Doubao-Seed-Code: ByteDance’s Coding Model Explained

Doubao-Seed-Code is ByteDance’s dedicated coding model, and its most practically useful property is not a benchmark score — it is that it speaks the Anthropic API format. That one compatibility decision means you can point Claude Code, Cline, Cursor, Codex CLI or veCLI at it with a base URL and a key, rather than rewriting an integration.

Released on 11 November 2025 through Volcengine’s Ark platform, it sits alongside the general Doubao models as a specialist, priced well below them. Here is what it is, what it costs, what ByteDance claims for it, how much of that claim you should weight, and what a developer outside China can actually do with it.

What it is

Doubao-Seed-Code is a model trained for software engineering rather than general conversation. ByteDance released it on Volcengine’s Ark platform — the mainland Chinese model service — on 11 November 2025, with a 256K-token context window.

The 256K figure matters more for coding than for chat. A general assistant rarely needs a quarter of a million tokens; an agent working across a repository routinely does, because it is holding file trees, diffs, test output and its own prior reasoning in the same window. Anything much smaller forces the harness to do aggressive context management, which is where a lot of agentic coding tools lose accuracy.

It has since been folded into the numbered generations. The Doubao-Seed-2.0 series, announced on 14 February 2026, consists of three general multimodal agent models — Pro, Lite and Mini — plus a dedicated Doubao-Seed-2.0-Code. So the coding line is now versioned alongside the flagship rather than sitting outside it. Our model family guide maps the generations.

The Anthropic-compatibility decision

This is the part worth reading even if you never intend to use the model.

Doubao-Seed-Code is compatible with the Anthropic API format and works with Claude Code, veCLI, Cursor, Cline and Codex CLI. Developers can point existing tooling at it with minimal configuration change — typically a base URL and an API key.

Almost every model provider offers OpenAI-shaped endpoints, because that has become the default wire format for chat completions. Anthropic compatibility is a different and more specific bet: it targets the agentic coding harnesses that grew up around Claude’s tool-use format. ByteDance is not competing for the chat-completions call. It is competing to be the model inside a coding agent someone has already chosen.

The international platform makes the same bet explicitly. BytePlus ModelArk’s documentation states that its API is compatible with both OpenAI and Anthropic interface protocols, listing https://ark.ap-southeast.bytepluses.com/api/v3 as the OpenAI-protocol base URL — for Chatbox, Cherry Studio, Cline, Cursor, Kilo Code, Roo Code, OpenCode and others — and https://ark.ap-southeast.bytepluses.com/api/compatible as the Anthropic-protocol base URL for Claude Code.

Pricing, and why it is cheap

Set that against the general flagship: Doubao-Seed-2.1 Pro is CNY 6.00 input and CNY 30.00 output per million tokens. The coding model is roughly a fifth of the input price and a quarter of the output price at the short-prompt tier. For an agent that makes many calls per task, that difference compounds fast.

Two things about the structure are worth planning around. The tiering is by prompt length, so a session that drifts past 32K tokens gets more expensive per token — trimming context is worth real money rather than merely being tidy. And output is priced six to seven times higher than input at every tier, which is the usual shape but especially punishing for a coding model that emits long diffs.

There is no official USD rate card for Doubao-Seed-Code. These are mainland-China SKUs in yuan. Internationally, the equivalent is listed under a Seed name — see below.

The 78.80% benchmark claim, read carefully

ByteDance reported that Doubao-Seed-Code scored 78.80% on SWE-bench Verified, describing it as a new state of the art. That number appears in a great deal of English coverage without the qualification that follows it, so here is the qualification.

The score was achieved running inside ByteDance’s own TRAE development environment. It is not a bare-model score. SWE-bench Verified results are notoriously harness-dependent: the same weights can differ by many points depending on the scaffolding around them — how files are retrieved, how tests are run, how many attempts are allowed, how errors are fed back. A vendor reporting a vendor-harness score is reporting the performance of a system, not of a model.

That does not make it dishonest or worthless. It makes it non-comparable to a number produced by a different harness, which is exactly how it is usually used. When you see 78.80% quoted next to another model’s SWE-bench figure, check whether both ran the same scaffold. Usually they did not.

