AI Streaming Bug Can Deliver Full Responses While Logging Zero Billable Tokens

·1 views

A billing vulnerability in AI streaming systems can cause a full response to be delivered to users while the client logs zero output tokens, making a paid API call appear free. The issue arises because streamed AI responses send content first and token-accounting data last, so if the final usage frame is dropped or malformed, no billing record is created. A developer has built an open-source Python tool called stream_billing_gate.py that reconciles delivered text against logged usage data offline to detect such discrepancies. The tool flags cases where text was clearly delivered but token count shows zero, exiting with a 'delivered-but-unbilled' error. The problem requires no API key to reproduce and can be tested against recorded stream logs, making it accessible for developers to audit their own systems.

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)

dataclass, TypedDict, or Pydantic: Choosing the Right Tool for JSON in Python

When converting JSON to typed Python objects, developers have three main options: dataclass, TypedDict, and Pydantic, each serving a distinct purpose. Dataclasses offer attribute-style access for trusted internal data but perform no runtime validation. TypedDict keeps data as plain dictionaries and provides type hints solely for static checkers, with no runtime enforcement. Pydantic, by contrast, both parses and validates data at runtime, making it the recommended choice for untrusted sources such as API payloads or webhooks. A key caveat across all three approaches is that auto-generated classes infer types from a single JSON example and may miss nullable fields, optional keys, or union types, so generated output should be reviewed before use.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer launches SheetSync, an open-source alternative to Google Sheets

A developer has released SheetSync, a free and open-source spreadsheet application designed as an alternative to Google Sheets. The app is built using Next.js, TypeScript, and Supabase for its backend and database infrastructure. SheetSync includes features such as formulas, cell merging, data validation, charts, and PDF and CSV export capabilities. It also supports organization and workspace management with invite links, along with pre-built templates for common use cases. The project is publicly available on GitHub, where the developer is welcoming community contributions and feedback.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

40-Year IT Veteran Argues Great Software Is Craftsmanship, Not Just Code

A software professional with four decades of experience contends that building truly effective software goes beyond technical skill and constitutes a form of craftsmanship akin to art. Drawing an analogy between a technician and an artisan, the author argues that knowing when to break design patterns and making a system elegant — not merely functional — cannot be taught in a course. The piece traces how knowledge-sharing in IT evolved from human mentors to books, then the internet, and now AI-powered 'Vibe Coding', which has lowered the barrier between an idea and a working solution. While acknowledging Vibe Coding as a genuine evolution, the author maintains that creativity, intuition, and the ability to envision what could exist remain distinctly human qualities. The central argument is that two developers given identical requirements will produce vastly different results, and the difference lies not in the brief but in the depth of craft each brings to the problem.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Why Using Too Many AI Tools Can Hurt Developer Productivity

A developer with experience building AI systems across multiple projects argues that workflow complexity, not tool scarcity, is the real productivity bottleneck for AI builders. Constantly adopting new tools introduces hidden learning costs, as mastering each platform's configuration, prompt behavior, and integrations takes far longer than the initial setup. Frequent context-switching between multiple applications fragments attention and undermines deep, focused work. The author contends that a moderately capable tool that integrates smoothly into an existing workflow often delivers more value than a more powerful one that creates daily friction. Rather than chasing every new release, the piece advocates building a lean, well-integrated toolset and treating reusable prompts as structured software assets rather than scattered chat history.

0 comments Read more at DEV Community