Qoder vs Claude Code: which Qoder, which endpoint, which bill

These are not the same kind of purchase. Claude Code is a Claude-only agent that can be pointed at any Anthropic-format endpoint with two environment variables and no subscription. Qoder is a credit-metered platform whose published model lists — for the standalone app, the IDE and the CLI — name no Claude model, so “run Claude in Qoder” means bring-your-own-key, and only on the surfaces that accept one. The single most expensive mistake on this comparison is treating “Qoder” as one product. It is not, and the surfaces disagree with each other about whether you can use your own endpoint.

First: which Qoder are you comparing?

As of September 21, 2026 the brand covers several separately-versioned surfaces, and the bring-your-own-key rules are different on each. The standalone app's own overview says it “originated from Quest mode in Qoder IDE and is now an independent desktop product”, that the two are available side by side, and that they “do not replace each other”.

Sources, all checked September 21, 2026: standalone app release notes (0.1.0 on August 27, 2026; 0.3.3 on September 18, 2026), IDE release notes (1.0.0 on May 15, 2026: “Qoder 1.0, evolving from AI IDE to autonomous development desktop”), CLI release notes, and the npm registry. Qoder is widely reported as Alibaba's, and Qoder for Enterprise is listed on the Alibaba Cloud Marketplace, but neither docs.qoder.com nor qoder.com names the legal entity, so ownership stays a press-sourced claim here rather than a first-party one. Note also that Claude Code's documentation moved: docs.claude.com/en/docs/claude-code/* now redirects to code.claude.com/docs/en/*.

Who should pick which

Pick Qoder if you want an opinionated platform that bundles the models with the tool, and if the models it bundles suit your work. Its selectors list Qwen, DeepSeek, GLM, Kimi and MiniMax families behind Auto, Ultimate, Performance and Efficient tiers, and you never think about an API key or a token rate — you watch a credit meter instead. The repo-aware extras it sells around the agent, such as Repo Wiki and Quest, are the reason to buy it: a custom base URL buys you a model, not those features.

Pick Claude Code if you want Claude models specifically, or if you want the same agent across several places. Anthropic's gateway connection page names the CLI, the VS Code extension, the desktop app, GitHub Actions, the Agent SDK and the cloud surfaces such as Slack and the web as the places it runs. And it separates cleanly from Anthropic's billing: Anthropic's gateway page states that while a gateway credential is active the claude.ai subscription “isn't used” and its usage limits do not apply.

On execution and permissions, this page did not run either tool, so it makes no claim about how either behaves in practice. The documented difference is where a custom model reaches. Qoder's model selector page says that once added under Settings → Models, “the model becomes available to task, Agent, and automation model selectors”, while its custom-models page carves out Repo Wiki, which uses a fixed model and consumes credits separately — so BYOK does not cover every feature. On Claude Code, a gateway credential covers inference everywhere except the claude.ai-identity features listed further down.

On plan and protocol limits, the gates are opposite in kind. Qoder's gate is commercial: its IDE and CLI custom-model pages restrict BYOK to an Individual plan, while the surface that takes an arbitrary URL has a docs page stating no edition requirement and a release note tying the feature to the Personal plan. Claude Code's gate is technical: three API formats, no plan requirement, and no OpenAI-compatible path at any price.

On cost path, Qoder is a prepaid allowance with an expiry and Claude Code is either a usage window or straight per-token billing. On migration cost, both directions are cheap to configure and expensive to unlearn: the endpoint is minutes, the agent's habits and your instruction files are not.

Side by side

The BYOK boundary, written twice because it is two boundaries

Claude Code accepts exactly three API formats, per Anthropic's gateway compatibility guide: “A gateway must expose at least one of the following API formats” — Anthropic Messages at /v1/messages with an optional /v1/messages/count_tokens, Amazon Bedrock InvokeModel, or Google Cloud's Agent Platform rawPredict. There is no OpenAI-compatible door, so a gateway's /v1/chat/completions route is irrelevant here. Anthropic also states that it “doesn't support routing Claude Code to non-Claude models through any gateway”.

Qoder depends entirely on which Qoder. The standalone app's custom-models page opens “Qoder supports accessing third-party provider model resources via API keys”, offers “OpenAI Compatible” and “Anthropic Compatible” provider types, and takes a Base URL, an API key and one or more model IDs, with optional display name, context window, and vision and thinking-mode capability flags. For the OpenAI type it also asks you to pick the Chat Completions API or the Responses API. The release note that added it is version 0.1.8, September 5, 2026: “Personal plan BYOK now supports custom Base URLs for any OpenAI- or Anthropic-compatible model service.” Qoder IDE is a different story — its custom-models page reads “Applicable For: Users with Individuals plans” and lists named providers only — Alibaba Cloud Model Studio, DeepSeek, Z.ai, Kimi, MiniMax and Xiaomi MIMO. The IDE's 1.27.0 note of August 29, 2026 says “Personal edition BYOK now supports OpenAI, Google, and OpenRouter providers”, which is a longer list of named providers rather than an endpoint field, and the strings Base URL, Anthropic and Claude do not appear on the IDE custom-models page at all.

