Companion to the list. The list tells you which gateway; this page tells you where the space is, how to measure anything in it, and where it is heading — with the evidence and a stated way to prove each claim wrong.
Last reviewed: 2026-08-12 · Star counts and repo states verified against the GitHub API on that date.
Languages: English · 简体中文
- The one-paragraph version
- What happened to the independent layer
- What is growing at the same time
- What it means when you pick today
- The evaluation standard: six checks
- Where it heads next
- How to argue with this page
The standalone AI gateway is being squeezed from both ends. Underneath, the hyperscalers now ship routing, spend caps and observability natively, so "buy a gateway" stopped being the obvious answer for single-cloud traffic. Above, the coding-agent wave produced a second class of gateway — subscription wrappers, protocol translators, token compressors — that out-stars most of the infrastructure the phrase "AI gateway" used to mean. The part that consolidated is the independent middle: a company whose entire product was the gateway. What survives that squeeze is open-source code you can keep, or a gateway absorbed into a platform you were already paying for.
| What | When | Evidence |
|---|---|---|
| TensorZero archived — VC-backed OSS LLMOps gateway ($7.3M seed) | 2026-06-11 | Repo read-only, verified archived: true; Apache-2.0 code and community forks remain usable |
| Pydantic AI Gateway archived, folded into Logfire | 2026-03-30 | Repo archived: true; the gateway became a feature of a platform |
| Helicone acquired by Mintlify, positioned as maintenance | 2026-03 | Repo still receiving commits — usable today; the open question is roadmap, not liveness |
| Portkey acquired by Palo Alto Networks (announced 2026-04-30) | closed 2026-05-29 | The gateway became the control plane for a security platform |
| Stripe reportedly in talks for OpenRouter (~$10B) | 2026-07-23 | WSJ report — unconfirmed; a signal about where value is accruing, not a fact |
| BricksLLM and Glide — quiet, never archived | since 2025-01 / 2024-08 | No commits. The common failure mode is silence, not a shutdown notice |
The pattern is not "open source lost." Three of the six kept their code alive and lost their company. That distinction is the actionable part.
Consolidation, not collapse — the same year produced clear winners:
- LiteLLM keeps compounding as the default self-hosted choice, and is the reference implementation most clients are tested against.
- Bifrost won on a number rather than a narrative: the lowest independently measured per-request overhead.
- Envoy AI Gateway reached v1.0 on 2026-06-23 — the first CNCF-backed, production-stable open-source option, which matters to buyers who need a foundation behind the project.
- The coding-agent tier — CLIProxyAPI, OmniRoute, sub2api, 9router — now out-stars most "serious" gateway infrastructure. Whatever one thinks of the terms-of-service risk, this is where the users went.
- Token-compression proxies (headroom, Paritok) are the fastest-rising category, and they are not routers at all. They attack the bill from the token side rather than the price side.
- Bus factor is a selection criterion, not a footnote. Two projects above went read-only while still being recommended in current blog posts. Check the last commit date before the feature matrix — the list now does that check for you and publishes the result: Maintenance signal lists every tracked repo that is archived or has gone six months without a commit, refreshed daily from the GitHub API. Stars are cumulative and never fall when a project stops shipping; that table is the correction.
- Prefer the layer you can keep. Permissively licensed code with forks survives its company; a hosted control plane does not. TensorZero's code is still usable, its company is not.
- If your traffic lives in one cloud, the native gateway is now a real answer. Bedrock, Azure API Management, Vertex and Databricks Unity all shipped routing plus spend caps during 2026. Worse portability, far less to operate — a trade worth making explicitly rather than by default.
- Star count is popularity, not fitness. The most-starred project in the list is a compression proxy that cannot route between providers. Read what a thing is before reading how popular it is.
Six checks decide whether a gateway is right, ordered by how early they usually bite. They apply to anything on any list — use them on entries here, and use them on the ones we have not found yet.
| # | Check | What good looks like | How to verify |
|---|---|---|---|
| 1 | Protocol fidelity — does it break your client? | Claude Code / Codex traffic survives translation with tools and streaming intact | The compatibility surface · measured results |
| 2 | Added latency — the tax on every single request | Single-digit milliseconds, measured rather than claimed | overhead.json |
| 3 | Cache economics — is the discount still reaching you? | Cache reads actually billed at the provider's discounted rate; hit rate above the break-even point | Caching economics |
| 4 | Metering honesty — does its count match the invoice? | Spend survives a crash mid-stream; reasoning and cached tokens counted correctly | Virtual keys, budgets and metering |
| 5 | Failure behaviour — what happens when a provider dies | Retries by default, no double billing on retry, a documented rate-limit contract | Failover and reliability |
| 6 | Trust surface — who sees the prompt, who signs the release | Stated retention, signed releases, patched CVEs, a real bus factor | Retention matrix · supply chain |
Scored per gateway across compliance, markup, security, stability and observability in the scorecard. Checks 1 and 2 are reproducible with llm-gateway-bench; the fidelity probe behind check 6 is canary_check.py.
A gateway that fails check 1 is disqualified regardless of the other five — a router that silently drops your tool calls is not cheaper, it is broken. Checks 3 and 4 are where money quietly leaks, and they are the two nobody markets against.
Three calls, each with the observation that would prove it wrong. They are dated so they can be scored later rather than quietly forgotten.
| Call | Proved wrong if… |
|---|---|
| The independent hosted gateway keeps consolidating; survivors are open-source-core or absorbed into a platform | A new independent hosted gateway reaches a public SLA and named enterprise customers without being acquired |
| Coding-agent traffic becomes the default meaning of "gateway", making protocol fidelity matter more than provider count | Clients converge on one wire format, so translation stops being a source of failure |
| Token compression becomes a standard gateway feature rather than a separate proxy | Two of the top self-hosted gateways ship it natively and the standalone compressors stop growing |
Every claim above is dated and sourced, which means it is falsifiable. If a call is wrong, the useful response is evidence, not a counter-opinion: a repo state, a measured number, a dated announcement. Open an issue with it and this page changes — that is the whole point of writing the disproof conditions down first.