OpenRouter Alternatives in 2026: LiteLLM, Vercel AI Gateway, Requesty, Portkey and Direct APIs Compared, With What Developers on Twitter (X) Say

Gateways sort into three groups. Hosted routers sell you one balance and one key for hundreds of models (OpenRouter, Requesty, Vercel AI Gateway). Gateways in front of your own keys add logging, caching, budgets and fallbacks but leave billing with each provider (Portkey, Cloudflare AI Gateway, a self-hosted LiteLLM proxy). Inference hosts serve models themselves (Together AI, Groq, Fireworks) or are the model maker (OpenAI, Anthropic, Google, xAI). The table uses each vendor's own wording.

How to read it. A percentage fee only matters relative to your model spend: on $100 of tokens, OpenRouter's 5.5% purchase fee is $5.50, Requesty's 5% markup is $5, and Vercel's is $0 plus whatever your card processor charges. For a gateway in front of your own keys, the cost is the plan (Portkey bills by recorded logs, not tokens) or the server you run (LiteLLM). For an inference host, there is no gateway fee, but you also lose the one-key, many-providers catalog.

Which fees we could not confirm. OpenRouter's FAQ page renders its credit fee as an unfilled template placeholder, so we took the 5.5% and 8% figures from openrouter.ai/pricing. Vercel's pricing page says every team gets a monthly free credit on a subset of models but does not show the amount, so we do not quote one. Portkey's pricing page does not say whether it marks up tokens; its docs have you add provider credentials to a Model Catalog, which means billing stays with your providers.

LiteLLM's docs describe its proxy as a self-hosted server that gives applications one endpoint for 100+ LLM providers and returns responses in OpenAI format: if a client can call the OpenAI API, it can call the proxy. The project is MIT-licensed (docs.litellm.ai/docs/simple_proxy). An Enterprise tier adds SSO, audit logs and multi-team management at custom pricing, with a 30-day self-hosted trial key (litellm.ai/enterprise).

The trade against OpenRouter is ownership against convenience. With LiteLLM, there is no one balance: you open an account with every provider you route to, store those keys in the proxy's config, and pay each provider at its list price. There is no gateway fee, and prompts never pass through a third party other than the model provider. In exchange, you run, scale and secure a server, and you keep provider keys alive on it, which is exactly where one leaked config bites: an unused key left in a running proxy keeps billing until someone notices.

When it wins. Teams with compliance rules about where prompts go, teams already paying several providers directly, and anyone who wants routing, budgets and logging under their own control. When OpenRouter wins. Solo builders and prototypes that want one key, one bill and hundreds of models without creating ten provider accounts.

On X, LiteLLM appeared in 8.9% of the 683 posts we read, and 54.1% of those also talked about price, fees or credits, the highest price share among the seven most-named options.

Vercel AI Gateway. Vercel's pricing doc says the gateway charges no markup and no platform fee on tokens, you pay the provider's list price from prepaid AI Gateway Credits, and you are responsible for any payment processing fees (vercel.com/docs/ai-gateway/pricing). Bring-your-own-key carries no markup either, but it is a paid-tier feature, and when a BYOK request fails, the gateway retries it with Vercel's own credentials and charges that fallback to your credits. The free tier covers a subset of models with lower per-model rate limits (a 429 when exceeded), and buying credits ends the monthly free credit. Optional add-ons are billed separately: team-wide zero data retention at $0.10 per 1,000 requests, a team-wide provider allowlist at $0.10 per 1,000 successful requests, and custom reporting at $0.075 per 1,000 writes and $5 per 1,000 queries.

Its OpenAI-compatible endpoint is https://ai-gateway.vercel.sh/v1 with Authorization: Bearer and an AI Gateway API key (or a Vercel OIDC token on Vercel deployments); chat/completions supports streaming, tool calling and structured outputs. Your app does not have to run on Vercel. If you already build with the AI SDK, model strings like openai/... or anthropic/... route through the gateway directly; our Vercel AI SDK guide shows that path with X data.

Requesty. Requesty's pricing page lists three plans: Free ($0, all free models, 200 requests per day, no card), Pay as you go (a 5% markup, 600+ models, bring your own keys) and Enterprise (custom). Its own example: a model that costs $10 per 1M tokens from OpenAI costs $10.50 through Requesty (requesty.ai/pricing). The quickstart gives https://router.requesty.ai/v1 (and an EU endpoint at router.eu.requesty.ai/v1), Bearer auth and provider/model IDs; a bare model name lets Requesty pick the provider per request.