# Claude Code — the documented shape, and what Kunavo's own setup page uses.
# ANTHROPIC_BASE_URL is an ORIGIN. Claude Code appends /v1/messages itself,
# so a value ending in /v1 produces /v1/v1/messages and fails.
export ANTHROPIC_BASE_URL=https://api.kunavo.com
export ANTHROPIC_AUTH_TOKEN=sk-kn-...     # Authorization: Bearer
# export ANTHROPIC_API_KEY=sk-kn-...      # x-api-key — pick ONE; the wrong one 401s

# Qoder, standalone desktop app — Settings -> Models -> add a custom model.
# Not env vars: three form fields, and the documented example Base URL
# CARRIES /v1, which is the opposite convention to the one above.
#   Provider type : "Anthropic Compatible"  (or "OpenAI Compatible")
#   Base URL      : https://api.example.com/v1   <- Qoder's own example
#   API Key       : your key
#   Model ID      : typed by hand, one or more. The page documents no
#                   model-list discovery for this field.

The URL suffix is a real trap, and it goes both ways. Anthropic's own verification snippet is curl -X POST "$ANTHROPIC_BASE_URL/v1/messages", which is only correct if the variable holds an origin; Qoder's documented example Base URL is https://api.example.com/v1. Copying one convention into the other's field produces a 404 that looks like an outage. The credential variable matters as much: Anthropic states that ANTHROPIC_AUTH_TOKEN goes in Authorization: Bearer, ANTHROPIC_API_KEY in x-api-key, apiKeyHelper in both, and “a credential in the wrong variable reaches the gateway in a header it doesn't read, and the request fails with 401”. ANTHROPIC_BASE_URL covers the origin rule in full.

What a custom endpoint costs you on Claude Code. This is the honest other side. Remote Control and voice dictation both rely on a claude.ai identity and are unavailable while a gateway credential is active; Remote Control is separately disabled whenever ANTHROPIC_BASE_URL points at a non-Anthropic host. The desktop app reads gateway routing from its own third-party inference configuration rather than from ANTHROPIC_BASE_URL, and with that configuration active it runs sessions locally only, with no SSH or Anthropic-hosted cloud environments. Anthropic's MCP page lists a custom ANTHROPIC_BASE_URL among the configurations without MCP tool search, fine-grained tool streaming is likewise off by default behind a custom base URL, and the fast-mode availability check calls api.anthropic.com directly regardless of the base URL. One thing you do keep: a gateway that serves /v1/models can populate the /model picker when a developer sets CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY=1 — a GET /v1/models?limit=1000 with a three-second default timeout, where any redirect counts as failure and Claude Code keeps only IDs containing claude or anthropic. Kunavo serves that endpoint, which makes this supported by Claude Code rather than something tested here.

Where Qoder's own sources disagree

Two disagreements matter to a buyer, and neither is resolved from outside the product.

Qoder CLI and custom URLs. The CLI custom-models page describes a catalogue, not an endpoint: “Available edition: Individual plan”, “Do not configure BYOK manually in settings.json; the available providers, models, and credential fields are determined by the BYOK catalog available to the current account”, with the wizard's options named as the source of truth. The strings Base URL, Custom URL, OpenAI, Anthropic, endpoint and compatible do not appear on it. But the CLI release notes say, at 1.1.50 on September 11, 2026, that “Personal edition users can now centrally configure and manage BYOK models with custom URL endpoints from the Custom page in /model”. Both pages were live on September 21, 2026, and a third, the CLI model page, describes that same Custom tab as “Connect models from your own provider subscriptions. Qoder supports a curated list of providers” — which reads with the catalogue, not with the release note. This guide settles none of it: check the /model Custom tab in your own installed CLI before planning around it. Either way the CLI is not a standalone client — its authentication page states “Authentication is required before using Qoder”, through browser sign-in or a QODER_PERSONAL_ACCESS_TOKEN — so even in the best case that path is your own key on top of a Qoder account, not instead of one. Claude Code has no equivalent requirement on the gateway path.

Which plan unlocks BYOK. Three vocabularies, no published mapping. The IDE page says “Individuals plans”, the CLI page says “Individual plan”, the release notes say “Personal edition” and “Personal plan”, and the standalone app's custom-models page states no edition requirement at all. The pricing page adds a fourth signal that fits none of them: “Bring Your Own Key (BYOK)” is a bullet under the Free plan and appears under none of the paid ones. Which surface that bullet refers to is not stated, so confirm it in your own account rather than from any article.

