Every Doubao AI Model Explained and Compared (Nov, 2026)

The Doubao family: Doubao, Seed and Dola

Three names describe the same models. Doubao is ByteDance's consumer assistant in China and the brand its cloud, Volcano Engine, uses for the models behind it.

Seed is the ByteDance research team that builds them, which is why every current ID reads doubao-seed-....

Dola is the international name: doubao.com points visitors outside mainland China to the Dola app, and BytePlus calls Seed 2.0 Lite "the first omnimodal understanding model in the Dola large model family" in its model release notes.

Our Doubao vs DeepSeek comparison covers the app side, including its 382 million monthly users in June 2026.

The six-digit suffix on each ID is the version date in YYMMDD form. BytePlus's long-term support table lists seed-2-0-lite-260428 with a version date of 2026-04-28, so 260915 means September 15, 2026.

Two models can share a name and differ in behavior: Seed 2.0 Lite exists as 260215, 260228 and 260428 builds, and only the last one hears audio.

Doubao Seed 2.1 Pro

ByteDance announced Seed 2.1 on June 23, 2026 as "a new generation of agent-capable models built for real-world productivity", and Pro is the top of it.

The September build, doubao-seed-2-1-pro-260915, reads 1,024K tokens and writes up to 256K, chain of thought included. It costs ¥6 in and ¥30 out per million tokens with no extra charge for long prompts, and cached input drops to ¥1.20.

ByteDance does not sell Pro on BytePlus, so a team outside China cannot call it through ByteDance's own cloud. The June build, 260628, is still listed with a 256K window.

Doubao Seed 2.1 Turbo

Turbo is the version built for volume.

Volcano Engine's release note calls it a low-cost, low-latency model whose results stand alongside Pro, and it is the only Seed 2.1 model on BytePlus, where it runs as dola-seed-2-1-turbo-260628 at $0.50 in and $2.50 out.

Its BytePlus model page lists a 256K window, 256K max output and a rate limit of 500 requests a minute, low next to the 30,000 RPM of Seed 2.0 Lite. Plan for that limit before you put a busy product on it.

In China the same model costs ¥3 in and ¥15 out.

Doubao Seed 2.1 Lite

Seed 2.1 Lite arrived with the September builds and is the cheapest way into the 1M window: ¥0.80 in and ¥2.70 out, flat across the whole context.

Volcano Engine describes it as a native all-modality model that understands text, images and audio in one pass. Volcano Engine bills audio input separately, at ¥12 per million tokens. Like Pro, it is a Volcano Engine model only.

Doubao-Seed-Evolving

Seed-Evolving trades a fixed version for speed of improvement. Its Volcano Engine page logs the launch on June 23, the move to a 1,024K context on July 14, and an update on November 5 aimed at multi-step agent tasks and factual consistency.

It costs the same ¥6 / ¥30 as Seed 2.1 Pro, defaults to high reasoning effort, reads images and video, and does not take audio.

Volcano Engine also uses it as the automatic replacement when older coding models retire, which tells you where ByteDance points coding agents.

The catch is the one in the name: the model behind the ID changes under you, so pin a dated model for anything you have to reproduce.

Seed 2.0 Pro, Lite, Mini and Code

The Seed 2.0 family arrived in February 2026 in three sizes.

The Seed2.0 page says Pro "focuses on long-chain reasoning", Lite "balances output quality and response speed" and Mini is "optimized for inference throughput and deployment density".

The April 28 builds of Lite and Mini added audio input and are now long-term support (LTS) versions, with an earliest retirement date of April 30, 2027 on BytePlus.

Code Preview, tuned for IDE agents such as Claude Code, and Seed 2.0 Pro both remain on BytePlus at $0.50 / $3.00, while their February builds in China end service on November 24.

Two narrower models round out the list on Volcano Engine. ByteDance built Seed Character (260628) for role-play and companion chat, with a 128K window and ¥0.80 / ¥2.00 pricing.

Seed Translation (250915) takes 1K tokens of input and writes 3K.

Doubao pricing compared across the lineup

Output tokens drive the bill, and on BytePlus the spread runs from $0.40 per million on Seed 2.0 Mini to $3.00 on Seed 2.0 Pro and Code Preview. Seed 2.1 Turbo sits at $2.50, cheaper than the older Seed 2.0 Pro.

That makes Seed 2.0 Pro hard to justify for new work outside China.

Three billing rules change the real number:

• Prompt length tiers. On BytePlus the Seed 2.0 models double in price once a prompt passes 128K tokens: Seed 2.0 Pro goes to $1.00 / $6.00. Volcano Engine splits Seed 2.0 into three tiers (up to 32K, 32K to 128K, 128K to 256K). The Seed 2.1 models charge one flat rate.
• Flex tier. Requests sent with service_tier=flex queue on spare capacity at about half price. Seed 2.1 Turbo drops to $0.25 / $1.25.
• Cache hits. Cached input on Turbo costs $0.10, a fifth of the normal input rate, plus $0.0083 per million tokens per hour of storage.

BytePlus also sells a Coding Plan for tools such as Claude Code: Lite at $10 a month and Pro at $50 a month.

For one developer running an agent all day, a flat plan can undercut per-token billing; for a product serving customers, the token price is the one to model.

How deep thinking and reasoning effort work

Deep thinking is on by default on the current Seed models. You switch it off with thinking.type=disabled in the Chat API or reasoning.effort=minimal in the Responses API.

Between those two states, the BytePlus deep thinking guide lists seven reasoning_effort values, from none to max, but the Seed models map them onto four levels: minimal, low, medium and high, with xhigh and max treated as high.

The default is medium, except on Seed-Evolving, which defaults to high.

Agent builders need to know one more rule.

On models that return a thinking summary, the API sends a summary plus an encrypted copy of the full reasoning, and in multi-turn tool calls you must pass that encrypted content back unchanged, or ByteDance says answer quality drops.

The Responses API handles it for you through previous_response_id. Seed 1.8 and some later models also keep reasoning in the context during tool calls, which raises input token counts on long agent runs.

Legacy and retired Doubao models

Doubao versions turn over fast.

Volcano Engine's model retirement notice says a version lives three to six months in most cases, and it runs retirements in batches: quotas shrink first, new endpoints stop a few days later, and on the service end date leftover endpoints move to a named replacement or shut down.

BytePlus runs the same process on its own schedule, so one model can retire on different dates in each cloud.

Most migrations take three changes: swap the model ID, then check the thinking and reasoning_effort settings, because defaults and accepted values differ between generations.

Re-test structured output if you move onto Seed 2.0 Pro or Code Preview, which lack it on BytePlus.

Code still on the Bot API or Context API has a fourth job: Volcano Engine says those calls fail after retirement unless you move them to the Responses API.