Seed 2.0 Pro Review: 98.3 AIME 2025, $0.47/1M (2026)
ByteDance Seed 2.0 Pro (February 2026): 272K context, 98.3 AIME 2025, 88.9 GPQA Diamond, 76.5 SWE-bench Verified. Omni-modal agent, $0.47/$2.37 per 1M.
Seed 2.0 Pro is ByteDance's flagship omni-modal agent model released February 14, 2026, scoring 98.3 on AIME 2025, 88.9 GPQA Diamond, 76.5 SWE-bench Verified, and 87.0 MMLU-Pro with a 272K-token context window and 3020 Codeforces rating. Priced at $0.47 per 1M input and $2.37 per 1M output tokens via Volcengine and BytePlus, it ranks 6th on the LMSYS Chatbot Arena overall leaderboard and 3rd on the Vision Arena.
Seed 2.0 Pro, released February 14, 2026 by ByteDance, is a flagship omni-modal agent model scoring 98.3 on AIME 2025, 88.9 GPQA Diamond, and 76.5 SWE-bench Verified with a 272K-token context window. It natively understands text, image, audio, and video. Priced at $0.47 per 1M input and $2.37 per 1M output tokens, it ranks 6th on the LMSYS Chatbot Arena overall leaderboard and 3rd on the Vision Arena.
Provider: ByteDance · Family: Seed 2.0
Context window: 272,000 tokens · Max output: 131,072
Input modalities: text, image, audio, video · Output: text, tool-calls
About Seed 2.0 Pro
Seed 2.0 Pro is ByteDance's flagship general-purpose omni-modal agent model, released February 14, 2026. It is the first model in the Seed foundation model series to natively unify understanding across text, image, audio, and video within a single inference pass. The architecture uses a Transformer backbone with a Mixture-of-Experts design following the Seed family pattern, though ByteDance has not disclosed precise parameter counts for the Pro variant. The model powers the flagship tier in the Doubao consumer app and the TRAE coding environment, sitting above the Seed 2.0 Lite and Mini variants in the lineup. On the 2026 benchmark stack, Seed 2.0 Pro scores 98.3 on AIME 2025, placing it in the global top tier for competition-level mathematics, alongside 88.9 GPQA Diamond, 76.5 SWE-bench Verified, and 87.0 MMLU-Pro. On competitive programming, it achieves a 3020 Codeforces rating approaching International Master level. On VideoMME, it scores 89.5, ranking 3rd on the LMSYS Vision Arena as of February 16, 2026. The overall LMSYS Chatbot Arena places it 6th in the text category. For context, SWE-bench Verified at 76.5 trails Claude Opus 4.5 at approximately 80.9, meaning Seed 2.0 Pro is a strong coder but cedes the top coding benchmark to Anthropic's flagship by roughly 4 points. Seed 2.0 Pro supports a 272K-token context window. The companion Seed 2.0 Lite confirms 262,144 tokens and 131,072 max output, suggesting similar constraints apply to Pro. ByteDance highlights coherent understanding and high-precision reasoning over hour-long videos with streaming real-time analysis, indicating strong long-context performance for video-centric workloads. No independent needle-in-haystack evaluation for the Pro variant has been published as of June 2026. The modality set is genuinely omni-modal on the input side: text, image, audio, and video are all accepted natively. Audio does not require pre-transcription by a separate ASR model. Video is accepted as a URL or encoded bytes through the Volcengine API. Function calling and tool-augmented execution are core features, with ByteDance describing Seed 2.0 Pro as an agent-era model optimized for multi-step workflows. GUI operation capability is noted in the Seed 2.0 series materials. Output is text and tool-calls only. At $0.47 per 1M input tokens and $2.37 per 1M output tokens, Seed 2.0 Pro is approximately 6 to 8 times cheaper than GPT-5 or Claude Opus 4.5 for equivalent reasoning tasks. A daily coding agent using 1M input and 200K output tokens costs roughly $0.94 per day, compared to $5-7 for comparable Western models. No batch API discount or prompt caching tier has been confirmed as of the launch date, limiting savings on repeat-context workloads compared to Anthropic's 90% cached input discount. API access is via Volcengine (ByteDance's