Doubao Large Model 2.1 Released, Daily Average Token Calls Surpass 180 Trillion
Alina Collins
Volcano Engine released its Doubao 2.1 model series, with daily token calls hitting 180 trillion — a more-than-tenfold jump year-on-year; the flagship Pro rivals or beats GPT-5.5 and Claude Opus 4.7 on multiple benchmarks, priced as low as ¥1.96 per million tokens.
What do 180 trillion daily token calls actually mean?
As of June this year, Doubao's daily token call volume topped 180 trillion, up more than 10× year-on-year.
Volcano Engine holds 49.5% of China's public-cloud MaaS market — model-as-a-service, selling AI model access to enterprises the way utilities sell electricity.
This means → Doubao is no longer a lab project. It is China's most-used commercial large-model platform, handling nearly half the market.
How low can the price go?
Pro pricing: ¥6 per million input tokens, ¥30 per million output tokens; with cache hits, input drops to just ¥1.2.
In coding and agent workflows combined, the effective cost compresses to ¥1.96 per million tokens.
Turbo cuts Pro's price by half, targeting high-frequency call scenarios.
In plain terms = for an enterprise writing code or running AI agents, a million tokens costs less than two yuan — the price war has hit the floor.
Can it really match GPT-5.5 and Claude Opus 4.7?
On NL2Repo-Bench and SciCode, Doubao 2.1 Pro beat GPT-5.5 and Claude Opus 4.7 respectively; its SciCode score of 59.8 covers real research problems across math, physics, chemistry, biology, and materials science.
On agent tasks, Doubao 2.1 Pro scored highest on OpenAI's GDPval benchmark; on ALE — covering 13 industry clusters and over 1,000 high-value real-world tasks — it topped Claude Opus 4.7.
It also reached global SOTA on tool-use (MCP-Atlas), image understanding (MMMU-Pro), and video temporal reasoning (TOMATO / LVBench).
This means → on public benchmarks at least, Doubao 2.1 Pro now sits in the same tier as OpenAI's and Anthropic's flagships — the "catching up" narrative no longer fits.
Where are video and audio headed next?
Video-generation model Seedance 2.5 is in late beta, with a full launch expected in early July; single-clip length reaches 30 seconds, with up to 50 multimodal assets as joint input.
Audio-generation model Doubao-Seed-Audio 1.0 launched the same day, producing up to 2 minutes of complete audio with multi-character dialogue, background music, and sound effects.
In plain terms = text, image, video, audio — Doubao is filling in one track at a time, aiming for a single model that can produce everything.
Can 180 trillion calls turn into real revenue?
Doubao 2.1 has integrated with partners including WPS, Dedao, and Unity, with plans to reach hundreds of millions of Doubao users.
A rolling-update version, Doubao-Seed-Evolving, ships upgrades 2–4 times per month; enterprises need not change their API endpoints.
It is already compatible with Claude Code, Codex, and other mainstream harness frameworks, plus Volcano's own tools TRAE and Coze.
This reflects a strategy of locking developers into the ecosystem with low prices and rapid iteration — but whether call volume can convert into sustainable revenue remains the central open question.
Content is for reference only, not financial advice.