A benchmark of 15 AI models reveals 99% cheaper, high-speed API options

An independent speed test of 15 models including Step-3.5-Flash and Qwen3-8B provides a clear guide to cutting AI costs, measuring Time to First Token and throughput.

THE ANSWER UP FRONT

This benchmark is for founders whose AI API bills are growing unsustainably. It provides a direct, data-backed list of fast, cheap models for common text generation tasks. Skip this if your product requires absolute state-of-the-art reasoning, as quality was not measured. The bottom line: viable, production-ready models exist that are an order of magnitude cheaper and faster than many popular premium options.

METHODOLOGY

This v0 review analyzes the benchmark results published by a developer on May 20, 2026. The source is a blog post detailing performance tests of 15 different AI models, available at https://dev.to/swift-logic-io218/speed-test-i-found-ai-apis-99-cheaper-than-premium-5cb. The author's methodology was consistent across all models: a prompt requesting a 200-word explanation of recursion was sent 10 times to each model via the https://global-apis.com/v1 endpoint, with tests originating from US East (Ohio) and Singapore. This review covers the author's published data on Time to First Token (TTFT) and tokens per second (tok/s). It does not include independent verification of these numbers, nor does it assess the qualitative output of the models, which was outside the scope of the original test.

WHAT IT DOES

Measures two critical speed metrics

The benchmark focuses on two user-facing speed metrics. First is Time to First Token (TTFT), which measures the delay between a user's request and the start of the model's streaming response. This is critical for making an application feel responsive. Second is tokens per second (tok/s), which measures the sustained generation speed once the response begins. This determines how quickly the full answer is delivered. The author correctly treats these as separate, important factors for user experience.

Ranks models by performance and cost

The core artifact is a data table ranking 15 models from providers like StepFun, DeepSeek, Qwen, and Tencent. The top performer for combined speed was StepFun's Step-3.5-Flash, with a reported 120ms TTFT and 80 tok/s. Other strong performers included DeepSeek V4 Flash (180ms TTFT, 60 tok/s) and Qwen3-8B (150ms TTFT, 70 tok/s). By including the cost per million output tokens for each model, the table allows for a direct comparison of price-to-performance ratios.

Identifies clear cost outliers

The most actionable finding is the identification of extremely low-cost models with competitive performance. Qwen's Qwen3-8B is the standout, delivering a reported 70 tok/s for just $0.01 per million output tokens. This is two orders of magnitude cheaper than the author's premium baseline of $3.00/M. This data provides a concrete alternative for teams looking to slash costs on high-volume, lower-complexity tasks.

WHAT'S INTERESTING / WHAT'S NOT

The data makes the commoditization of capable, small models concrete. The existence of a model like Qwen3-8B at $0.01/M tokens is a massive data point for founders. It fundamentally changes the economics of building and scaling AI features. The benchmark's dual focus on TTFT and tok/s is also pragmatic, reflecting the two distinct components of perceived user speed. The methodology is transparent and reproducible, which lends credibility to the results.

The most significant omission is any measure of quality. The test uses a single, simple prompt ("Explain recursion...") and does not evaluate the correctness, coherence, or instruction-following capabilities of the models. A model can be fast and cheap but useless if its output is poor. The reliance on a single aggregator (Global API) also means the results measure the performance of the model plus the aggregator's routing, not just the model in isolation. This is a reasonable proxy for a real-world setup but not a pure measure of the model's raw performance.

PRICING

The benchmark's pricing data, captured on May 20, 2026, reveals a wide spectrum. The most notable data points are for the high-performers: Step-3.5-Flash is listed at $0.15/M output tokens, and DeepSeek V4 Flash at $0.25/M. The most extreme budget option, Qwen3-8B, is priced at just $0.01/M output tokens. These figures stand in stark contrast to the author's reference point of premium models costing $3.00/M tokens, illustrating a potential cost reduction of over 99%. Input token costs were not included in the analysis.

VERDICT

For founders building features where speed and cost are the primary constraints, this benchmark is an invaluable resource. It proves that for many common tasks, defaulting to expensive, "premium" models is fiscally irresponsible. The data points to specific models like Step-3.5-Flash for balanced speed and cost, and Qwen3-8B for applications where cost is the absolute priority. However, this benchmark is a starting point, not a final answer. The lack of quality assessment means every team must still validate these models against their specific use case before deploying to production.

WHAT WE'D TEST NEXT

A follow-up analysis should focus on what this benchmark omits. First, we would run a qualitative analysis on a basket of prompts representing common business tasks (e.g., summarization, email drafting, simple code generation) to create a speed-vs-quality matrix. Second, we would test a subset of these models directly from their native APIs to isolate any latency introduced by the Global API aggregator. Finally, we would include input token costs to provide a more complete total cost of ownership picture for different workloads.

The investor read

The benchmark is a clear signal of the rapid commoditization of foundation models for a large swath of B2B and B2C use cases. The value is migrating from the model providers to the orchestration and application layers. Companies that can intelligently route requests to the most cost-effective model for a given task in real-time have a significant advantage. This data suggests that the moat for many AI application companies will not be model access, but rather data, distribution, and workflow integration. Any startup pitching a high-margin business built solely on reselling a premium API is a questionable investment.