ByteDance’s Seed 2.0 Pro Beats GPT‑5.2 High in Math Benchmarks

ByteDance’s Seed team released the Seed 2.0 series, comprising three variants—Pro, Lite, and Mini—each optimized for different usage scenarios.

Benchmark Performance

Mathematical reasoning: Seed 2.0 Pro scored 89.8 on the MathVista benchmark, exceeding GPT‑5.2 High (83.1) and Claude Opus 4.5 (80.6). On the MathVision test it achieved 88.8, also leading the competitors.

Code generation: Seed 2.0 Pro attained a Codeforces rating of 3020, slightly below GPT‑5.2 High’s 3148 but far above Claude Sonnet 4.5’s 1485. On LiveCodeBench it received 87.8 points, essentially matching GPT‑5.2 High’s 87.7.

Model variants: Pro focuses on long‑chain reasoning and stability on complex tasks; Lite balances generation quality with response speed; Mini targets high‑concurrency and batch‑generation workloads.

Multimodal Capabilities

Official demonstrations show the model handling complex visual inputs, extracting structured information from images, and generating interactive content.

Pricing

Seed 2.0 Pro’s input cost is roughly one‑tenth of Claude Opus and about four times cheaper than GPT‑5.2 High. Seed 2.0 Lite is priced at $0.09 per million tokens. A pipeline processing 100 million tokens per day drops from $500 to $47 under the Lite pricing.

Remaining Gaps

On the SWE‑bench programming benchmark, Seed 2.0 scored 76.5 % versus Opus’s 80.9 %. On the NL2Repo task it achieved 27.9 points, well below Opus’s 43.2. Fact‑accuracy on SimpleQA‑Verified was 36.0, compared with Gemini’s 72.1.

Availability

The models are accessible via API, allowing developers to test the capabilities in real‑world applications.