Requesty Review (2026): Pricing, Features & Honest Verdict

TLDR

Requesty is an LLM gateway that puts one OpenAI-compatible endpoint in front of hundreds of models, with failover, prompt caching, budget caps and EU data residency. It is the tidiest managed option for teams that do not want to babysit their own proxy. Best for: engineering teams shipping multi-model AI features. Price: free tier, then a 5% markup on model costs. Rating: 7.3/10.

What is Requesty?

Requesty is an AI gateway. You point your existing OpenAI SDK at router.requesty.ai, swap the API key, and every request now flows through a single router that can reach a catalogue Requesty advertises at 600+ models across roughly 30 providers. OpenAI, Anthropic, Google, Mistral, DeepSeek and the long tail all sit behind one endpoint, one set of logs and one invoice. Nothing else in your code has to change.

The company is based in London and was founded by Daniel Trugman, Thibault Jaigu and Leslie Nooteboom. It raised a $3 million seed round led by 20VC in September 2025. Requesty says it now moves more than 90 billion tokens a day for over 70,000 developers. Those are the company’s own figures, not audited ones. The positioning is easy to read though: Requesty wants to be the governed, European-friendly answer to OpenRouter, with data residency and access controls treated as core product rather than an enterprise upsell.

What Are Requesty’s Key Features?

Routing policies and automatic failover

This is the part that earns its keep. You can distribute traffic across providers by cost, by latency, or by custom weights you set yourself. When a provider throws a 429 or falls over, Requesty reroutes the request to a backup model instead of returning an error to your user. If you have ever had an Anthropic rate limit take down a production feature at 2am, you already understand why this matters. We found the weighted load balancing genuinely useful for splitting traffic between a cheap model and a strong one without writing routing logic yourself.

Prompt caching that runs by default

Requesty caches repeated prompts with no configuration required, and claims this cuts token spend on repeat prompts by up to 80%. That number will only be real for workloads with heavy prompt reuse: RAG systems with fixed system prompts, classification pipelines, coding agents that resend the same context. If every request is unique, ignore the claim entirely. Budget caps and per-key spend limits sit alongside it, so you can hard-stop a runaway agent before it burns a month of budget in an afternoon.

Governance, guardrails and EU data residency

The Enterprise tier adds SSO through Okta, Azure AD, Google Workspace or custom OIDC, full role-based access control with audit logs, an approved-models whitelist, teams and groups with their own spend controls, PII detection, and service accounts for CI/CD. EU data residency is available on the free and pay-as-you-go tiers too, which is unusual. Most gateways make you sign an enterprise contract before they will talk about where your prompts land.

MCP gateway and bring-your-own-keys

Requesty ships an MCP gateway on the pay-as-you-go tier, which lets you expose Model Context Protocol servers to your agents through the same governed pipe as your model calls. That matters more than it sounds: MCP is how most agent stacks now reach internal tools, and it is usually the least supervised part of the system. You can also bring your own provider API keys and keep any negotiated rates or committed-spend discounts you already have with OpenAI or Anthropic, while still getting Requesty’s routing and logging on top.

Observability and SDK coverage

Analytics break spend, latency and error rates down per key, per user, per model and per project. You can tag requests with metadata such as feature name, user ID or trace ID, which makes cost attribution across a product actually possible. Requesty works with LangChain, the Vercel AI SDK, LlamaIndex and Pydantic AI, and it exposes an Anthropic-compatible route at /anthropic/v1/messages for teams already using Anthropic’s SDK.

How Much Does Requesty Cost?

Requesty starts free and then charges a flat 5% markup on model costs. The free tier gives you 200 requests per day against free models only, with routing, caching, fallbacks and EU data residency included and no credit card required. The pay-as-you-go tier opens up the full catalogue plus routing policies, spend limits, budget caps, the MCP gateway and advanced observability. Requesty’s own worked example: a model that costs $10 per million tokens from OpenAI costs $10.50 through Requesty. There are no seats and no minimum spend.

Against the field, that 5% is competitive but not the cheapest possible answer. OpenRouter charges 5.5% with a $0.80 minimum on credit purchases, dropping to 0.5% if you bring your own provider keys. LiteLLM is open source and adds zero markup if you self-host, though you then pay for the servers and the on-call rota. Portkey is free to 10,000 logs a month, then $49 a month for 100,000 logs and $9 per additional 100,000, which is a very different shape of bill: you pay for observability volume rather than a slice of model spend.