Versus OpenRouter. OpenRouter's pricing page puts its fee on the credit purchase (5.5% Standard, 8% Business, never on individual requests), lists 600+ models from 80+ providers, sets no plan-based rate limit on paid models, and gives BYOK users $25,000 of list-price inference a month with no fee, then 5%. So for a team spending heavily on its own provider keys, OpenRouter's BYOK allowance can cost less than Requesty's markup, while for a team buying credits, Vercel's zero markup is the lowest stated fee of the three.

Portkey and Cloudflare AI Gateway are less about the model catalog and more about what happens around each call: logs, caching, retries, fallbacks, guardrails and budgets.

Portkey. The pricing page lists an Open Source self-hosted gateway, a Developer plan that is free with 10K recorded logs a month (3-day log retention, labeled not suitable for production), Production at $49 a month with 100K recorded logs and $9 per additional 100K, and custom Enterprise. Exceeding the log limit does not block requests; logs past the limit are simply not recorded (portkey.ai/pricing). The REST base URL is https://api.portkey.ai/v1, and the OpenAI SDKs work by changing the base URL and adding Portkey's headers.

Cloudflare AI Gateway. Cloudflare's pricing reference says the core features (analytics, caching, rate limiting) are free; Workers Free accounts keep 100,000 legacy logs in total, Workers Paid 10,000,000 per gateway; and if you buy credits through Unified Billing instead of using your own provider keys, a 5% fee applies, so a $100 credit purchase is a $105 charge (developers.cloudflare.com/ai-gateway/reference/pricing). Its OpenAI-compatible endpoint takes the form https://gateway.ai.cloudflare.com/v1/{account_id}/default/compat/chat/completions with {provider}/{model} IDs and an optional cf-aig-authorization header.

In our X sample, Cloudflare came up in 9.1% of posts, and many of those were about Workers AI free allowances rather than the gateway: 51.6% mentioned free models or a free tier. Portkey appeared in 3.1%.

If your app has settled on two or three models, a router can be one hop you do not need. Inference hosts publish OpenAI-compatible endpoints of their own: Together AI at https://api.together.ai/v1 (its compatibility matrix lists tools as supported, and model IDs follow <provider>/<model_name>), and Groq at https://api.groq.com/openai/v1, where logprobs, logit_bias, top_logprobs and messages[].name return a 400 and n must be 1 (console.groq.com/docs/openai).

Going direct to the model maker (OpenAI, Anthropic, Google, xAI) gives list price, the newest features first and one data processor fewer. The costs are operational: a key and a bill per provider, your own retry and fallback code, and no single place to compare spend. A pattern that shows up in the posts we read is a hybrid one: a router while you explore models, then direct contracts once volume concentrates on one provider.

Groq was the single most-named option in our sample (21.7% of posts), and its context was mostly free access: 53.4% of Groq posts mentioned free models or a free tier and 27.0% rate limits or quotas, which is what free-tier lists look like. Local or self-hosted models (Ollama, LM Studio, llama.cpp, vLLM) were second at 15.4%, and 52.4% of those posts were about control or self-hosting. The provider's own API by name appeared in 3.5%.

We ran two searches on X: one for OpenRouter next to switching words (alternative, instead of, switched, moved to, migrate, leaving), the other for OpenRouter next to a named rival (LiteLLM, Portkey, Requesty, "ai gateway", Together, Helicone, Fireworks, Groq, "direct api"). After removing posts that were only a link or a mention, 683 posts remained: 244 from the switching search and 479 from the rival search (a post can be in both). Keyword rules then tagged the alternatives named and the reasons given.

Money first, then reliability. Price, fees or credits appeared in 49.2% of switching posts against 41.3% of all posts, the largest single reason. Reliability, errors and outages rose the most between the two groups, from 15.7% overall to 22.5% when someone was talking about switching. Latency followed at 20.9%.

Privacy is quieter than the guides suggest. Comparison articles often lead with data retention and compliance, but only 7.8% of switching posts mentioned privacy, data retention or compliance, and BYOK just 3.3%. Free models pulled the other way: they came up more in the overall sample (23.7%) than in switching posts (18.4%), which fits OpenRouter's free models being something people stay for.

Who is talking. The median post had 4 likes and 555 views; 18.9% were replies. English was 72.3% of posts, Chinese 10.7% and Japanese 2.4%; 11.7% had no detectable language (link or media only).

Across the 683 posts, 61.1% named at least one alternative, and 70.7% of those also gave a reason. Because the second search was built around rival names, these shares show what is discussed next to OpenRouter, not market share.

Official accounts at collection. These are organization accounts, so we name them. Follower counts say more about each company's wider audience than about its gateway: Vercel's account covers its whole platform, and Together AI's covers its inference cloud.

