ByteDance’s New Code Model Beats Claude and GPT‑5 on Benchmarks, Raising Anthropic Concerns

Introduction

A recent release from ByteDance, the tech giant behind TikTok, has taken the AI coding community by surprise. Their new model—often referred to as Dubau Seed Code—outperforms leading competitors such as Anthropic’s Claude and the yet‑unreleased GPT‑5 on several coding benchmarks, all while costing a fraction of the price. The rapid rise of this model may explain why Anthropic has reportedly restricted access for the Trey code editor, a ByteDance product that previously leveraged Claude.

ByteDance and Its AI Ecosystem

ByteDance is not only a social media powerhouse; it has been quietly building a suite of AI tools, including:

• Trey – an AI‑assisted code editor praised for its intuitive interface and “solo mode” workflow.
• Volcano API – a platform that exposes ByteDance’s language models to developers, albeit currently limited to Chinese users.
• Dubau Seed Code – the latest large language model (LLM) focused on software engineering tasks.

These offerings illustrate ByteDance’s ambition to compete directly with established players like OpenAI, Anthropic, and Google.

The Trey Code Editor and Its Relationship with Anthropic

Trey gained popularity for its robust code‑completion capabilities and the ability to run a variety of models, some of which were initially free. However, Anthropic abruptly cut off Trey’s access to Claude models, a move reminiscent of earlier, controversial decisions by Anthropic against other third‑party services. While the exact motivations remain opaque, internal testing suggests that Anthropic may feel threatened by ByteDance’s emerging coding model.

Benchmark Performance: SWE‑Bench Verified

One of the most respected evaluations for code‑generation models is the SWE‑Bench Verified benchmark. Anthropic has historically highlighted its performance on this test, making any challenge to its ranking particularly sensitive.

Results Overview

• Dubau Seed Code topped the leaderboard, surpassing Anthropic’s Claude‑Sonnet by roughly 8 %.
• The model also outperformed GPT‑5‑style baselines and other leading systems such as Gemini 3 checkpoints.
• Overall, Dubau Seed Code secured the 15th position among all participants, with the top four spots occupied by Gemini variants.

These results demonstrate that a relatively inexpensive model can compete with, and even exceed, premium offerings on a critical coding benchmark.

Cost and Speed Advantages

Beyond raw performance, Dubau Seed Code stands out for its affordable pricing and rapid inference:

• Pricing: $17‑$12 per million tokens (approximately 15× cheaper than Claude‑Sonnet).
• Throughput: Around 80 tokens per second, enabling near‑real‑time responses for interactive coding sessions.
• Multimodal Support: The model can process images and video, expanding its utility beyond pure text generation.

These attributes make the model attractive for developers and enterprises seeking cost‑effective AI assistance.

Accessing the Model Outside China

While the Volcano API requires a Chinese mobile number, developers worldwide can still experiment with Dubau Seed Code via ZenMox (an open‑router‑style platform). ZenMox provides:

• Free trial credits for new users.
• An Anthropic‑compatible API endpoint, allowing existing Claude‑based workflows to switch to Dubau Seed Code with minimal code changes.

This accessibility has facilitated broader community testing and contributed to the model’s rapid adoption.

Real‑World Evaluation

The author conducted a series of practical tests to gauge the model’s capabilities across different domains.

Coding and Graphics Tasks

• Floor‑plan generation: Produced correct code, though visual quality was modest.
• SVG Panda with burger: Recognizable graphics; interaction between elements could improve.
• 3‑JS Pokéball: Accurate colors and shapes; missing interactive button.
• Autoplay chessboard: Failed to function as expected.
• Minecraft‑style map (Kandinsky influence): Generated impressive depth effects and random terrain, outperforming Sonnet in visual richness.
• Butterfly animation: Smooth flight animation and appealing environment, despite a less detailed butterfly model.
• Rust CLI tool: Functioned correctly.
• Blender script: Did not execute successfully.

Overall, the model achieved a respectable 15th place on the SWE‑Bench leaderboard, especially notable given its low cost.

Agentic Benchmarks (Claw‑Code Integration)

When paired with Claw‑Code, a toolset for evaluating AI agents, results were mixed:

• Movie tracker app: Non‑functional, riddled with bugs.
• God‑game simulation: Numerous errors prevented successful execution.
• Go TUI calculator: Outstanding performance; generated a fully functional, aesthetically pleasing UI.
• Spelt app, Nux app, Open‑Code repository query: All failed to produce usable outcomes.

These outcomes placed the model at 12th overall, surpassing some commercial agents like Cursor Composer but trailing behind specialized systems such as Kimmy and Quen Code. The author notes that the model appears optimized for Trey’s workflow, and the reliance on terminal commands rather than edit‑diff operations may have hindered performance.

Implications for Anthropic and the Wider Market

The emergence of a high‑performing, low‑cost coding model from a Chinese provider challenges the prevailing narrative that premium pricing guarantees superior capability. Anthropic’s decision to restrict Trey’s access to Claude could be interpreted as a defensive maneuver to protect market share.

For developers, the key takeaway is that affordable alternatives now exist without sacrificing much in terms of quality. This shift could drive broader adoption of AI‑assisted development tools, especially among startups and cost‑conscious enterprises.

Conclusion

ByteDance’s Dubau Seed Code model delivers a compelling combination of benchmark‑leading performance, multimodal capabilities, and an exceptionally low price point. Its success on SWE‑Bench Verified and competitive results on agentic tasks demonstrate that a well‑tuned, smaller model can rival industry heavyweights like Claude‑Sonnet and the forthcoming GPT‑5.

The model’s availability through platforms such as ZenMox ensures that developers worldwide can experiment with it, potentially reshaping the landscape of AI‑driven software engineering. As the market reacts, we may see increased pressure on established providers to reconsider pricing structures and accessibility, ultimately benefiting the broader developer community.

---