China cloud platform), BytePlus (international enterprise), OpenRouter, and DeepInfra. No AWS Bedrock, Google Vertex, or Azure deployment exists as of the February 2026 launch. The model is API-only with no open weights. Python, Go, Java, and TypeScript SDKs are available through the Volcengine client libraries, and OpenRouter provides an OpenAI-compatible endpoint for easier migration from Western providers. ByteDance has not published a system card for Seed 2.0 Pro comparable to those from Anthropic or Google. The training data cutoff is January 2024, older than comparable Western frontier releases. RLHF alignment is used; red-teaming partners are not publicly named. SOC 2 Type II, ISO 27001, and HIPAA certification status for BytePlus have not been confirmed, making Seed 2.0 Pro less suitable for regulated industries without a direct enterprise agreement review. Seed 2.0 Pro is the right choice for research and engineering teams needing top-5 benchmark math and science performance, cost-sensitive frontier reasoning at 6-8x savings, or video analytics capabilities that outclass most Western rivals. It is not the right choice for teams with strict US or EU data sovereignty requirements, those needing enterprise SLAs comparable to AWS or Azure, or anyone querying post-January 2024 events without retrieval augmentation. For regulated industries, GPT-5 on Azure or Claude Opus 4.5 on Bedrock remain safer options. The Seed 2.0 series is a major generational step from Seed 1.5, which introduced Seed1.5-Thinking (chain-of-thought reasoning) and Seed1.5-VL (vision-language with 20B active MoE parameters). Seed 2.0 Pro integrates the multimodal strengths of both into a single unified omni-modal system. ByteDance upgraded Seed 2.0 Lite in late April 2026, indicating ongoing development velocity. No Seed 2.0 Pro-specific patch has been documented between February and June 2026, and no formal Seed 3.0 roadmap has been announced.
Pricing
$0.47 per 1M input tokens and $2.37 per 1M output tokens via Volcengine reference pricing. BytePlus international pricing may vary. No prompt caching tier confirmed as of February 2026.
Key Features
- Omni-Modal Understanding: First model in the Seed family to natively process text, image, audio, and video in a single inference pass, with no external preprocessing required.
- 98.3 AIME 2025 Math Reasoning: Near-perfect score on competition-level mathematics, placing Seed 2.0 Pro in the global top tier for quantitative reasoning tasks.
- 272K Token Context Window: Handles long documents, codebases, and hour-long videos within a single prompt with coherent reasoning across the full context depth.
- 3020 Codeforces Rating: Reaches International Master-level competitive programming performance, reflecting high accuracy on algorithmic problem solving at contest difficulty.
- 6-8x Cost Advantage vs Western Frontier Models: At $0.47/$2.37 per 1M tokens, delivers top-5 benchmark performance at a fraction of GPT-5 or Claude Opus 4.5 pricing for reasoning-heavy workloads.
Pros
- 98.3 AIME 2025 and 88.9 GPQA Diamond place it in the global top 3 for math and graduate-level science reasoning.
- Omni-modal video understanding with 89.5 VideoMME enables hour-long video analysis that most Western rivals cannot match.
- $0.47/$2.37 per 1M tokens is 6-8x cheaper than GPT-5 or Claude Opus 4.5 for equivalent reasoning-heavy workloads.
Cons
- API infrastructure is primarily China-based (Volcengine) with thinner enterprise SLA coverage in US and EU than AWS Bedrock or Azure OpenAI.
- No confirmed prompt caching, raising costs for repeat-context agent loops compared to Anthropic's 90% cached input discount.
- Training cutoff of January 2024 is older than recent Western frontier releases, requiring RAG for post-cutoff factual grounding.
Benchmarks
- mmlu pro: 87
- aime 2025: 98.3
- video mme: 89.5
- gpqa diamond: 88.9
- lmarena rank: 6
- codeforces rating: 3020
- swe bench verified: 76.5
Frequently Asked Questions
What is Seed 2.0 Pro and who built it?
