Radox AI for Chat, Writing and Coding in 2026: When One Endpoint Fits and When It Does Not
Radox AI is one OpenAI-compatible endpoint that reaches 33 chat models. That shape suits some jobs very well (a chat feature that may change models, a writing pipeline, a coding agent with a custom endpoint) and suits others less. This guide sorts the situations so you can decide quickly. This article focuses on chat/text. JAVIS has not run Radox AI hands-on; this guide is an editorial read of the official pages.
| Best for | Code that already speaks OpenAI |
| Chat | Streaming over SSE |
| Writing | Long context on most models |
| Coding | 7 tools with setup guides |
| Start | Free: 10M tokens/mo, no card |
Radox fits best when the thing you need is choice of model, not a new tool.
If your chat feature, writing script or coding agent already talks to an OpenAI-style API, Radox changes two settings and gives you 33 models to try.
That is the whole case for it, and it is a strong one for the right jobs. Radox's homepage puts it in one line: "Any coding agent or app that accepts a custom OpenAI endpoint works out of the box." The rest of this guide is about checking whether your job is one of those, and how to size the plan.
Six situations, and what Radox gives you in each.
| Situation | Why a gateway fits | Start with |
|---|---|---|
| A chat feature in your own app | One client library, streaming replies, model changes without a rewrite | Free to test, then a plan by volume |
| A writing or summarizing script | Models with large context windows on the same key | Free or Starter |
| A coding agent or editor with a custom endpoint | Seven tools have documented setup guides | Starter to Pro |
| Comparing several models on your own prompts | Change one string, rerun the same test | Free |
| A small team sharing one bill | One account, usage tracked per key and model | Basic or Pro |
| Agents that run for hours | Long-lived streaming connections, no monthly token cap on the top plan | Scale or Unlimited |
The same situations as filters on the catalog
Each task maps to one or two labels in the catalog, so the list of models worth testing is short.
| Task | Filter on | Models | Example IDs to try first |
|---|---|---|---|
| Chat feature in your app | Streaming (33) and Tools (32) | 33 / 32 | deepseek-v4-flash-0731, gpt-5.6-luna, claude-sonnet-5 |
| Writing or summarizing long text | Context window of 1000K or more | 27 | gpt-5.6-luna, kimi-k2.6, gemini-3.7-flash |
| Coding agent or editor | Tools label and a catalog note that mentions coding | 18 | deepseek-v4-pro-0813, claude-sonnet-5, gpt-5.6-terra |
| Screenshots, mockups, diagrams as input | Vision label | 28 | gpt-5.6-luna, seed-2.1-turbo, kimi-k2.6 |
| The model should search the web itself | Web label | 11 | kimi-k2.6, gemini-3.7-flash, glm-5.2 |
What a chat feature needs: streaming, a familiar shape, and the option to switch.
The homepage lists "Streaming over SSE with hour-long connections" and says streaming uses "server-sent events with keep-alive pings". For a chat interface, that means tokens can appear as they are generated, and a long answer does not need a special client.
Radox also states that "Request and response shapes match the OpenAI spec, including streaming chunks, tool calls, and error envelopes." If your chat code already handles those shapes, the integration work is the base URL and the key.
Streaming is one of six labels the catalog carries, and the labels are what decide whether a model suits a chat feature at all.
| Feature | What Radox says | Models with the label |
|---|---|---|
| Streaming | Server-sent events with keep-alive pings and connections that stay open up to an hour | 33 |
| Tool calling | Full round trips, parallel calls and streamed argument deltas | 32 |
| Vision input | Screenshots, mockups and diagrams as base64 or URLs | 28 |
| Reasoning | Chain-of-thought arrives in a separate field from the answer | 32 |
| JSON output | Structured JSON output | 33 |
| Web search | Built-in web search | 11 |
| Labels carried | Models |
|---|---|
| Vision and Web | 11 |
| Vision only | 17 |
| Web only | 0 |
| Neither | 5 |
This is the shape of a streamed chat call, with the model ID as the only value you change when you switch.
Writing work is mostly about context length and cost per million tokens.
