Cheap LLM API routes reported inflated input token counts for a fixed prompt

·1 views

A daily check sent the same roughly 90-token prompt to 107 model routes across four hosts on 2026-10-12. Many routes reported input counts in the thousands, with some GPT routes near 4,469 tokens and two GLM routes at 14,111. The author suggests hosts may be prepending hidden system prompts that the customer is billed for on each request. No host is named, and the author notes other explanations are possible. Developers can test their own providers with a one-token request and compare the reported prompt_tokens against a baseline.

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)

Sloth Sort: local Gemma 3 tool automatically renames messy screenshots and downloads

Sloth Sort is an open-source tool that watches the Desktop and Downloads folders for files with generic names like screenshots and scans. When such a file appears, Gemma 3 running locally through Ollama identifies its contents and renames it with a date-first, descriptive name. Only throwaway names are changed, and every rename can be reversed from a review page on the local machine. The model runs entirely on the user's computer so sensitive screenshots never leave it, and it unloads after five idle minutes. The project was submitted to the Hacktoberfest Open-Source AI Challenge.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Asynchronous JavaScript: How Non-Blocking Operations Work

Some JavaScript tasks, such as fetching data from an API or waiting on a timer, take time to finish. Asynchronous JavaScript lets the language handle these tasks without halting other code from running. Synchronous code executes line by line in order, while asynchronous code can schedule work, like a setTimeout callback, to run later. The article's example shows a log statement printing before a delayed callback, illustrating this behavior. Common uses include fetching API data, loading products on shopping sites, sending login details, uploading files, and communicating with databases through a backend.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

ShortSingh: Android app's model search now hides models your phone cannot run

Off Grid AI Mobile (OGAM) has added a filter to its Android model search so users avoid downloading AI models their phones cannot load. Results are filtered by the phone's RAM, showing only models that will load and marking those that only just fit as Tight. Each model appears once even when many uploads exist, and models needing a newer engine, speed-up helper files and image generators are left out of chat search. The changes arrived in versions 0.0.116 and 0.0.117, and search is free with no account required. Once downloaded, models run on the phone without an internet connection.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Arrange My Files AI Builds Reviewable Plans to Organize Documents Without Touching Originals

Arrange My Files AI is a tool that turns cluttered folders into organization plans users can review before anything is copied. Original files stay in place, and the included batch is copied into a new output session. Files with clear content go into category folders, while uncertain or unsupported files are placed unchanged in a Needs Review folder. The author tested the app with nine dummy files, where the engine renamed some files based on document titles while preserving descriptive names. The developer notes that the system performs quick, bounded analysis rather than full document understanding, so some files still require manual review.

0 comments Read more at DEV Community