GLM-52 897 —
GPT-56SC 873 —
CL-OP5X 865 —
GROK-46H 865 —
GEM-37FH 865 —
GPT-56T 861 —
GLM-5 856 —
MUSE-SPK 841 —
QWEN-38X 824 —
GPT-6A 820 —
KIMI-K3X 810 —
CL-FAB5H 787 —
CL-OP5H 764 —
CL-OP46H 742 —
CL-OP47H 733 —
GEM-38FH 676 —
CL-OP47 583 -0.7%
INKL 531 —
CL-OP46 496 -0.2%
CL-OP48 490 -0.2%
GLM-52 897 —
GPT-56SC 873 —
CL-OP5X 865 —
GROK-46H 865 —
GEM-37FH 865 —
GPT-56T 861 —
GLM-5 856 —
MUSE-SPK 841 —
QWEN-38X 824 —
GPT-6A 820 —
KIMI-K3X 810 —
CL-FAB5H 787 —
CL-OP5H 764 —
CL-OP46H 742 —
CL-OP47H 733 —
GEM-38FH 676 —
CL-OP47 583 -0.7%
INKL 531 —
CL-OP46 496 -0.2%
CL-OP48 490 -0.2%
← Back to feed

ByteDance Launches Seed 2.1: Elo 1539 on Arena Frontend, Claims Top GDPVal, Seedance 2.5 Ships Alongside

ByteDance launched Seed 2.1 today on Volcano Engine and Doubao, introducing four variants: Seed 2.1 Pro (flagship deep-thinking), Seed 2.1 Turbo (half the price, production-scale), Seed-Evolving (weekly model updates under a fixed API ID), and Doubao-seed-character (role-model fine-tune with video understanding and sticker recognition).

The official launch follows a preview phase that started June 19, when ByteDance posted Seed 2.1 Pro-Preview to Arena AI’s Code leaderboard.

Benchmark Numbers

The strongest public data point is the Code Arena: Frontend result. Seed 2.1 Pro-Preview ranks #8 globally with an Elo of 1539, on par with Claude Opus 4.6 in that category. It reaches the top 10 in five of seven frontend subcategories — #6 in Brand & Marketing and #7 in React. Its weakest category is raw HTML (#14), which suggests the model trades hand-coded markup fluency for polished UI generation.

On GDPVal (Artificial Analysis’s composite agentic score), Seed 2.1 Pro claims the highest score among models evaluated to date. GDPVal measures response quality across coding, agent task completion, and reasoning in live user workflows, not static benchmarks. ByteDance explicitly deprioritized static benchmark optimisation in the launch post, pointing to GDPVal and Workspace Bench as its evaluations of record.

Three capabilities anchor the release: end-to-end coding delivery (requirement analysis through result validation), general agent multi-step task execution across tools and environments, and multimodal understanding for visual and video inputs. The company did not publish SWE-bench or tau-bench scores at launch.

Pricing

ModelInputOutput
Seed 2.1 Pro¥6/M (~$0.82)¥30/M (~$4.12)
Seed 2.1 Turbo¥3/M (~$0.41)¥15/M (~$2.06)

Turbo pricing makes Seed 2.1 cheaper than GPT-5.5 ($5/M input, $30/M output) by roughly 6x on input and 7x on output at current RMB/USD rates. Neither model is available outside Volcano Engine and Doubao at this time.

Seedance 2.5

ByteDance also launched Seedance 2.5, which generates complete 30-second videos end-to-end — a step up from the segment-and-stitch approach typical of current video AI. SeedDance 2.0 had already held both Arena Video top spots at Elo 1450. Whether 2.5 extends that lead will depend on when ByteDance posts it to Arena’s video leaderboard.

The Position

ByteDance has dominated video AI and consumer chat (Doubao reached over 130 million MAUs in China), but has been absent from frontier text and coding competition until now. Seed 2.1 is the first model in the ByteDance family making a credible push at the GPT-5.5 / Claude Opus 4.8 tier — at lower cost, on Chinese infrastructure, with no export exposure. The GDPVal claim and the Arena Frontend placement are real signals. The absence of SWE-bench and tau-bench at launch leaves the full picture incomplete.