LiteLLM alternatives

9 tools we tested head to head against LiteLLM, ranked — and what each one actually does differently.

last reviewed 17 aug 2026 · from our best 10 llm routers and ai gateways ·list curated by Onur Ozcanx in

first — what you'd be leaving

LiteLLM ranks #6 of 10 in our llm gateways testing. the widest provider coverage, with the governance behind a paywall.

78/100

the broadest coverage of any self-hostable gateway, but the features most teams deploy it for sit under a proprietary licence with no published price.

why people look for an alternative

−sso, rbac, audit logs and multi-tenancy are all enterprise-licensed
−enterprise pricing is quote-only, sold per seat
−the open-core split means an unlicensed production rollout is easy to stumble into

stay with LiteLLM if 100+ providers behind one openai-format interface is the thing you care about most — nothing below beats it on that.

the short version

best alternativeOpenRouterteams who want one key across everything and no gateway bill at all90/100 best free optionBifrostlatency-sensitive proxying where the gateway must not be the bottleneck72/100

OpenRouter

#1 in llm gateways · the default, and the only one whose catalogue is the product

90/100

verdictgenuine pass-through pricing on inference with no subscription, undercut slightly by a card fee that most comparisons never mention.

OpenRouter vs LiteLLM

	LiteLLM	OpenRouter
price	free to self-host — enterprise is quote-only	$0 — no platform fee
free tier	yes	yes
markup on tokens	none — self-hosted	none — 5.5% on card top-ups
platform fee	quote only	none
self-host	yes — mit core, proprietary enterprise/	no
models carried	100+ providers	not published

switch forteams who want one key across everything and no gateway bill at all

pros

+true pass-through on inference — anthropic list price, unchanged
+no subscription and no platform fee at any volume
+the broadest practical catalogue in the category
+usable free tier of open-weight models with documented rate caps
+byok supported for teams that want to keep their own provider relationships

cons

−5.5% card fee on credit top-ups undercuts the no-markup claim
−byok is charged 5% above the monthly allowance
−no self-host, and governance features are undocumented

visit OpenRouter →full OpenRouter review →

Portkey

#2 in llm gateways · mit-licensed, and billed per log rather than per token

87/100

verdictthe only gateway here with a complete feature set under a genuinely permissive licence, priced in a unit that decouples your gateway bill from your token spend.

Portkey vs LiteLLM

	LiteLLM	Portkey
price	free to self-host — enterprise is quote-only	$49/mo (production)
free tier	yes	yes
markup on tokens	none — self-hosted	none disclosed
platform fee	quote only	$49/mo (production)
self-host	yes — mit core, proprietary enterprise/	yes — mit
models carried	100+ providers	1,600+ claimed

switch forteams who want the full governance stack and the option to run it themselves

pros

+mit licence with no enterprise carve-out — genuinely self-hostable
+gateway bill is independent of token volume
+fallbacks, retries and timeouts on every tier including free
+virtual keys with budgets for per-team spend control
+guardrails documented rather than sold as an enterprise conversation

cons

−per-log billing is hard to forecast for high-request workloads
−free tier keeps logs for three days, which is too short to debug with
−publishes a broken '0+ llms' stat next to its 1,600+ claim

visit Portkey →full Portkey review →

Helicone

#3 in llm gateways · the gateway and the observability layer in one apache 2.0 codebase

84/100

verdictthe cleanest open-source licence in the category paired with a real 0% token markup, let down by overage rates you cannot read off the page.

Helicone vs LiteLLM

	LiteLLM	Helicone
price	free to self-host — enterprise is quote-only	$79/mo (pro)
free tier	yes	yes
markup on tokens	none — self-hosted	0% — byok supported
platform fee	quote only	$79/mo (pro)
self-host	yes — mit core, proprietary enterprise/	yes — apache 2.0
models carried	100+ providers	100+ providers

switch forteams who want request-level observability without running two systems

pros

+apache 2.0 throughout with no enterprise carve-out found
+0% markup on tokens, stated in the docs, byok supported
+gateway and observability in one system rather than two
+prompt management and versioning deployable through the gateway
+unlimited seats on the paid tier

cons

−overage rates live in a calculator, not on the pricing page
−data retention is one month on pro, three on team
−$79 entry price is the highest of the pass-through gateways

visit Helicone →full Helicone review →

Vercel AI Gateway

#4 in llm gateways · zero markup stated plainly, including on byok

82/100

verdictthe most explicit no-markup commitment in the category, with the compliance controls you would need at scale metered on top.

Vercel AI Gateway vs LiteLLM

	LiteLLM	Vercel AI Gateway
price	free to self-host — enterprise is quote-only	$0 — usable on the free hobby plan
free tier	yes	yes
markup on tokens	none — self-hosted	none — stated, incl. byok
platform fee	quote only	$0 hobby / $20 pro
self-host	yes — mit core, proprietary enterprise/	no
models carried	100+ providers	hundreds — not counted

switch forteams already deploying on vercel who want failover without a second vendor

pros

+explicit zero markup on tokens, including byok
+usable from a free hobby account
+automatic cross-provider retry and failover
+accepts openai and anthropic request formats unchanged
+hundreds of models plus embeddings, image and video on one balance

cons

−allowlist, zero-data-retention and reporting all meter separately
−free-tier credit amount is not published
−no self-host, and you inherit vercel as a platform dependency

visit Vercel AI Gateway →full Vercel AI Gateway review →

Cloudflare AI Gateway

#5 in llm gateways · free caching, retries and logging in front of fourteen providers

80/100

verdictthe cheapest way to put caching and observability in front of a model api, limited by a narrower provider list than the catalogue gateways.

