Seed 2.0 Code Guide: A Coding Variant With No Scorecard

A model with "Code" in its name invites an assumption: somewhere there is a table showing it beating its general-purpose sibling on coding benchmarks. For Seed 2.0 Code, that table does not exist publicly. It carries the Seed 2.0 name and a description of what it was tuned for, but the coding scores ByteDance publishes on its Seed 2.0 page belong to the general Pro and Lite models.

That does not make it a bad model. It means you cannot pick it from a leaderboard, and the only defensible way to choose it is to test it on your own repository. This guide sets out what is actually known and how to run that test.

Where it came from

ByteDance's official Seed 2.0 page presents three general-purpose agent models, Pro, Lite and Mini, and does not mention a Code variant at all. What public information exists comes from the API listings rather than from a launch post with a scorecard.

OpenRouter added it on 12 August 2026 and describes it as optimised for agentic coding, suited to frontend development, multilingual programming and coding-agent workflows in tools such as Claude Code. The listing's canonical slug carries a 20260730 build date, and the only provider shown at the time of writing is ByteDance Seed itself, serving at FP8. No weights for it appear on ByteDance Seed's Hugging Face account.

What ByteDance has and has not published

The official Seed 2.0 page publishes coding results for the general models only. On SWE-Bench Pro it lists 46.9% for Pro (0215), 46.0% for Lite (0215) and 46.6% for Lite (0428). On Terminal Bench 2.0 it lists 55.8%, 45.0% and 43.3% respectively. There is no row for Code.

You might reasonably expect Code to land at or above Lite, since it is a specialised build on the same base. That is an expectation, not a result. Specialised coding builds sometimes trade general instruction following for code style, and a model tuned for one IDE's harness can behave differently in another. If someone quotes you a SWE-bench number for Seed 2.0 Code, ask where it came from. Vendor-claimed vs independent benchmarks explains why even published numbers deserve that question.

What the endpoint offers an agent harness

The OpenRouter listing is a good fit for coding agents on paper:

• 262,144 tokens of context and up to 131,072 tokens of output, enough for large multi-file edits in a single turn.
• Tool calling, JSON mode and structured outputs, which an agent loop needs for file edits and shell commands.
• A reasoning effort control, so you can spend less thought on mechanical edits and more on debugging.
• Image and video input, useful for implementing a UI from a mockup or a screenshot of a broken page, which fits the frontend focus in OpenRouter's description.

Pricing on OpenRouter at the time of writing is $0.50 per million input tokens and $3.00 per million output. Above 128K tokens of prompt, both double, to $1.00 and $6.00. That threshold matters more for coding agents than for most workloads, because agent sessions accumulate context with every file read and tool result. A long session can cross 128K without anyone deciding it should. Compaction or a hard context cap in the harness keeps you in the cheaper tier.

The short window

OpenRouter lists an expiration date of 11 November 2026 for this endpoint, under three months after it was added. On the same day it was listed, OpenRouter also added Seed 2.1 Turbo, which it describes as a multimodal model for coding and long-horizon agent workflows. Read together, that suggests the coding role is moving to the 2.1 generation; see the Seed 2.1 Turbo guide.

Practically, this changes the calculation. Integrating a model takes work: prompt tuning, harness quirks, cost monitoring. Doing that work for an endpoint with weeks left is hard to justify unless Seed 2.0 Code is already in your stack. If it is, use the remaining time to evaluate its replacement on the same tasks.

When to pick it over the general Seed 2.0 models

Against Seed 2.0 Lite, which lists at $0.25 input and $2.00 output, Code costs twice as much for input and 1.5 times as much for output. That premium is worth paying only if Code completes your tasks in fewer turns, because fewer turns cut the input bill that dominates agent sessions. If it does not, Lite is the better default for coding work too.

Pick Code for frontend tasks driven by visual input and for multi-language repositories, but only if your own tests confirm the listing's description. Pick Lite for mixed workloads where coding is one job among several.

How to evaluate it

1. Take ten closed issues from your own repository, each with a test that failed before the fix and passed after.
2. Run Seed 2.0 Code and Seed 2.0 Lite through the same agent harness with the same tools and the same reasoning effort.
3. Record pass rate, turns to completion, total input tokens and whether any session crossed 128K.
4. Compute cost per solved task, not cost per token. A model that is cheaper per token can be dearer per fix.

How to benchmark LLMs on your own work has the fuller method. With no public scorecard, your ten issues are the scorecard.