Seed 2.0 Pro is ByteDance's flagship foundation model, released February 14, 2026, as the top variant in the Seed 2.0 series alongside Lite, Mini, and Code. Built by the ByteDance Seed team, it is the first model in the Seed family to achieve full omni-modal understanding, natively processing text, image, audio, and video in a single inference pass. ByteDance uses a Transformer architecture with a Mixture-of-Experts design, following the pattern of Seed 1.5-VL (which used 20B active parameters in a MoE configuration), though precise parameter counts for Pro remain undisclosed. On benchmarks, it scores 98.3 on AIME 2025, 88.9 GPQA Diamond, 76.5 SWE-bench Verified, and 87.0 MMLU-Pro, placing it in the global top 5 for math and science reasoning. It was designed for the agent era: long-horizon multi-step tasks where models orchestrate tools, process rich media, and reason over extended 272K-token context windows. The model powers ByteDance's Doubao consumer app and TRAE coding environment, and is available via API on Volcengine and BytePlus. Compared to GPT-5, Claude Opus 4.5, and Gemini 2.5 Pro at similar benchmark levels, Seed 2.0 Pro targets a 6-8x cost advantage at $0.47/$2.37 per 1M tokens.
How much does Seed 2.0 Pro cost per 1M tokens?
Seed 2.0 Pro is priced at $0.47 per 1M input tokens and $2.37 per 1M output tokens via Volcengine reference pricing as of February 2026. BytePlus international pricing may vary slightly by region and enterprise agreement tier. No batch API discount or prompt caching tier has been publicly confirmed for Seed 2.0 Pro, unlike Claude (which offers 50% batch discount and 90% cached input savings) or GPT-5 (which supports prompt caching). Worked cost examples: analyzing a 100K-token research document costs approximately $0.05; a daily coding agent consuming 1M input and 200K output tokens costs $0.94 per day; processing 1,000 customer support turns at 2K input and 500 output per turn costs approximately $2.12. Compared to GPT-5 (approximately $15/M output) and Claude Opus 4.5 (approximately $15/M output), Seed 2.0 Pro delivers roughly 6x output cost savings for equivalent reasoning tasks. For teams in China accessing via Volcengine, Yuan-denominated pricing may provide further savings. No self-hosted option exists, so all compute costs are API-based.
What is Seed 2.0 Pro's context window and max output?
Seed 2.0 Pro supports a 272K-token context window, enabling analysis of long documents, large codebases, and hour-long videos within a single prompt. The companion Seed 2.0 Lite variant has a confirmed 262,144-token (2^18) context and 131,072-token max output, suggesting similar constraints apply to Pro. ByteDance specifically highlights 'coherent understanding and high-precision reasoning over hour-long videos with streaming real-time analysis,' indicating the long-context architecture is optimized for video-centric workloads. No independent needle-in-haystack evaluation has been published for the Pro variant's long-context recall, unlike Claude Opus (99% accuracy at 200K) or Gemini 2.5 Pro (published results at 1M context). For comparison, GPT-5 supports up to 256K context and Claude Opus 4.5 supports 200K, making Seed 2.0 Pro's 272K competitive at the top of the 2026 frontier. No extended-context tier beyond 272K is documented as of the launch date. PDF and multi-file document inputs are accepted natively via the API.
How does Seed 2.0 Pro compare on benchmarks vs GPT-5 and Claude Opus 4.5?
On AIME 2025, Seed 2.0 Pro scores 98.3, placing it in the global top tier for competition-level mathematics; GPT-5 and Claude Opus 4.5 have not published directly comparable AIME 2025 scores at this level. For GPQA Diamond (graduate-level science reasoning), Seed 2.0 Pro achieves 88.9, competitive with the frontier but near the level of Gemini 2.5 Pro (~90.0) on this benchmark. On SWE-bench Verified (agentic coding), Seed 2.0 Pro scores 76.5, behind Claude Opus 4.5 at approximately 80.9, meaning Claude wins the coding benchmark by about 4 points, equivalent to roughly 1 in 25 coding tasks going wrong for Seed where Claude would succeed. MMLU-Pro at 87.0 sits within the frontier pack alongside Gemini and GPT-5 variants. On VideoMME, Seed 2.0 Pro scores 89.5, a strong result that outperforms most text-focused Western rivals on video-native tasks. The LMSYS Chatbot Arena places it 6th overall and 3rd on Vision Arena as of February 16, 2026, reflecting strong crowd-sourced human preference. Benchmark caveat: several scores are vendor-reported and independent third-party verification has not been fully completed for all Seed 2.0 Pro results as of the upload date.