Drafting, summarizing and rewriting are token-heavy because the input (a document, a thread) is often longer than the reply. Two catalog values matter: the context window and the listed rate. Most of the 33 models list a context window of 1000K or more, and the rates run from $0.115 to $16.606 per 1M tokens, per the models page.
Tokens "cover both your prompt and the model's reply, combined", so a long source document counts every time you send it. For a writing pipeline, a practical order is: test a lower-band model on a few real documents, check the output yourself, and move up a band only if the draft quality needs it.
Long inputs across the catalog
| Rate band | Models | 1000K or more | 500K | 262K or less |
|---|---|---|---|---|
| Under $1 | 6 | 3 | 0 | 3 |
| $1 to under $3 | 12 | 9 | 2 | 1 |
| $3 to under $9 | 13 | 13 | 0 | 0 |
| $9 and above | 2 | 2 | 0 | 0 |
Rates and labels by band
| Rate band | Models | Reasoning | Vision | Tools | Web |
|---|---|---|---|---|---|
| Under $1 | 6 | 5 | 2 | 5 | 0 |
| $1 to under $3 | 12 | 12 | 11 | 12 | 9 |
| $3 to under $9 | 13 | 13 | 13 | 13 | 2 |
| $9 and above | 2 | 2 | 2 | 2 | 0 |
How many drafts a plan covers
| Plan | Tokens / month | Round trips of 10,000 tokens |
|---|---|---|
| Free | 10M | 1,000 |
| Starter | 100M | 10,000 |
| Basic | 200M | 20,000 |
| Pro | 350M | 35,000 |
| Scale | 500M | 50,000 |
| Unlimited | No monthly cap | No monthly cap |
If you go past the allowance, the same arithmetic works on the balance side.
Seven coding tools have documented setup guides.
The docs section on coding agents describes "Step-by-step setup for Kilo Code, Cline, Cursor, opencode, Claude Code, Continue, and Aider." The homepage FAQ states the recipe: set the provider to "OpenAI Compatible", paste the base URL and key, and pick a model. It adds that all seven are "verified".
Choose "OpenAI Compatible" in the tool's provider settings.
Base URL https://radoxai.com/v1 and your key.
Type a model ID from the catalog, for example claude-sonnet-5.
Two features matter for coding tools: tool calling (the mechanism "behind AI coding agents", per the docs list) and reasoning models, where "thinking never contaminates your answer" because chain-of-thought arrives in a separate field. The docs also have a page on how to budget tokens for reasoning models, which is worth reading before a long agent session.
Kilo Code
Cline
Cursor
opencode
Claude CodeAll seven tools, and one more on the homepage
| Tool | Category on the Radox homepage | In the docs setup guide | In the FAQ's verified list |
|---|---|---|---|
| Kilo Code | VS Code agent | Yes | Yes |
| Cline | VS Code agent | Yes | Yes |
| Cursor | AI code editor | Yes | Yes |
| opencode | Terminal agent | Yes | Yes |
| Claude Code | Terminal agent | Yes | Yes |
| Continue | Not described on the homepage | Yes | Yes |
| Aider | Not described on the homepage | Yes | Yes |
| OpenClaw | Personal agent | Not named in the docs list | Not named in the FAQ list |
What the catalog says about coding
Radox writes a one-line note for each model. These are the entries that mention coding.