One structural detail worth understanding: because Requesty charges a percentage of model spend, your gateway bill scales with your token bill rather than with request volume. That is friendly if you run millions of tiny classification calls on cheap models, and less friendly if you push long contexts through frontier models all day. Portkey’s per-log pricing behaves in exactly the opposite way. There is no single right answer, only the one that matches your traffic shape.

The maths flips at scale. At $2,000 a month of model spend, Requesty’s markup costs you $100 and saves you a week of infrastructure work. At $50,000 a month, the same markup is $2,500 and self-hosting LiteLLM starts looking obviously correct. Enterprise pricing is custom and comes with volume discounts, so if you are in that band, negotiate rather than accepting list.

Who is Requesty Best For?

Use Requesty if you are a small or mid-sized engineering team running real production traffic across more than one model provider, and you want failover, spend caps and per-team access control without hiring someone to run a proxy. It is a particularly good fit if you are in the EU or selling to EU customers, because data residency is available from the free tier up rather than gated behind a sales call.

It also suits agencies and consultancies building AI features for several clients at once. Per-project spend tracking, request tagging and separate keys mean you can attribute cost per client without inventing your own accounting layer, and the absence of per-seat pricing means adding a junior developer to a project does not cost you anything extra.

Skip Requesty if any of the following are true:

• You call exactly one model from one provider and have no plans to change. The gateway is overhead you do not need.
• Your model spend is deep into five figures a month and you have DevOps capacity. Self-hosted LiteLLM will be cheaper.
• You need deep prompt management, evals and experimentation tooling as the main event. Portkey and dedicated LLMOps platforms go further there.
• You are prototyping over a weekend and just want the widest model list with the least friction. OpenRouter is simpler.

Best Requesty Alternatives

OpenRouter

The obvious comparison and the category’s default. OpenRouter is a model marketplace first: enormous catalogue, dead simple to start, and a set of free models with daily rate limits for prototyping. Pricing is a 5.5% fee on credit purchases with a $0.80 minimum, falling to 0.5% when you bring your own keys. What you give up is governance. Requesty’s RBAC, audit logs, approved-model lists and PII detection are more complete. Pick OpenRouter for speed of experimentation, Requesty when someone in compliance starts asking questions.

LiteLLM

The open-source option, and the right one if you have infrastructure people. LiteLLM’s proxy is free to self-host with no markup at all, supports the same OpenAI-compatible pattern, and has a large provider list. The enterprise tier with SSO and support is quoted on request. The trade is real: you own the uptime, the upgrades and the 3am pages. Below roughly $2,000 a month of model spend, the hosting and maintenance cost usually exceeds what Requesty’s 5% would have charged you.

Portkey

Portkey overlaps with Requesty on routing and caching but leans harder into observability, prompt management and guardrails. Its free tier covers 10,000 logs a month, paid starts at $49 a month for 100,000 logs with $9 per extra 100,000, and its gateway is open source if you want to self-host that layer. Because Portkey bills per logged request rather than as a percentage of spend, it can be cheaper for expensive models and pricier for chatty, cheap ones. Model the two against your actual traffic before choosing.

Final Verdict: Is Requesty Worth It?

Requesty does a narrow job well. The routing and failover are solid, the spend controls are the kind you actually want when an agent loop misbehaves, and shipping EU data residency on the free tier is a real differentiator rather than marketing. For a team that wants a managed gateway with grown-up access controls and does not want to run infrastructure, it is a straightforward yes.

What we would watch is the catalogue and the caching. The advertised model count on the marketing site is larger than the figure in the documentation, which is a small thing but the kind of small thing worth checking against the models you actually need before you migrate. Likewise, treat the caching savings as a hypothesis to test rather than a promise, because they depend entirely on how much of your prompt payload repeats.

The reasons it lands at 7.3 rather than higher are competitive, not technical. OpenRouter has a bigger catalogue and more mindshare, LiteLLM is free if you can operate it, and Requesty is a young company on a $3 million seed, which is worth weighing if you are about to route all your production traffic through it. Start on the free tier, run a real workload through the router for a week, and check whether the caching and failover numbers hold up on your traffic before you commit.