SShortSingh.
TopTechnologyAI & MLStartupsProgrammingScienceBusinessFinanceCrypto & Web3World
Back to feed
Programming

A 17-line Python script reveals how much your LLM gateway marks up token prices

0
·1 views

LLM gateways simplify multi-vendor AI access by routing requests through a single API endpoint, but their per-token pricing relative to direct vendor rates is rarely compared in one place. A developer has published a short Python script that queries a gateway's models endpoint, applies a user-defined input/output token mix, and calculates the effective cost against published vendor list prices. When tested on August 14, 2026, the script found one gateway charging roughly 15% below list price for both GPT and Claude Sonnet models. For gateways that do not expose pricing via their API, the author recommends dividing last month's invoice by logged token counts to derive a real effective rate. The piece also cautions users to verify gateway compatibility with streaming, tool calls, and usage reporting before assuming a simple base URL swap is sufficient.

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)

Log in to join the discussion and vote.

Log in

Related stories

0
ProgrammingDEV Community ·

Gitoza Stores Task Tickets as YAML in Git So AI Tools Can Check If Code Matches

A tool called Gitoza proposes storing project task tickets as plain-text YAML files directly in Git repositories, rather than in cloud-based project management platforms like Jira. The approach aims to close the gap between ticket status on a board and the actual state of code in a repository, a problem the developers call 'ticket–code drift.' Because the files live on disk alongside the codebase, AI-powered IDE agents such as Cursor or VS Code can read both the tickets and the source code in the same session. This allows developers to ask practical questions—such as whether a specific ticket has already been implemented or whether a bug has already been filed—without querying an external API. Wiki documentation is also stored as nested Markdown files in the same Git repo, keeping planning artifacts version-controlled and locally searchable.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

AI sysadmin at Finnish hosting firm benchmarks five self-hosted S3 servers with live tests

Väinämöinen, an AI sysadmin deployed by Finnish seedbox company Pulsed Media, conducted live benchmarks of five self-hosted S3-compatible storage servers — rclone serve s3, versitygw, MinIO, SeaweedFS, and S3Proxy — on a real customer-grade storage box. Each server was tested individually using the same client, object size, and methodology to ensure fair comparison, with median results reported across three runs. The AI also verified a full Restic backup-and-restore cycle, confirming byte-identical results via SHA-256 checksums rather than relying on vendor claims. S3Proxy was left unscored in the benchmark table after it failed to cold-start within the measurement window three times, with the AI choosing transparency over fabricating an estimate. The project argues that infrastructure documentation built from actual test results — including its gaps — is more trustworthy than conventionally written docs that fill every field with unverified confidence.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

How to Use Claude Code as a Reliable Engineering Collaborator, Not a Prompt Tool

A developer workflow guide published on DEV Community outlines how to get consistent, high-quality results from Claude Code by treating it as a structured engineering collaborator. The approach begins with a CLAUDE.md file that documents project conventions, commands, and architectural boundaries so context does not have to be repeated each session. Before making changes, the agent is asked to trace relevant code paths and identify the smallest responsible component, reducing the risk of implementing fixes in the wrong layer. For high-risk tasks touching areas like authentication, migrations, or public APIs, a read-only planning phase is recommended before any files are modified. The workflow concludes with a verification checklist covering tests, linting, diff review, and edge cases such as idempotency and human handoff points in browser flows.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Developer Builds AI Agent Memory-Sharing Platform, Pivots After Privacy and Design Concerns

A developer set out to explore agent memory by building Monet, a platform allowing AI agents to share learned memories across users and devices. While dogfooding the product, they realized shared agent memories also contained personal traces of their own interactions, raising privacy concerns they had initially overlooked. A discussion with another developer on Dev.to sparked a new hypothesis: rather than passing full conversation transcripts to a model each turn, organizing dialogue into structured memory states could reduce noise and improve response accuracy. Acting on this, the developer rebuilt the system using a Brain_DB-backed memory engine with an MCP layer, naming the main agent Stig, and found that sessions rarely exceeded 20% context usage thanks to reliable memory retrieval. The project then caught the attention of their workplace team, where a similar second-brain approach was already being explored by a colleague using past session distillation.

0 comments Read more at DEV Community
A 17-line Python script reveals how much your LLM gateway marks up token prices · ShortSingh