Published Updated
In one or two sentences, before anything else.
Kilo Gateway is the OpenAI-compatible model API behind the Kilo Code agent, run by Kilo Code, Inc., which Anaconda acquired in July 2026. It bills tokens at the upstream rate plus a 5% fee on credit purchases, and most of its 400 models match OpenRouter's IDs and prices.
What it can do, one sentence each, and whether we actually tested it.
One OpenAI-compatible endpoint for models from Anthropic, OpenAI, Google, xAI, DeepSeek, Moonshot, Z.ai, MiniMax and Qwen; its public model list held 400 entries on October 21, 2026.
Kilo's pages say 500+ models; the public list is smaller.
Tokens are deducted at the upstream provider's rate, and a 5% payment-processing fee is charged on each credit purchase, auto top-up and Kilo Pass payment.
Per Kilo's billing docs. The gateway page's "market rates" wording does not mention the fee.
24 bring-your-own-key options, including Anthropic, OpenAI, Bedrock, Z.ai Coding Plan, Kimi Code and OpenCode Go; usage is tracked at 0 USD on Kilo's side.
Per Kilo's authentication docs. A failed BYOK request does not fall back to Kilo's own keys.
Models tagged :free and the kilo-auto/free router work without signing in, at 200 requests an hour per IP address; 13 :free model IDs were listed on October 21, 2026.
Kilo's docs warn that Auto Free may route to providers that log prompts and use them for training.
Virtual models that pick a model per mode or task: Frontier sends plan and ask modes to Claude Opus 4.7 and code to Claude Sonnet 4.6; Efficient falls back to GLM-5.3 Flash.
Per Kilo's model docs; the mappings change server-side.
Kilo Pass Starter costs 19 USD a month plus 5% and shows up to 26.60 USD of monthly credit with bonuses; bought credits never expire.
Per Kilo's Kilo Pass page, bonus credits expire at the end of each monthly cycle.
Automatic repair of duplicated or orphaned tool calls, and fill-in-the-middle code completions through Mistral Codestral.
Per Kilo's gateway docs.
A remote MCP server at mcp.kiloapps.io/mcp with two tools, search and call, over Kilo's platform API: usage and spend, agent profiles and cloud agents.
Per Kilo's blog (October 13, 2026). It manages your Kilo account; it does not run models.
The people it helps and the problems it solves, and where it ranks.
Built for
Problems it solves
Graded in
Public tests other people ran: what they gave it and how it did.
Case 1
Input
10 agentic sessions across 8 models in VS Code, WebStorm and the CLI, then Anthropic's billing dashboard checked against Kilo's usage log.
What happened
The bills matched to the cent. One request went through OpenRouter despite a direct Anthropic key, Orchestrator mode used three to four times the tokens of a single chat, and about 11 USD of Kilo Pass bonus credit expired unused. The site carries affiliate links; one reviewer.
Case 2
Input
createOpenAI with baseURL https://ai-gateway.kilo.ai/api/v1 and a KILO_API_KEY, then streamText with the model anthropic/claude-sonnet-4.5.
What happened
Any OpenAI-compatible client works by changing the base URL. Only chat completions is documented, so tools that speak Anthropic Messages or OpenAI Responses, such as Claude Code or Codex, cannot point at it directly.
Case 3
Input
A simple question in a Kilo Code session that kept re-reading files.
What happened
The loop used 8.5 million tokens, about 7.59 USD. Kilo support refunded it and pointed to a max-cost setting, which the user asked to be on by default.
What it does badly, with the evidence, and what we could not settle.
Weak
It is largely a layer on OpenRouter.
Evidence: In GitHub issue #2279 a Kilo collaborator wrote in November 2025: "Currently all Kilo Gateway traffic goes through OpenRouter." On October 21, 2026, 388 of its 400 model IDs also appeared in OpenRouter's list, 323 of them at identical input and output prices.
Its price wording differs from page to page.
Evidence: The gateway page says "market rates" and, for your own keys, "Pay only for the routing, not the tokens"; the billing docs add a 5% fee to every purchase and charge nothing on your own keys; a comparison page says zero markup without the fee.
Open-weight models are listed at the maker's rate, not the cheapest host's.
Evidence: Kilo lists DeepSeek V4.1 Flash at 0.30 / 1.20 USD per million tokens and GLM-5.3 at 1.40 / 4.40 USD, while OpenRouter's cheapest hosts listed 0.024 / 0.375 USD and 0.079 / 1.40 USD on October 21, 2026.
