EvoLink is an AI model gateway that puts LLM, image, video and audio models from many providers behind a single OpenAI-style API, one balance and one dashboard, so engineering teams can switch models without rewriting integration code.
Used this tool? Rate it
Used this tool? Rate it
EvoLink is an AI model API platform built for teams that have already moved past the prototype stage. Its own positioning statement is unusually precise about this: the site describes itself as an AI Model API Platform for Production, and invites developers to access leading LLM, image, and video models through one EvoLink API. That single sentence contains the entire product thesis. Instead of maintaining a separate account, billing relationship, SDK and failure mode for every model vendor an application touches, a team points its code at one endpoint and treats model choice as a configuration value rather than an architectural commitment.
The platform is operated by EVO GLOBAL TECHNOLOGIES LIMITED, a company founded in 2025 and based in Hong Kong, whose stated mission is to be the model access layer for AI teams shipping to production, summarised on the site as One API. Every model. Production ready. The legal terms confirm the same entity and give a full registered address in Mong Kok, Hong Kong, with Hong Kong law governing the agreement and Hong Kong courts holding exclusive jurisdiction over disputes. For a category where many aggregators publish nothing but a landing page, that level of entity disclosure is worth noting up front.
It helps to be clear about what EvoLink is not. It is not an end-user creative tool: there is no editor, no canvas, no timeline. The primary calls to action on the homepage are to get an API key and to read the docs. The people who benefit from it are the ones writing the code that calls a model, plus the people who have to explain the resulting bill. Everything below should be read through that lens.
EvoLink uses pure usage-based pricing with no subscription layer at all. The pricing page states plainly that there is no subscription, no seat fee, and no minimum monthly spend, and that a single balance covers every model: one balance, every model, with usage metered per call and visible in the dashboard in real time. The minimum top-up is $10, and account and key creation requires no credit card, so evaluation costs nothing until a call is actually made.
Billing units follow the modality rather than a single abstract credit: video and audio bill per second, images per image, and LLMs per token. The model catalogue publishes unit prices per model, which vary widely across the range — from fractions of a cent per image for lightweight models to several dollars per million tokens for frontier language models. One operationally useful detail is the failure policy: if a generation fails due to an upstream error, timeout, or capacity issue, it never hits your balance. Volume discounts are described for teams at roughly the $1,000-per-month level and above. The vendor also states that its prices sit below official provider APIs because it buys capacity at volume rates and passes most of the discount on; that is a commercial claim from the vendor rather than an independently verified benchmark, and it should be tested against your own workload. Note also that all purchases are final and non-refundable, including unused balance.
The unified-gateway category is real and increasingly crowded, and the honest comparison is about trade-offs rather than a ranking. The most common alternative is going direct to each model provider: you get first-party support, day-one access to new releases and a contractual relationship with the vendor, at the cost of maintaining a separate integration and invoice for every provider you use. Other aggregation layers occupy the same conceptual position as EvoLink, competing on breadth of catalogue, routing behaviour, pricing transparency and maturity. Media-focused inference hosts specialise in image and video workloads and often expose richer model-specific parameters, but cover language models less comprehensively. Self-hosted routing libraries and open-source gateways offer maximum control and no vendor markup, in exchange for the operational burden landing on your team.
Two evaluation criteria matter more than feature lists here. First, track record: EvoLink was founded in 2025 and, as of this writing, has no independent coverage from established technology media, so its reliability claims rest on self-reporting. Second, exit cost: because the interface follows OpenAI-style conventions, the same property that makes adoption cheap also makes leaving cheap, which is a genuine argument for trying it on a subset of traffic. Weigh a young operator's convenience against the maturity of a direct provider relationship according to how critical the workload is.
There is no free tier in the sense of free inference, but there is no cost to evaluate the platform either. Creating an account and issuing an API key requires no credit card, and model playgrounds allow previews before purchase. Actual generation draws down a prepaid balance, with a minimum top-up of $10. There is no subscription, no seat fee and no minimum monthly spend, so the total cost of an idle account is zero.
Usually not extensively. The documentation states that the platform is compatible with OpenAI-style interfaces and mainstream SDKs, making migration from existing applications easier, so for many teams the change amounts to a new base URL and a new key. That said, compatibility in the general case is not the same as parity on every parameter, so test the specific endpoints, streaming behaviour and provider-specific options your application relies on before cutting traffic over.
The catalogue spans four modalities — text, image, video and audio generation — drawing on providers including Google, Anthropic, OpenAI, BytePlus, Alibaba, xAI, DeepSeek, Midjourney and Suno. The exact roster changes as models are released and retired, so check the models directory for current availability rather than relying on a published total.
Billing is per call against a single prepaid balance, with units that follow the modality: video and audio per second, images per image, and language models per token. Each model in the catalogue carries a published unit price. Usage is metered per call and visible in the dashboard in real time. Notably, failed generations caused by upstream errors, timeouts or capacity limits do not draw down the balance.
No. The terms are explicit that all purchases including top-ups are final and non-refundable, and that unused credit balances are not refunded under any circumstances. Filing a chargeback triggers immediate and permanent account termination. The practical consequence is to top up in amounts matched to near-term consumption rather than pre-funding a large balance.
Not in the published terms. The 99.9% figure is described as observed uptime in the vendor's own material, and multi-channel routing with automatic failover is presented as an architectural feature rather than a contractual service level. Teams with hard availability requirements should design their own fallback rather than treating the number as a guarantee.
The published privacy policy does not say. It documents collection of account data, usage data including API request logs and model consumption records, transaction data and technical data, and it names Stripe and Cloudflare as third-party providers — but it does not address whether prompt content or generated outputs are retained or used for training. If your workload involves confidential input, treat this as an open question to resolve with the vendor directly before integrating.
The terms prohibit illegal activities, generating sexually explicit content (NSFW), and promoting violence. They also prohibit reverse engineering, scraping, and reselling the API without permission. Separately, individual upstream providers apply their own policies, so a request that the gateway permits may still be refused by the model behind it.
They run asynchronously. The API operates in asynchronous processing mode: the initial call returns a task ID, and the caller polls for the finished result rather than waiting on an open connection. Generated image links remain valid for 24 hours, so completed assets should be downloaded into your own storage as part of the job rather than referenced by URL later.
The platform is operated by EVO GLOBAL TECHNOLOGIES LIMITED, founded in 2025 and based in Hong Kong, with a full registered address published in the terms of service. The agreement is governed by Hong Kong law with exclusive jurisdiction in Hong Kong courts. Because the company is young and has no independent technology media coverage to date, its operational claims currently rest on self-reporting.