| Model ID | Maker | In / 1M | Context | Radox's catalog note |
|---|---|---|---|---|
deepseek-v4-pro-0813 | DeepSeek | $0.552 | 1049K | “Excellent reasoning and coding, 1.6T params” |
seed-2.1-turbo | ByteDance Seed | $0.92 | 262K | “solid reasoning and coding” |
seed-2.1-pro | ByteDance Seed | $1.15 | 262K | “Strong reasoning and coding” |
kimi-k2.6 | Moonshot AI | $1.219 | 1000K | “Strong reasoning and coding, 1.0T params” |
glm-5.2 | Zhipu AI | $1.357 | 1000K | “Good reasoning and coding, 753B params” |
qwen3.7-max | Alibaba Qwen | $1.76 | 1000K | “Strong reasoning and coding performance” |
grok-4.6 | xAI | $2.806 | 500K | “Improved reasoning and coding over 4.5” |
claude-sonnet-5 | Anthropic | $3.324 | 1000K | “Great coding and reasoning at a lower cost” |
gpt-5.6-terra | OpenAI | $3.577 | 1050K | “Strong reasoning and coding at a mid price point” |
gemini-3.1-pro | $4.474 | 1049K | “Strong reasoning and coding with high Code Arena score” | |
kimi-k3 | Moonshot AI | $4.98 | 1000K | “Strong reasoning and coding, 2.8T params” |
claude-opus-4.7 | Anthropic | $8.303 | 1000K | “Strong coding and reasoning, high Code Arena score” |
claude-opus-4.8 | Anthropic | $8.303 | 1000K | “Reliable reasoning and coding performance” |
gpt-5.5 | OpenAI | $8.947 | 1050K | “Solid reasoning and coding performance” |
gpt-5.5-pro | OpenAI | $8.947 | 1050K | “Enhanced reasoning and coding over base 5.5” |
gpt-5.6-sol | OpenAI | $8.947 | 1050K | “Top-tier reasoning, coding, and agentic capability” |
claude-fable-5 | Anthropic | $16.606 | 1000K | “Outstanding coding and reasoning with top Code Arena score” |
claude-fable-5.1 | Anthropic | $16.606 | 1000K | “Outstanding coding and reasoning with top Code Arena score” |
Who the 18 coding-note models come from
| Maker | Under $1 | $1 to under $3 | $3 to under $9 | $9 and above | Total |
|---|---|---|---|---|---|
| Anthropic | 0 | 0 | 3 | 2 | 5 |
| OpenAI | 0 | 0 | 4 | 0 | 4 |
| 0 | 0 | 1 | 0 | 1 | |
| DeepSeek | 1 | 0 | 0 | 0 | 1 |
| Alibaba Qwen | 0 | 1 | 0 | 0 | 1 |
| ByteDance Seed | 1 | 1 | 0 | 0 | 2 |
| Moonshot AI | 0 | 1 | 1 | 0 | 2 |
| Zhipu AI | 0 | 1 | 0 | 0 | 1 |
| xAI | 0 | 1 | 0 | 0 | 1 |
Coding-note models against the rest
Match the plan to how many tokens you send, not to the tool you use.
Every plan reaches all 33 models; only monthly tokens and per-minute limits differ. The table pairs each workload with the plan whose allowance and limits fit it. Prices are on the pricing page and are not repeated here.
| Workload | Allowance to look at | Per-minute limit | Plan |
|---|---|---|---|
| Trying models, a script now and then | 10M tokens / month | 60 requests | Free |
| Side project, one editor with an agent | 100M to 200M tokens / month | 120 to 300 requests | Starter or Basic |
| A production feature or a small team | 350M tokens / month | 600 requests | Pro |
| Several services or long agent runs | Scale: 500M tokens; Unlimited: no monthly cap | 1,200 to 3,000 requests | Scale or Unlimited |
If you run past the allowance, requests bill from your balance at the model's rate, and with an empty balance you get HTTP 402, per the pricing FAQ. The allowance "resets monthly" and does not carry over.
Allowances and limits side by side
| Plan | Tokens / month | Requests / min | Tokens / min |
|---|---|---|---|
| Free | 10M | 60 | 100,000 |
| Starter | 100M | 120 | 200,000 |
| Basic | 200M | 300 | 500,000 |
| Pro | 350M | 600 | 1,000,000 |
| Scale | 500M | 1,200 | 2,000,000 |
| Unlimited | No monthly cap | 3,000 | 5,000,000 |
Order as described in the pricing FAQ and homepage FAQ, Oct 15, 2026.