Only an OpenAI Chat Completions API.
Evidence: Kilo's API reference documents chat completions, FIM completions and the model list, with no Anthropic Messages or OpenAI Responses endpoint.
Customer data was exposed through a vendor in August 2026.
Evidence: Kilo says an attacker reached customer records at its analytics vendor Metabase: some names and emails and a few Slack tokens, with no payment data.
Refunds are not guaranteed.
Evidence: No refund policy appears in Kilo's docs. Trustpilot shows 2.5 out of 5 from 18 reviews, with refused refunds for a 19 USD Kilo Pass renewal and a 100 USD credit purchase; one GitHub user did get a refund.
Not for teams that need Anthropic- or Responses-format endpoints, a written refund policy, or a gateway that is not itself built on another reseller.
Evidence: The API reference, the missing refund policy and the OpenRouter overlap above.
Uncertain
What an open-weight request actually costs.
Why unsure: The billing docs say cost follows the upstream provider's pricing, but the list shows the maker's rate and the docs give no example.
Why some Google models are listed below Google's price.
Why unsure: Kilo lists Gemini 3.1 Pro Preview at 1 / 6 USD per million tokens against Google's 2 / 12 USD, and gives no reason.
How much traffic still goes through OpenRouter.
Why unsure: Kilo's gateway page now names direct providers and gateways like OpenRouter and Vercel as upstreams but gives no split.
Prices after the Anaconda deal.
Why unsure: Anaconda said on July 15, 2026 that models and pricing stay the same, with no time commitment.
The numbers, each with its source and the date we checked it.
| Spec | Value | Note and source |
|---|---|---|
| Operator and owner | Kilo Code, Inc., maker of the open-source Kilo Code agent; acquired by Anaconda (announced July 15, 2026, terms undisclosed) | Source: anaconda.comchecked |
| Price against the makers | Upstream provider's rate with no token markup, e.g. Claude Opus 5.5 at 4 / 20 USD and GPT-5.4 at 2.50 / 15 USD per million input / output tokens | Anthropic and OpenAI list the same prices. Source: ai-gateway.kilo.aichecked |
| Platform fee | 5% payment-processing fee on credit purchases, auto top-ups and Kilo Pass; 0 USD on your own keys | Source: kilo.aichecked |
| Minimum top up | Not stated for individuals; organisation auto top-up from 100 USD; mobile packs of 10, 50, 100 or 500 USD | Source: kilo.aichecked |
| Payment methods | Card through Stripe on the web; App Store and Google Play in the mobile apps; invoices for Enterprise | Source: kilo.aichecked |
| Free tier | Models tagged :free and kilo-auto/free, no account needed, 200 requests an hour per IP | Source: kilo.aichecked |
| Models | 400 in the public list (Kilo says 500+); 388 of them also in OpenRouter's list | Source: ai-gateway.kilo.aichecked |
| API formats | OpenAI Chat Completions with streaming, tool calling and response_format; FIM completions; no Anthropic Messages or Responses endpoint documented | Source: kilo.aichecked |
| MCP | Kilo platform MCP server at mcp.kiloapps.io/mcp (usage, spend, agent profiles, cloud agents); no MCP features in the gateway API | Source: blog.kilo.aichecked |
| Agent Skills | Kilo-Org/skills: 3 official skills (data investigation, Snowflake query debugging, Jupyter notebooks), none for the gateway; no Anthropic Skills API | Source: github.comchecked |
| Coding agents | Built into Kilo Code (VS Code, JetBrains, CLI, cloud agents, mobile); Kilo maintains a Pi provider; OpenClaw documents a Kilo provider | Source: kilo.aichecked |
| Privacy | Organisation data-collection policy; free routes may log and train on prompts; your keys encrypted at rest; no SOC 2 or zero-retention statement found for the gateway | Source: kilo.aichecked |
| Status and incidents | status.kilo.ai: gateway degraded 07:08 to about 07:24 UTC on September 17, 2026; Anthropic models unavailable through the gateway for about two hours on September 22, 2026 | Source: status.kilo.aichecked |
| Refunds and expiry | No refund policy in the docs; bought credits never expire; Kilo Pass bonus credits expire monthly | Source: kilo.aichecked |
Answer first, then why, an example, the limits and what to do next.