OpenRouter's account (157,733 followers) is more than 25 times the size of LiteLLM's (6,160) and over 150 times Requesty's (959), which matches what the posts show: most builders meet OpenRouter first and compare from there.

Whichever gateway you choose, the way a model gets fresh X data is the same: you define a tool, the model asks for it, your code calls an API and passes the result back. OpenAI-style tools on chat/completions is part of the format these gateways accept (Vercel and Together document it explicitly; check your gateway's docs for the model you pick), so one search_x function works across them; only the base URL, key and model ID change. The full script is in the code block at the end of this page.

The tool. search_x(query, max_tweets) calls GET https://api.twitterapi.io/twitter/tweet/advanced_search with the X-API-Key header, query in X's search syntax and queryType=Latest, follows next_cursor while has_next_page is true, and returns compact rows (author.userName, text, likeCount, createdAt, url). Reads need no X account and no X developer app.

The loop. Send the conversation with tools; if the reply has tool_calls, run each one, append a role: tool message with the matching tool_call_id, and call the gateway again; stop when the model answers in plain text or after a few rounds. The pattern is the same one our OpenAI Agents and Claude tool-use guides use, just pointed at a gateway.

Or use MCP. If your client speaks MCP (Claude Code, Cursor, Codex and others), our hosted server at https://mcp.twitterapi.io/mcp exposes the same search as ready-made tools with your API key as a Bearer token, so you do not write the tool at all; the MCP server guide has the setup per client.

Writes, if the agent should post. Posting runs through an authenticated session: one /twitter/user_login_v2 login ($0.10, with a proxy you supply) returns login_cookies, then each /twitter/create_tweet_v2 call is 60 credits ($0.0006), billed per HTTP 200 call. The call acts as that account, so X can rate-limit, lock or suspend it; keep a human pace and expect the session to need a re-login now and then.

The X data bill next to the gateway fee. Gateway fees are a percentage of model spend; X data is a separate meter, and it is usually the one that scales with what the agent does. TwitterAPI.io charges 15 credits per tweet returned, $0.15 per 1,000 tweets, with a 15-credit minimum per request: a full 20-tweet page costs $0.003, and a search that finds nothing costs the $0.00015 minimum. The official X API's pay-per-usage rate card (docs.x.com/x-api/getting-started/pricing) lists $0.005 per post read, $5.00 per 1,000, about 33 times more, with pay-per-usage capped at 3 million post reads a month before Enterprise.

Put next to the gateway. On $100 of model spend, the gateway fee is $0 on Vercel (plus payment fees), $5 on Requesty and $5.50 in OpenRouter's credit fee; the same agent reading 10,000 tweets adds $1.50 of X data. Profiles are 18 credits per user ($0.18 per 1,000) against $0.010 per user read on the X rate card. New accounts that pass the automatic bot check get up to $1 of free credits (at most 100,000 credits, no card, one per person or network, expiring after 31 days); purchased credits never expire. Rate limits are 0.2 requests per second before the first payment and 20 once paid; plans are on the pricing page.

Cost lever. In a tool-calling agent the model decides how many tweets to fetch, so cap max_tweets in the tool schema and in your code. That one number bounds both the X bill and the tokens you pay the gateway to read the results.

Question. When developers on X mention OpenRouter next to a switching word or a rival gateway, which alternatives do they name and which reasons do they give?

Queries. Two advanced_search queries with -filter:retweets: OpenRouter plus switching words (alternative, instead of, switched, switching, moved to, migrate, ditched, leaving) and OpenRouter plus rival names (LiteLLM, Portkey, Requesty, "ai gateway", Together, Helicone, Fireworks, Groq, "direct api"), each as Latest and Top. Plus /twitter/user/info for six organization accounts: OpenRouter, LiteLLM, PortkeyAI, RequestyAI, togethercompute and vercel.

Window and sample. The switching search's Latest results spanned about 9.6 days, the rival search's about 37.8 days, both collected in 2026. After de-duplication, 779 posts from 627 distinct authors; 683 had at least 15 characters of text once links and mentions were removed, and all shares use that base. 55 API requests in total.

Method. Keyword rules on post text tag each post with the options it names and the reasons it gives; a post can carry several labels. The script, protocol and de-identified rows are in data/studies/openrouter-alternatives/ in our repository.

What the sample cannot show. Keyword rules miss sarcasm and context (a post can mention fees to praise OpenRouter), the rival search pulls toward the names in it, and posts are not usage: none of this measures which gateway has more customers. Gateway fees on this page come from vendor pages, not from the sample.

Privacy. Aggregates only. Search results come from private individuals, so raw pages were not kept in our repository; the committed rows carry no post IDs, URLs, text, author IDs or handles. Only organization accounts are named.