For contrast, the one ByteDance coding-adjacent number anyone can independently reproduce is the open-weight one: the model card for Seed-OSS-36B-Instruct reports 56 on SWE-bench Verified under the OpenHands harness. Self-reported, but the weights are public under Apache-2.0, so anyone can check it. And as of 23 August 2026, no Doubao or Seed model appears on the public arena.ai leaderboard at all, so there is no third-party placement to fall back on.

Using it from outside China

The coding model is reachable internationally, but not under the Doubao name and not from everywhere.

One further gap for planning purposes: BytePlus’s Billing Center documentation lists 64 countries and regions where a card payment method can be added, and as of 23 August 2026 the mainland United States is not among them — the only US-affiliated entry is the United States Virgin Islands. Canada, India, Brazil and mainland China are also absent. Singapore, the United Kingdom, Germany, Japan, Australia and Hong Kong are included.

How the price compares

Cheapness is the clearest argument for this model, so it is worth putting numbers beside numbers. Every figure below is from the vendor’s own published rate card.

The yuan conversions are indicative only — we include them for scale, not for budgeting, and exchange rates move. Read the CNY figures as the real ones.

What the table shows is that ByteDance is not alone at the cheap end any more. GPT-5.6 Luna at $0.20 input and $1.20 output is in the same territory with a four-times larger context window, and BytePlus’s own seed-2-0-mini undercuts everything at $0.10. The specific case for the coding model is not that it is the cheapest token available; it is that it is a purpose-trained coding model at general-purpose-cheap prices, with a harness compatibility that removes the switching cost.

When it is worth evaluating

We have not run this model against a workload, so what follows is a framework for your own evaluation rather than a recommendation.

• You already run an agentic coding harness. The Anthropic-format compatibility only pays off if you have a Claude Code, Cline or Cursor setup to point at it. Starting from nothing, the compatibility is worth little.
• Your volume is high enough for price to matter. At a few hundred calls a month the difference between CNY 1.20 and $4.00 per million input tokens is noise. At agent scale, where a single task can make dozens of calls carrying a large context, it is the whole budget.
• You are not in the United States. BytePlus ModelArk excludes the US from both its availability list and its card-payment list, which makes the international route awkward for American developers regardless of the model’s merits.
• Your work is not subject to data-residency constraints you would struggle to justify. ModelArk’s own region page warns that although requests are primarily routed to the region where the endpoint is created, some may be routed to inference resources in other regions. That is worth reading before you commit regulated code to it — see our data-handling article and the API guide.
• You can tolerate an unverifiable quality claim. The honest position is that nobody outside ByteDance has published a comparable measurement, so an evaluation on your own repository is the only meaningful test.

What we cannot tell you

• We have not benchmarked it. We are not a model-evaluation lab, and every performance figure in this article is ByteDance’s own, reported with ByteDance’s name attached.
• There is no independent placement. No Doubao or Seed model appears on the public arena.ai leaderboard, so the vendor claims cannot be triangulated against a third party.
• The SWE-bench figure is harness-bound. 78.80% was measured inside ByteDance’s own TRAE environment and is not directly comparable to a number from another scaffold.
• Model IDs move. The naming diverges between Volcengine (doubao-*) and BytePlus/OpenRouter (seed-*), and date-suffixed IDs change with each release. Check the live catalogue before writing one into code.
• We could not find a published USD rate card for the coding model specifically. The prices above are mainland-China yuan SKUs.

Frequently asked questions

Where to go next

For endpoints, keys, regions and the full pricing picture, read the Doubao API guide. For how this model sits among the others, see the models behind Doubao. If you are evaluating ByteDance against alternatives, Doubao vs DeepSeek compares developer pricing directly. New to the subject? Start with what Doubao is or our complete Doubao guide.

Primary sources referenced in this article: Volcengine’s Ark model documentation, BytePlus ModelArk’s documentation, and OpenRouter’s ByteDance Seed listings.

Last updated 23 August 2026. Every product claim in this article is traced to a primary source — ByteDance and Volcengine documentation, official app-store listings, published policies, or a named research firm with a date attached. Prices and model IDs reflect the published catalogues on 23 August 2026. Doubao changes quickly; if something here has gone stale, write to [email protected] and we will correct it. See our editorial policy for how we source and correct.