What a month of allowance buys, task by task
| Task (assumed size) | Free | Starter | Basic | Pro | Scale |
|---|---|---|---|---|---|
| Chat reply, about 2,000 tokens in and out | 5,000 | 50,000 | 100,000 | 175,000 | 250,000 |
| 1,000-word draft: about 3,000 in plus 1,500 out | 2,222 | 22,222 | 44,444 | 77,777 | 111,111 |
| Summary of a 50,000-token document plus a 1,000-token reply | 196 | 1,960 | 3,921 | 6,862 | 9,803 |
| Coding-agent session of about 200,000 tokens | 50 | 500 | 1,000 | 1,750 | 2,500 |
How long the allowance lasts at full speed
Cases where a direct contract with the model maker suits better.
Radox's Terms are plain that it is "an API gateway" that routes requests "to third-party AI model providers". An intermediary adds one more party between you and the model. For some work that is fine; for other work you may want the maker's own terms.
- You need a contract that names the model maker directly.
- Your data rules require choosing the exact provider and region.
- You need a model the catalog does not list.
- You want several makers on one key and one bill.
- You switch models often while testing.
- Your tools accept a custom OpenAI endpoint.
Good habits for any intermediary AI service.
These apply to every gateway, and Radox's Terms say much the same: "Keep keys server-side." and "Scope each key to only what it needs, and rotate keys you suspect are exposed."
- Keep keys out of browser code, public repositories and shared notebooks.
- Do not paste passwords, keys or sensitive personal data into prompts.
- Check AI output before you rely on it; the Terms put that responsibility on you.
- The Privacy Policy notes that prompts go to "the upstream provider serving the model you selected", where that provider's own retention policy applies.
Pick a path by what you are building.
Change the base URL and key, start on Free, and test two models.
Check that it is one of the seven documented tools, or that it accepts a custom OpenAI endpoint.
Size by monthly tokens and requests per minute, then read the usage dashboard after the first month.
Use the maker's own API.
The decision as a flow
nocontinue
yescontinue
Match the need before comparing prices.
| If you mostly need... | Look at |
|---|---|
| A direct contract and terms with the model maker | The maker's own API, such as Anthropic, OpenAI or Google |
| Another hosted multi-model gateway | OpenRouter |
| A gateway you host and control yourself | LiteLLM proxy (open source) |
| A first look at Claude before choosing a model | Our Claude review |
How this guide was made.
Links labelled as Radox AI calls to action on this page use JAVIS's referral link; if you sign up through them, JAVIS may receive a referral reward at no extra cost to you.
This guide is built from Radox AI's official pages (home, pricing, models, docs, Terms, Privacy), dated Oct 15, 2026. This article focuses on chat/text.
Start on the free plan; no card required per the pricing page.
Create a free accountRead the related guides below first.
Choose the next articleIf you are deciding how to use a gateway for daily work.
Questions people ask before choosing a gateway for daily work
Is Radox AI good for chat features?
It fits when your code already uses an OpenAI-style client: streaming over server-sent events is documented, and models change by editing one string.
Which coding tools does Radox document?
Kilo Code, Cline, Cursor, opencode, Claude Code, Continue and Aider.
Can I use it for long documents?
Most catalog models list a 1000K or larger context window; check the model page for the exact value.
What does it cost to start?
The Free plan gives 10M tokens a month with no card required, per the pricing page.
What happens if I use up my allowance?
Requests bill from your balance at the model's rate; with an empty balance you get HTTP 402.
Is my prompt stored by Radox?
Its Privacy Policy says it does not store prompt or reply content, and that the upstream provider's own retention policy applies.
Should I use Radox or the maker's API?
Use the maker's API if you need a direct contract or provider control; use a gateway if model choice on one key matters more.
Where this came from, and the date it applies to.
Endpoint, streaming, coding-tool recipe, compatibility statement: radoxai.com (home) and radoxai.com/docs, Oct 15, 2026.
Plans, limits, billing order, 402: radoxai.com/pricing, Oct 15, 2026.
Context windows and rates: radoxai.com/models, Oct 15, 2026.
Gateway description, key handling, responsibility for output: Radox Terms of Service, effective 23 August 2026.
Data handling: Radox Privacy Policy, effective 23 August 2026.
Images: logos hosted on radoxai.com (branding, coding tools).
Added charts and tables: counts, bands and cross-counts are calculated by JAVIS from the models and pricing pages dated Oct 15, 2026.
