SShortSingh.
TopTechnologyAI & MLStartupsProgrammingScienceBusinessFinanceCrypto & Web3World
Back to feed
Programming

Developer Tests 8 AI APIs on 50 Prompts: Cheapest Is 26x Less Than Most Expensive

0
·2 views

Yu Kan, developer of the NovAI API gateway, ran 50 standardised prompts across 8 AI APIs in July 2026 to compare real-world costs and output quality. The cheapest model, Qwen-Plus via NovAI, cost just $0.011 for all 50 prompts, while Anthropic Claude Sonnet 4.6 was the most expensive at $0.287. Despite the 26-fold price difference, the quality gap between the cheapest and most expensive models was only 0.8 points on a 5-point scale. Testing covered five task categories including coding, reasoning, writing, translation, and long-context summarisation, with token counts and latency recorded from actual API responses. The author also confirmed that routing DeepSeek through the NovAI gateway added no cost overhead and only a negligible 42ms latency compared to DeepSeek's official API.

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 ·

Microsoft's Go-Based TypeScript Compiler Delivers Up to 10x Speed Gains in TypeScript 7

Microsoft has ported the TypeScript compiler from TypeScript to Go under Project Corsa, delivering performance improvements of roughly 5x to 12x depending on project size, hardware, and workload. The decision followed extensive evaluation of multiple languages and architectures, including hybrid approaches that would keep only performance-critical parts in native code. Go was chosen primarily for compatibility reasons, as both the existing and new compilers need to coexist and share fixes during a transition period. The team also studied native tools like swc, oxc, and esbuild before concluding that a full native port would yield the greatest long-term benefits. TypeScript 7 is expected to bring these gains to developers, though the team cautions that real-world results will vary and some caveats remain before upgrading large codebases.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Eight newsletter tools compared: costs, mechanics, and which problem each actually solves

A detailed breakdown published on 2 August 2026 compares eight newsletter management tools, noting that prices and behaviours were verified against each vendor's own website on that date. The analysis draws a sharp distinction between two separate problems: inbox clutter, which requires routing newsletters elsewhere, and volume overload, which requires condensing or filtering content. Four core technical approaches appear across the tools, including dedicated intake addresses, mailbox connections via OAuth, email-to-Atom conversion, and forwarding filters. Kill the Newsletter stands out as the only tool that converts incoming emails into a plain Atom feed URL, though its FAQ warns that some publishers block relay addresses and recommends a forwarding workaround. The author discloses upfront that one of the eight tools reviewed is their own product, describing it as new and unproven, while presenting the remaining seven as established options for most readers.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Silent 5,000-row API cap caused dashboard to silently misreport search rank data

A developer discovered that an internal search-rank dashboard was silently displaying incorrect data due to a hard 5,000-row cap in a generic query endpoint, while the actual table contained 9,104 rows. The API returned a normal 200 response with no truncation warning, causing 4,104 rows to be dropped without any visible error. Because rank-change calculations depended on comparing two full time windows of data, the missing rows caused the dashboard to render dashes instead of meaningful changes. The fix involved shifting data aggregation to SQL using Postgres DISTINCT ON, returning one row per keyword rather than raw daily rows, keeping response size bounded by keyword count. The developer outlined key lessons: silent caps are silent lie generators, LIMIT without ORDER BY is nondeterministic sampling, and missing values must never render identically to zero-change values.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Fluttorch 1.0 catches silent accuracy loss when ML models are quantized for mobile

When machine learning models are exported and quantized for mobile deployment, their accuracy can quietly degrade without any build or runtime error flagging the problem. A related issue arises when preprocessing logic is written separately in Python for training and in Dart for serving, allowing the two versions to silently diverge over time. Developer Fluttorch addresses both gaps with a manifest-based system that records model shapes, data types, preprocessing constants, and a content hash at export time, preventing the Dart runtime from restating or contradicting those values independently. The tool's code generator, parity gate, and runtime all read from the single manifest, turning previously silent mismatches into compile-time errors. Version 1.0.0 of Fluttorch launched this week, and its test suite replays export-time predictions against the quantized on-device artifact to verify the two models still agree.

0 comments Read more at DEV Community