What are the best OpenRouter alternatives in 2026?

Why OpenRouter alternatives are being re-examined this week

On 21 October, LiteLLM published an incident report saying its release pipeline was compromised and pushed a malicious v1.107.0 to PyPI and three container registries. The PyPI build, all seven wheels, was installable from 11:39:51 to 13:23 UTC. Docker Hub and GHCR tags v1.107.0, latest, main-latest and main-stable pointed at the poisoned image from 11:46 to 14:20 UTC.

The detail that matters for anyone shopping for a gateway: the official v1.107.0 source tag on GitHub does not contain the malicious files. The attacker held GitHub CLI credentials belonging to three developers, including one organization owner, disabled branch protection, and pointed the release workflow at a lookalike repository, Berri-AI/litellm. LiteLLM's own pipeline then built, signed and published the artifacts. The company says the tokens were harvested from logs of an internal LiteLLM deployment exposed to the internet, and that it observed Claude Code embedding a user's GitHub CLI OAuth token in a response body.

So the usual reassurance, "the vendor's signed build is clean," does not survive this incident on its own. Neither does "self-hosting is the safe option." A self-hosted proxy is software you install from a package index, upgraded on your schedule, and run with your provider keys and cloud credentials in reach. The security work moves to you. That is the tradeoff worth pricing before you pick a layer.

What an OpenRouter alternative actually replaces

OpenRouter is a hosted router: one API key, one OpenAI-compatible endpoint, 500+ models across 80+ providers on paid plans, and 25+ free models on the free plan. An alternative has to reproduce some subset of four jobs.

• Routing and fallbacks: send a request to a model, retry on another provider when one 429s.
• Key custody: keep provider keys in one place instead of in every service.
• Spend control: budgets, per-key limits, usage attribution.
• Observability: request logs, latency, cost per feature.

Three architectures deliver those jobs. Hosted gateways (Portkey, Helicone, Vercel AI Gateway) sit in the request path like OpenRouter. An edge gateway such as Cloudflare AI Gateway puts caching and rate limiting in front of providers while your own keys pass through. A self-hosted proxy such as LiteLLM runs inside your network, which is where the engineering time and the supply-chain risk both land.

OpenRouter alternatives compared on verified pricing

The table below uses each vendor's published price on 22 October 2026. Two numbers are easy to confuse. OpenRouter's 5.5% is charged when you buy credits, not per token. Cloudflare's 5% is also a credits-purchase fee under Unified Billing, and a gateway that only forwards your own provider keys does not touch it.

A worked example on the OpenRouter side: $1,000 of credits costs $1,055 on Standard and $1,080 on Business. Bring-your-own-key traffic is different. Standard and Business include $25,000 per month of list-price inference with no OpenRouter fee, and charge 5% above that. If your team already pays providers directly, that allowance is why the headline percentage often does not apply to you.

Vercel's gateway is the only one in this list that states zero markup and zero platform fee on tokens, including bring-your-own-key traffic, with $5 of monthly credits on the free tier. That free tier is limited to free-tier-eligible models, and rate limits are lower per model than on the paid tier.

What the platform fees cost at three spend levels

Platform fees only, with provider token costs excluded, because those are identical across all six options when the same model runs on the same provider.

• $500 a month of inference bought as credits: OpenRouter Standard $27.50, Business $40, Cloudflare with Unified Billing $25, Vercel $0, Portkey $49, Helicone Pro $79.
• $2,000 a month: OpenRouter $110 or $160, Cloudflare $100, Vercel $0, Portkey $49, Helicone $79.
• $10,000 a month: OpenRouter $550 or $800, Cloudflare $500, Vercel $0, Portkey $49 plus $9 per extra 100,000 requests over the log allowance, Helicone $79.

The crossover is early. Portkey's flat $49 is already cheaper than OpenRouter's credit fee at $900 a month of purchased credits, and Helicone's $79 beats it at about $1,450. Below that, a percentage fee is genuinely the cheapest way to get routing, and paying a fixed platform fee to save $20 is a bad trade.

Portkey and Helicone: the observability-first option

Portkey's free Developer tier is explicitly not for production: 10,000 recorded logs a month, 3-day log retention, three prompt templates. Production costs $49/mo for 100,000 recorded logs with 30-day log retention, and overage is $9 per additional 100,000 requests. Going over the log cap does not drop requests, it just stops recording them, which is the failure mode to watch if logs are your audit trail.

Helicone prices per request instead: Hobby is free at 10,000 requests a month with 1 GB of storage and one seat, Pro is $79/mo with unlimited seats, and Team is $799/mo for five organizations plus SOC-2 and HIPAA compliance. Its gateway documentation covers routing through Helicone's proxy, so the observability layer and the router are the same install. For a team that wants cost-per-customer reporting more than a model catalogue, that is a shorter path than assembling a router plus an analytics tool.

