupported, and importing code from GitHub is not supported. Halfcycle, which is its own product and so a biased source, positions its tool as a way to have AI build ordinary code the user owns. The article concludes that quick internal tools may suit Base44, while software a business must own or move should be built as standalone code.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

AI video generation costs per second in 2026: prices across 13 models at 720p

At 720p, AI video generation in October 2026 ranges from roughly $0.06 to $0.63 per second of finished footage, depending on the model. Ten-second 720p clips cost about $0.92 on Grok Imagine and up to $3.67 on Kling 2.6, with VEO 3.1 limited to 8-second clips priced at $2.00 or $5.00. The figures come from the authors' own billing formula, which is checked against a platform that runs these models for ad production, rather than from vendors' API rates. The authors argue that headline "from $X per video" prices are misleading because they hide clip length and resolution, and they publish the data as an open CC BY 4.0 dataset. Moving from 720p to 1080p raises costs for some models by 40 to 100 percent, while others charge the same at both resolutions.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Node.js Script Automates Multi-Platform Article Publishing, Cutting Manual Work by About 80%

A developer working on a content-automation project found that publishing one article to Medium, Substack, and Bluesky required manually reformatting files and keeping metadata in sync. Their first fix, a Bash script, failed due to Unicode encoding problems, hard-coded paths, and missing error handling. They replaced it with a Node.js generator that builds platform-specific files from a single template, with Markdown for Medium and Substack and JSON for Bluesky. According to the author, the change reduced manual copy-paste time by roughly 80% and keeps publishing metadata consistent. The approach also resolved a CI failure caused by a metadata flag that had not been updated after a post went live.

0 comments Read more at DEV Community