Cloudflare AI Gateway vs LiteLLM

	LiteLLM	Cloudflare AI Gateway
price	free to self-host — enterprise is quote-only	$0 — free on all cloudflare plans
free tier	yes	yes
markup on tokens	none — self-hosted	none — 5% on unified credits
platform fee	quote only	$0 / $5 workers paid
self-host	yes — mit core, proprietary enterprise/	no
models carried	100+ providers	14 providers

switch forteams already on cloudflare who want a cache and a retry layer for nothing

pros

+free on every cloudflare plan, including the free tier
+documented pass-through with no token markup on byok
+caching, retries, fallback and rate limiting included
+$5/mo buys 10m persistent logs per gateway
+cost and token analytics without a separate observability vendor

cons

−fourteen providers, far narrower than the catalogue gateways
−5% added if you fund calls through unified billing credits
−a cloudflare account feature rather than a portable product

visit Cloudflare AI Gateway →full Cloudflare AI Gateway review →

Requesty

#7 in llm gateways · 5% on model spend, stated plainly and applied consistently

75/100

verdictthe most transparent pricing in the category and the only one that charges for inference, which makes it costlier than the pass-through gateways on identical traffic.

Requesty vs LiteLLM

	LiteLLM	Requesty
price	free to self-host — enterprise is quote-only	$0 free tier — pay-as-you-go is 5% on model cost
free tier	yes	yes
markup on tokens	none — self-hosted	5% on model cost
platform fee	quote only	none
self-host	yes — mit core, proprietary enterprise/	no
models carried	100+ providers	600+ / 300+ — conflicting

switch forteams who would rather pay one legible percentage than model a platform fee

pros

+one legible number — 5% on model spend, no seats or minimum
+eu data residency available on every plan
+fallbacks and spend limits included at the first paid tier
+routing and caching on the free tier
+openai-compatible endpoint with a documented base url

cons

−the only gateway here that marks up inference at all
−guardrails, pii detection and sso are enterprise-only at custom pricing
−publishes two different model counts on its own properties

visit Requesty →full Requesty review →

Bifrost

#8 in llm gateways · apache 2.0 on paper, open-core in the build

72/100

verdictthe fastest thing here and the most permissively licensed on paper, but the features that make a gateway safe in production are not in the apache build.

Bifrost vs LiteLLM

	LiteLLM	Bifrost
price	free to self-host — enterprise is quote-only	free (self-host) — enterprise is custom
free tier	yes	yes
markup on tokens	none — self-hosted	none — self-hosted
platform fee	quote only	none (oss)
self-host	yes — mit core, proprietary enterprise/	yes — apache 2.0
models carried	100+ providers	23+ providers

switch forlatency-sensitive proxying where the gateway must not be the bottleneck

pros

+unmodified apache 2.0 licence with no carve-out in the repository
+roughly 20µs added latency, the lowest published here
+drop-in replacement for the openai, anthropic and google sdks
+hierarchical budgets across virtual keys, teams and customers
+runs from a single go binary with zero-config startup

cons

−guardrails, clustering and the mcp gateway are enterprise-only despite the licence
−provider and model counts contradict between repo and product page
−enterprise pricing is quote-only with no published figure

visit Bifrost →full Bifrost review →

TrueFoundry

#9 in llm gateways · air-gapped deployment, at an entry price that jumps hard

68/100

verdictthe strongest isolation story in the category, priced with a step from free to $499 that has nothing in between.

TrueFoundry vs LiteLLM

	LiteLLM	TrueFoundry
price	free to self-host — enterprise is quote-only	$0 (developer, 50k requests/mo)
free tier	yes	yes
markup on tokens	none — self-hosted	none published — unverified
platform fee	quote only	$499/mo (pro)
self-host	yes — mit core, proprietary enterprise/	enterprise tier only
models carried	100+ providers	1,600+

switch forregulated teams that need the gateway inside their own vpc

pros

+vpc and air-gapped deployment so prompts stay in your infrastructure
+1,600+ models on one openai-compatible endpoint
+guardrails including pii filtering and toxicity detection
+request-metered, so token-heavy workloads do not scale the bill
+cost attribution and budgets across teams

cons

−no published statement that it takes no cut of token spend
−free tier to $499/mo with nothing in between
−self-hosting is enterprise-tier only and has no published licence

visit TrueFoundry →full TrueFoundry review →

Kong AI Gateway

#10 in llm gateways · enterprise api governance applied to model traffic, priced opaquely

64/100

verdictthe right answer if kong is already your api layer, and the hardest product here to price before you talk to sales.

Kong AI Gateway vs LiteLLM

	LiteLLM	Kong AI Gateway
price	free to self-host — enterprise is quote-only	not published
free tier	yes	yes
markup on tokens	none — self-hosted	none — pass-through
platform fee	quote only	not published
self-host	yes — mit core, proprietary enterprise/	yes — apache 2.0 core
models carried	100+ providers	not published

switch fororganisations already running kong who want ai traffic under the same policy layer

pros

+ai traffic governed by the same policy layer as the rest of your apis
+kong gateway core is apache 2.0, verified from the licence file
+no markup on tokens — proxies to your own provider credentials
+mcp server generation and agent-to-agent traffic management
+30-day trial with full enterprise functionality and no gateway limits

cons

−no published base price for the entry paid tier
−the oss-versus-enterprise split on the ai plugins is undocumented
−provider and model coverage is not listed anywhere on the product page

visit Kong AI Gateway →full Kong AI Gateway review →

how these were compared

every tool on this page went through the same test as LiteLLM — same tasks, same order, scored the same way. the comparison tables are the figures from that testing, not vendor spec sheets.

the llm gateways test in full →

was this useful?