Kilo Gateway has relied heavily on OpenRouter: a Kilo collaborator wrote in November 2025 that all gateway traffic went through it. Kilo's gateway page now names direct providers and gateways like OpenRouter and Vercel as upstreams, and on October 21, 2026, 388 of its 400 model IDs also appeared in OpenRouter's list.
Kilo Gateway is not cheaper than the official APIs for most models: tokens cost the upstream rate and every credit purchase adds 5%. Claude Opus 5.5 is 4 / 20 USD per million tokens on both. Using your own keys or a coding plan through Kilo costs nothing extra on Kilo's side.
Limits
A few Google models are listed below Google's price, such as Gemini 3.1 Pro Preview at 1 / 6 USD against 2 / 12 USD, without explanation.
Kilo Gateway's free models carry a :free tag and work without an account, at 200 requests an hour per IP address. On October 21, 2026 the list had 13, including MiniMax M3.1 Flash, NVIDIA Nemotron models and Poolside Laguna, plus the kilo-auto/free router.
Limits
Kilo warns that free routes may send prompts to providers that log them and use them for training.
Kilo Gateway is run by a funded company now owned by Anaconda, but free routes may log prompts and a vendor breach exposed some customer data in August 2026. Kilo says no payment data was involved. No SOC 2 or zero-retention statement covers the gateway, and refunds are not guaranteed.
OpenRouter is the better choice outside Kilo's own tools, while Kilo Gateway makes most sense inside Kilo Code. OpenRouter offers more models, Anthropic and Responses endpoints and cheaper open-model hosts; Kilo charges 5% against OpenRouter's 5.5% and routes coding plans such as Z.ai and Kimi Code at no charge.
Kilo Gateway itself has no MCP features, but Kilo publishes a platform MCP server. The server at mcp.kiloapps.io/mcp lets an MCP client read usage and spend and manage agent profiles and cloud agents, and Kilo Code works as an MCP client for other servers.
Anaconda announced on July 15, 2026 that it had acquired Kilo Code, the company behind Kilo Gateway. Terms were not disclosed, and Anaconda said the models, pricing and open-source repository stay the same.
This page is built from Kilo's own gateway page, pricing pages, docs, status page and public model list, the makers' price pages, OpenRouter's public model list, a hands-on review by OpenAI Tools Hub, GitHub issues, Trustpilot and Hacker News, all checked on October 21, 2026. Kilo's claims are labelled as its own.
Kilo Gateway is the model API inside Kilo Code, the open-source coding agent launched in March 2025, whose CEO and co-founder is Scott Breitenother. Kilo raised an 8 million USD seed round in December 2025 (Pulse 2.0), and Anaconda acquired Kilo in July 2026, saying models and pricing stay the same.
One OpenAI-compatible endpoint, https://ai-gateway.kilo.ai/api/v1, serves Kilo Code in VS Code, JetBrains, the CLI, cloud agents and the mobile apps, and any OpenAI-compatible client (docs). Its public model list held 400 entries on October 21, 2026, almost all text models; Kilo's pages say 500+. You can pay with credits, a Kilo Pass subscription, or your own keys and coding plans from 24 providers, which Kilo routes at no charge (authentication docs).
In September 2025 a user asked Kilo to be upfront that its API routes through OpenRouter; in November 2025 a Kilo collaborator replied in issue #2279: "Currently all Kilo Gateway traffic goes through OpenRouter." Kilo's gateway page now says it load-balances between direct providers and gateways like OpenRouter and Vercel. The public lists still overlap heavily: 388 of Kilo's 400 model IDs also appeared in OpenRouter's list on October 21, 2026, and 323 of them carried identical input and output prices. In issue #13894 (September 2026) requests for a model labelled OpenRouter in the picker went to Kilo Gateway and asked for Kilo credits despite a connected OpenRouter key.
Kilo's billing docs say tokens are deducted at the upstream provider's rate with no markup, and a 5% payment-processing fee is charged on each purchase, so 20 USD of credit costs 21 USD. Requests on your own keys cost 0 USD on Kilo's side.
| Model, per million input / output tokens | Kilo Gateway | Maker's list price |
|---|---|---|
| Claude Opus 5.5 | 4 / 20 USD | 4 / 20 USD (Anthropic) |
| GPT-5.4 | 2.50 / 15 USD | 2.50 / 15 USD (OpenAI) |
| Gemini 3.1 Pro Preview | 1 / 6 USD | 2 / 12 USD (Google) |
| DeepSeek V4.1 Flash | 0.30 / 1.20 USD | 0.30 / 1.20 USD at peak hours, half off-peak (DeepSeek) |
Add 5% to every row. Kilo does not explain why Gemini 3.1 Pro Preview is listed at half Google's price. For open-weight models Kilo shows the maker's rate, while OpenRouter's endpoint list had DeepSeek V4.1 Flash hosts from 0.024 / 0.375 USD; because billing follows the upstream provider, the listed figure may differ from the charge. Kilo Pass Starter costs 19 USD a month plus 5% and shows up to 26.60 USD of monthly credit with bonuses, but bonus credits expire at the end of each month.