Is Seed 2.0 Pro open source or proprietary?
Seed 2.0 Pro is fully proprietary: model weights are not released and it is accessible only through ByteDance's commercial APIs. Access is available via Volcengine (ByteDance's cloud, primarily serving China), BytePlus (ByteDance's international enterprise cloud), OpenRouter (global aggregator), and DeepInfra (aggregator). Unlike Meta's Llama 4 or DeepSeek's MIT-licensed releases, there is no HuggingFace download, no quantization option, and no self-hosting path. The license is ByteDance's commercial API terms, governed primarily under Chinese law for Volcengine and under a separate BytePlus agreement for international users. There are no open-source variants of Seed 2.0 Pro, though ByteDance has released some earlier Seed model components for research use, so future openness cannot be ruled out. For teams requiring open weights for air-gapped deployment, compliance, or fine-tuning, alternatives include Llama 4 Scout/Maverick or Qwen3-72B. Commercial use is permitted under BytePlus API terms for most standard applications.
What modalities does Seed 2.0 Pro support?
Seed 2.0 Pro is ByteDance's first omni-modal understanding model, accepting text, image, audio, and video inputs natively in a single prompt. On the output side, the model produces text and tool-call outputs; it does not generate images, audio, or video (those are handled by separate Seed models: Seedance 2.0 for video generation, Seedream for image generation). Vision capabilities include image understanding, chart reading, spatial reasoning, and multi-image comparison. Audio input supports spoken language reasoning natively without pre-transcription by a separate ASR model. Video input is a core differentiator: the model handles hour-long video with streaming output and scores 89.5 on VideoMME, outperforming most Western rivals on video-native tasks. Function calling and tool use are core features, designed for agent-era multi-step task execution with structured JSON output support. PDF and document inputs are accepted via the API for long-document analysis workflows.
Does Seed 2.0 Pro train on user data?
ByteDance states that Seed 2.0 Pro does not train on API inputs by default, consistent with standard enterprise API terms. The default data retention policy for Volcengine and BytePlus API calls follows ByteDance's standard terms of service; the exact retention window is not specified in public documentation as of February 2026. Enterprise zero-retention options are available under BytePlus enterprise agreements and must be negotiated directly with ByteDance. SOC 2 Type II, ISO 27001, and HIPAA certification status for the BytePlus API have not been publicly confirmed, unlike AWS Bedrock or Google Vertex which inherit full cloud certifications. GDPR compliance for international BytePlus API calls should be verified directly with ByteDance, as data flows to Chinese infrastructure introduce additional considerations for EU data subjects. There is no EU AI Act systemic risk declaration filed by ByteDance as of the February 2026 model release. Teams in regulated industries (finance, healthcare, legal) should conduct a formal data processing agreement review before deploying Seed 2.0 Pro in production.
Who is Seed 2.0 Pro best for and who should avoid it?
Seed 2.0 Pro is the best choice for research teams and data scientists needing top-tier math and science reasoning (98.3 AIME 2025, 88.9 GPQA Diamond) at 6-8x lower cost than Western frontier alternatives. Video analytics teams and media companies benefit from its class-leading 89.5 VideoMME score and hour-long video comprehension, an area where GPT-5 and Claude Opus 4.5 are materially weaker. Cost-sensitive engineering teams running high-volume agent workloads get the most from its $0.47/$2.37 per 1M pricing at top-5 benchmark quality. Teams that should avoid it include those with strict US or EU data sovereignty requirements, since ByteDance's compliance certifications are not publicly documented and infrastructure is primarily China-based. Enterprise teams requiring SLAs with clear uptime guarantees and legal accountability under US or EU law should use AWS Bedrock-hosted GPT-5 or Azure OpenAI for Claude Opus 4.5. Voice-first product teams cannot use Seed 2.0 Pro for audio output and will need a separate TTS model. Teams needing knowledge of events after January 2024 without retrieval augmentation should note the training cutoff, where Claude Opus 4.5 and Gemini 2.5 Pro have more recent knowledge.