Seed 1.6 Guide: Adaptive Thinking and a Retirement Date

Reasoning models have a well-known waste problem: they think hard about questions that do not need it. Ask one for the capital of France and you may pay for several hundred tokens of deliberation before a one-word answer. Seed 1.6 was ByteDance's attempt to train that habit out, by letting the model decide per prompt whether to reason at all.

It is now more than a year old, and on OpenRouter it has a scheduled end date. That makes it worth understanding for two reasons: the adaptive-thinking idea still matters, and anyone running it in production has a migration to plan.

What Seed 1.6 is

ByteDance Seed described the Seed 1.6 series in a technical post dated 25 June 2025. The base model is a sparse mixture-of-experts with 230B total parameters and 23B active per token, building on the MoE work from Seed 1.5.

Pre-training ran in three stages: text only, then a multimodal stage mixing visual data with discipline, code and reasoning text, then a long-context stage that extended the sequence length from 32K to 256K. The result is a model that takes images and video as well as text, with a 256K window. OpenRouter lists it at 262,144 tokens of context and 32,768 tokens of output.

The family had siblings. Seed 1.6-Thinking was the always-reasoning variant, and Seed 1.6 Flash was the fast, cheap one. The model simply called Seed 1.6 is the adaptive one, which ByteDance also calls Seed 1.6 (Adaptive CoT).

How adaptive chain of thought works

The technique is called AdaCoT. During reinforcement learning, the reward penalises unnecessary thinking and rewards thinking where it helps, so the model learns a policy for when to reason rather than having one hard-coded. ByteDance exposes three modes through configuration:

• FullCoT — always thinks first. ByteDance says results are comparable to Seed 1.6-Thinking with shorter chains of thought.
• NoCoT — answers directly. Fastest and cheapest.
• AdaCoT — decides per prompt.

The interesting data is the trigger rate in adaptive mode. ByteDance reports the model chose to think on about 37% of MMLU questions, about 70% of MMLU-Pro, and 90 to 100% on AIME and BeyondAIME. In other words, it reasons roughly in proportion to difficulty, which is exactly the behaviour you want if you are paying for output tokens.

The practical consequence is that cost per request varies more than with a fixed-mode model. A batch of easy classification prompts and a batch of maths problems can differ several-fold in output tokens on the same endpoint. Budget on measured averages from your own traffic, not on a single test prompt, because reasoning spend lands almost entirely in output tokens.

What it was good at

ByteDance's own claims leaned on exam-style evaluation. It reported unofficial results on the 2025 Chinese Gaokao and on India's JEE Advanced, where Seed 1.6-Thinking scored full marks in maths across five sampled rounds. Those are vendor-run, image-input exam papers, and they say more about STEM reasoning over scanned questions than about agentic work.

The post also claims parity with or gains over Seed 1.5-VL on visual tasks, and support for GUI interaction, without publishing GUI figures in the text. Treat the visual strength as plausible and test it on your own images.

Why the listing date looks wrong

OpenRouter shows Seed 1.6 with a creation date of 23 December 2025, yet the technical post dates from June 2025. The date is when the listing was added. The listing's canonical slug, seed-1.6-20250625, carries a suffix that matches the June post, which tells you which build sits behind the friendly name.

A dated build is a snapshot, which is what you want for reproducible evaluations, while a friendly name can in principle be repointed to a newer build. If your outputs are tested or audited, record the build you validated against and recheck it whenever the listing changes.

The retirement date changes the decision

At the time of writing, OpenRouter lists Seed 1.6 at $0.25 per million input tokens and $2.00 per million output, doubling to $0.50 and $4.00 once a prompt exceeds 128K tokens. It also lists an expiration date of 11 November 2026 for the endpoint.

Here is the part worth noticing: Seed 2.0 Lite is listed at exactly the same prices, including the long-prompt tier. ByteDance positions Lite as its production-grade general model, balancing output quality and response speed. Same cost, a newer generation, and no retirement notice. There is little reason to start anything new on Seed 1.6.

If you want to cut cost rather than hold it flat, Seed 2.0 Mini is the other candidate. OpenRouter describes it as delivering performance comparable to Seed 1.6, at $0.10 input and $0.40 output.

A migration checklist

1. Pull a few hundred real prompts from your logs, weighted toward the ones where Seed 1.6 chose to think.
2. Run them through Seed 2.0 Lite and Seed 2.0 Mini with reasoning effort set explicitly, since the 2.0 models expose an effort control rather than an adaptive switch.
3. Compare quality, output tokens per request and latency side by side, not just pass rates.
4. Switch before 11 November, and keep the old model as a fallback only until that date. Model deprecation and migration has the longer version of this process.

The adaptive-thinking behaviour is the one thing that does not carry over automatically. If your prompts relied on the model skipping reasoning for easy cases, you now have to make that choice yourself, per route, with an effort setting.