Chinese and open-weight AI models dominate developer usage on OpenRouter, data shows

·1 views

A month-long daily tracking of OpenRouter's public usage rankings reveals that the top five most-used AI models by token volume are either Chinese-developed or open-weight, with DeepSeek V4 Flash leading at 4.72 trillion tokens per week. OpenRouter is a neutral API marketplace where developers pay per token and freely choose any model, making its rankings a real-world signal of developer demand. Notably, the first OpenAI model appears at rank 12 and the first Google Gemini model at rank 13, though the data excludes first-party consumer traffic from platforms like ChatGPT or claude.ai. A stark pricing gap — DeepSeek V4 Flash costs roughly $0.09 per million input tokens versus $5 for Claude Opus — appears to be a key driver, especially for token-intensive workloads like agent pipelines and batch processing. The analyst concludes that Chinese and open-weight models are now sufficiently capable for production API workloads while costing significantly less than leading US flagship models.

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)

M.Sc. Student Builds Affordable ATS Resume Scanner Starting at €3.99/Month

A master's student developed ClearScan, an ATS resume scanning tool, after finding existing options like Jobscan too expensive during his own job search. The platform allows users to paste a resume and job description to see how closely they match, with transparent scoring that explains the results. ClearScan launched today and has already attracted its first paying customers. A free tier offers two scans per month, while paid plans start at €3.99/month, deliberately priced with students in mind. The tool is now live at clearscan.fyi, with the developer actively seeking feedback from users familiar with ATS screening challenges.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Qualcomm buys Modular for $4B as Cursor acquires Continue and funding rounds surge

The commercial open-source software space saw a wave of consolidation this week, with Qualcomm agreeing to acquire AI infrastructure startup Modular for nearly $4 billion. Cursor, the AI-powered code editor, quietly acquired Continue, an open-source alternative to GitHub Copilot. Elastic also expanded its AI portfolio by acquiring site reliability engineering startup Deductive AI for up to $85 million. On the funding side, DeepSeek closed a $7 billion-plus round, Timefold secured a $13 million Series A for its scheduling optimization platform, and Moonshot AI is reportedly seeking a $30 billion valuation in new funding talks. Additionally, Sentient Foundation pledged $42 million to advance open-source AGI, while Vercel launched an open-source agentic framework called eve and Daytona announced a shift to closed source.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Dev fixes OpenAI Assistants API timeout errors by making limits configurable

A developer discovered that their production AI assistant was crashing with timeout errors during a live client demo due to a hardcoded 60-second polling limit. The OpenAI runs were not actually failing — they were simply taking longer than expected as session history grew, causing the app to quit too early. The fix involved moving the timeout value to an environment variable and extending it to 150 seconds, while also updating the polling loop to handle all five terminal run states. The developer noted that AI workload durations vary significantly by session length, making hardcoded limits unreliable in real-world use. The update was deployed successfully, eliminating the timeout errors in production.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

PromptOT Lets Teams Manage LLM Prompts as Versioned Blocks Without Code Deploys

A developer built PromptOT after a production bug traced back to an undocumented one-line edit inside a 200-line hardcoded prompt caused a support bot to promise incorrect refund timelines. The platform breaks monolithic prompt strings into typed, independently versioned blocks covering role, context, instructions, guardrails, and output format. Each block can be toggled, edited, and rolled back separately via a dashboard, without requiring a new code deployment. Apps retrieve the compiled prompt through a simple REST API call, with variables resolved at fetch time. PromptOT also ships an MCP server with 23 tools, allowing AI assistants like Claude to draft and save prompt changes directly in chat; a free tier supporting three projects and 1,000 API calls per month is available at promptot.com.

0 comments Read more at DEV Community