LiteLLM self-hosted: cheapest per token, highest operational cost

LiteLLM is the option with no vendor fee at all. The repository's core is MIT licensed, with the enterprise/ directory carved out under a separate license, which is why GitHub reports the repo license as NOASSERTION rather than MIT. It had 61,799 stars on 22 October 2026. You pay for a server and for provider tokens, and nothing else.

After this week, self-hosting also means owning a dependency you upgrade deliberately. LiteLLM's report gives a checklist that applies to any self-hosted gateway, not just theirs:

• Pin an exact version. LiteLLM names 1.106.2 as the safe release, and its GitHub releases show v1.106.2 published on 21 October.
• Install with hash checking, from a lockfile, so a rebuilt wheel with the same version number fails verification.
• Treat latest and main-stable tags as unreviewed code. Both were compromised in this incident.
• Search site-packages for cook.pth, sitecustomize.py and stray .pth files, since Python runs import lines in those files in every process.
• If a host ran the affected build, rotate the IAM role, every provider key in the proxy database, the DATABASE_URL password, and anything else on that host's environment. Deleting the marker file proves nothing.

The economics still favour LiteLLM at scale, and the incident does not change that. What it changes is the default posture: an internet-facing LiteLLM proxy holding keys for a dozen providers is a high-value target, and the report says the initial access came through exactly such a deployment.

Cloudflare AI Gateway: the free layer most teams underuse

Cloudflare AI Gateway is available on every plan, and the core features (dashboard analytics, caching, rate limiting) are free with a Cloudflare account and one line of code change. Log storage is where the plan matters. Accounts that created their first gateway on or after 24 September 2026 store logs in Workers Logs and follow Workers Logs pricing. Older accounts use Legacy Logs, capped at 100,000 logs total across all gateways on Workers Free and 10,000,000 logs per gateway on Workers Paid. Data Loss Prevention scanning is free on all plans, and guardrails are billed as Workers AI tokens.

The reason to pick it over a hosted router: you keep calling providers with your own keys, so no third party holds them, and caching plus rate limiting happen before the request leaves Cloudflare's network. The reason not to: it is a gateway, not a spend-tracking product. There are no per-feature budgets or virtual keys to hand out to five teams, and Unified Billing adds a 5% fee if you buy credits through it rather than paying providers directly.

How to choose between them

Start from the constraint that already exists in your stack, not from the cheapest headline.

• One provider, one app: skip the gateway. Call the provider's API directly and use its own spend dashboard. A router adds a hop, a dependency and a key to rotate.
• Several providers, you want the widest catalogue and no ops: OpenRouter. 5.5% on purchased credits is the price of not running anything.
• Already on Vercel, or you want zero token markup on your own keys: Vercel AI Gateway. Watch the free-tier model list and rate limits.
• You need caching and limits at the edge with keys you control: Cloudflare AI Gateway, and pay providers directly to avoid the 5% billing fee.
• Guardrails, prompt versioning and per-team governance: Portkey at $49/mo, or Helicone if per-request observability with SOC-2 matters more.
• Data cannot leave your network, or volume is high: LiteLLM, with pinned versions, hash-checked installs and a proxy that is not exposed to the open internet.

The honest catch across all six: every hosted router is a single point of failure that also sees your prompts, and every self-hosted proxy is a security surface you now patch. Neither problem disappears, you just pick which one you would rather staff for.

At a glance

Pricing verified on each vendor's official page, 22 October 2026

FAQ

Is OpenRouter free?

The free plan is free and includes 25+ free models with no platform fee. Free-model traffic is capped at 50 requests a day without purchased credits and 1,000 a day once the account has bought $10 of credits. Paid plans charge 5.5% (Standard) or 8% (Business) when you buy credits, with no markup on token prices.

Is LiteLLM safe to use after the 1.107.0 compromise?

LiteLLM says all versions before 1.107.0 are safe and asks users to pin 1.106.2. The malicious artifacts were built from an attacker-controlled repository, not from the official source tag. If a host ran litellm 1.107.0 or the affected Docker tags between 11:46 and 14:20 UTC on 21 October, treat it as compromised, delete cook.pth and sitecustomize.py from site-packages, kill the detached daemon, and rotate the keys that host could reach.

Which OpenRouter alternative has no platform fee?

Vercel AI Gateway states zero markup and no platform fee on tokens, including bring-your-own-key traffic, and includes $5 of monthly credits. Cloudflare AI Gateway is free for core features but adds 5% when you buy credits through Unified Billing. LiteLLM has no vendor fee at all, though you pay for the server and the engineering time.

Related reading

LiteLLM vs OpenRouter: the 5.5% fee explained, Cloudflare AI Gateway pricing, cheapest LLM API, What is OpenRouter Fusion, and is it worth the cost?, Which AI gateway should you run in 2026: managed or self-hosted?

Sources

• LiteLLM incident report, 21 October 2026
• OpenRouter pricing and fees
• Portkey pricing
• Cloudflare AI Gateway pricing docs
• Vercel AI Gateway pricing docs
• Helicone pricing
• SafeDep technical analysis of the LiteLLM compromise