No independent latency, throughput or uptime test of Kilo Gateway has been published. Kilo's status history shows the gateway degraded from 07:08 to about 07:24 UTC on September 17, 2026, and Anthropic models unavailable through the gateway for about two hours on September 22, 2026.
Trustpilot shows 2.5 out of 5 from 18 reviews: praise for model choice and pairing with Z.ai plans, complaints about reliability, token burn, refused refunds for a 19 USD Kilo Pass renewal and a 100 USD purchase, and configured providers not being respected. In issue #3767 an agent loop burned 8.5 million tokens on a simple question; Kilo refunded the 7.59 USD. On Hacker News commenters on the Anaconda deal expected pressure to make Kilo more profitable.
https://mcp.kiloapps.io/mcp on October 13, 2026. It has two tools, search and call, over Kilo's platform API (usage and spend, agent profiles, cloud agents), signs in by OAuth and is documented for the Kilo CLI and VS Code extension. It does not run models, and the gateway API has no MCP features. Kilo Code itself is an MCP client.npx skills add Kilo-Org/skills; none covers the gateway. The Kilo Marketplace collects skills, MCP servers and modes. With no Anthropic endpoint, Anthropic's Skills API does not apply.response_format, plus FIM completions (API reference). Kilo has no SDK of its own; its docs use the Vercel AI SDK's OpenAI provider and the OpenAI SDK, and publish an llms.txt.Organisations can allow or deny providers that collect data, but some free models require allowing it, and Kilo warns that Auto Free may send prompts to providers that log them and train on them, including NVIDIA trial endpoints (model docs). Your own keys are encrypted at rest. No SOC 2 or zero-retention statement for the gateway was found. In August 2026 an attacker reached Kilo customer records at its vendor Metabase: some names and emails and a few Slack tokens, with no payment data, according to Kilo. Kilo's docs give no refund policy; bought credits never expire.
Skip Kilo Gateway if you do not use Kilo Code, need an Anthropic- or Responses-format endpoint, want a written refund policy, or would rather pay the upstream relay directly. For Kilo Code users who bring a coding plan or their own keys, it adds no cost.
Sources
Changelog
APIMODELS is a pay-as-you-go AI API relay that sells about 150 models through one key at its own prices, mostly below the makers' list prices, and spends the margin on backup channels: 98.62% of requests succeeded over 30 days, moderation rejections excluded. APIMODELS is our own product: it is made by the team behind GenGrades.
ZenMux is an LLM gateway, operated under its terms by NexaMind Singapore Pte. Ltd., that sells Claude, GPT, Gemini and many Chinese models through OpenAI, Anthropic and Vertex-style endpoints at the makers' list token prices. It adds automatic credits for bad or slow answers, which it calls insurance, and 20 to 200 USD monthly coding plans.
Vercel AI Gateway is Vercel's hosted API for about 480 text, image, video and speech models behind one key, billed from prepaid credits at the providers' list prices with no markup, including when you bring your own keys. It became generally available on August 21, 2025, and gives each Vercel team 5 USD of free credit a month.
Requesty is a London-based LLM gateway, run by Requesty Ltd and seed-funded by 20VC in 2025, that sells OpenAI- and Anthropic-compatible access to more than 600 models at list price plus a 5% top-up fee. It stands out for routing policies, spend controls and an EU endpoint in Frankfurt.
Poe API is Quora's developer API for the bots in its Poe chat app, opened on July 31, 2025, with OpenAI-compatible endpoints and an Anthropic-compatible one for official Claude bots. Calls spend the same points as a Poe subscription, so there is no separate API bill; current Claude bots cost about 1% more than Anthropic's list price.
OpenRouter is a New York company's API relay that sells access to more than 680 models from over 100 hosting providers through one key, at each provider's list price plus a 5.5% fee on credit purchases. Stripe agreed to buy it on August 19, 2026; no closing had been announced by October 21, 2026.