Two bills, two units

Qoder's credits page defines credits as “the resource quota for AI models in Qoder” measuring “the computing resources consumed when Qoder products perform tasks”, and publishes per-task estimates: roughly 50 credits for an Agent-mode task at 200K context, 3–4 for Editor Ask, 7–12 for Editor Agent, 50 for Quest Agent, 75 for Quest Experts and 50 for a Repo Wiki build. It then says plainly that “the values shown in the table are just estimates” and “actual credits deductions may vary”. The same page does tie credits to tokens — “the credits deducted are calculated based on the model used and the total token consumption for both input and output” — but publishes no credit-per-token rate for any model, so the conversion stays one-way. That estimate disclaimer is why the per-task figure further down is presented as allowance drawn under a stated assumption, and never as what a task costs.

Read from docs.qoder.com/account/pricing on September 19, 2026. Two things that table hides. A top-up credit costs about a third more than an in-plan one, so running out mid-month is a price increase, not just an interruption — and the pack expires a month after purchase, so unused credits are cash already spent. Qoder's Teams and Enterprise pricing sits on a separate page that this check did not open itself; and the Alibaba Cloud Marketplace listing for Qoder for Enterprise quotes a different per-credit rate again — “additional Credits at $0.02/Credit with no commitment”, twice the in-plan rate above, against 3,000 credits per seat per month and credits “valid for 24 months from redemption”. Take the number from whichever checkout you will actually buy through.

An illustrative estimate, not a measurement. Assume one agent session that sends 600,000 uncached input tokens and receives 40,000 output tokens, with no cache reads, no cache writes and no retries. Rates are live Kunavo catalog prices per million tokens.

On the other side, Qoder's own published figures give ~50 credits for an Agent-mode task at 200K context, which at the Pro plan's effective rate is about $0.50 of allowance. Do not read those two numbers as a comparison. They price different units of different work under different assumptions, one of which Qoder itself labels an estimate that varies — and neither says anything about which tool finishes your task in fewer attempts. Cheapest listed price and lowest cost to finish the work are separate questions, and only the second one is worth budgeting.

Three mechanics worth carrying into that decision. First, Anthropic's cost documentation states that on Teams and Enterprise plans each member's Claude Code usage draws from a per-seat allowance resetting on a rolling five-hour window and a weekly window, shared with Claude chat — so a plan is not a token allowance you can price directly either.

Second, the same page publishes the only real-world spend anchor either side of this comparison offers: across enterprise deployments “the average cost is around $13 per developer per active day and $150-250 per developer per month, with costs remaining below $30 per active day for 90% of users”. That is Anthropic's reporting on its own billing, and Qoder publishes no equivalent, so it anchors one column rather than settling a comparison.

Third, Kunavo's catalog amount is a billing floor rather than a cap: when the upstream reports its charge, the bill is the greater of catalog cost and upstream cost multiplied by the applicable markup. Cache activity, tool calls and retries all sit outside the estimate above. The minimum top-up is $10 in prepaid credit, a funding minimum rather than a task fee or a subscription — see billing details.

What this page did not test

Neither product was run for this guide. Nothing here is a compatibility result or a benchmark: there is no speed, quality or task-success comparison between Qoder and Claude Code on this page because no such measurement was made. In particular, no configuration of Kunavo as a Qoder custom provider was tested under either provider type, and whether Qoder forwards prompt-cache control, thinking fields or Anthropic beta headers to a custom endpoint is undocumented — which matters for cost, so it is left open rather than guessed. Every Qoder statement above is “per Qoder's documentation”.

If the Claude Code route is the one you want, it is two environment variables and per-token billing from a prepaid balance with no subscription. Start at the Claude Code setup guide — a published configuration reference, not a compatibility test — then create a Kunavo account, run one bounded task, and compare the charge your account actually records against what you pay today. Deeper cost work lives in Claude Code pricing, the best API for Claude Code and Claude Code without a subscription. If the Qwen family Qoder bundles is what attracted you, Qwen Code pricing covers that lane directly.

FAQ

Checked September 21, 2026 by fetching docs.qoder.com (the standalone app's overview, custom-models and model-selector pages; the IDE custom-models and model-tier-selector pages; the CLI custom-models, model and authentication pages; account pricing; credits; and the release notes for all three surfaces), the Alibaba Cloud Marketplace listing, code.claude.com's gateway, MCP and cost pages, and the npm registry. Not checked here and marked as such above: Qoder's corporate ownership, its Teams and Enterprise pricing page, Claude plan prices, and any runtime behaviour of either tool. Kunavo rates are read from the live catalog; every dollar figure is illustrative token arithmetic rather than a measured task cost.