Models

Rebyte Agents use one of three model families: autoregressive LLMs, diffusion-based LLMs, and decision models. Pick the family first, then choose a model ID within it.

Use the exact model ID below when creating or updating an Agent. The same catalog is used by the web app, iOS, and Agent API. The API requires an explicit model when defining an Agent; new product UI Agents start with deepseek-flash.

Autoregressive LLMs

Conventional language models that generate one token at a time, left to right. This is the default family for most workloads.

GPT-6 Astra requires paid-model access. Its reasoning.effort accepts low, medium, high, xhigh, and max; none and minimal are rejected. Omit the effort to use the provider's default.

Claude Sonnet 5.5

Use claude-sonnet-5-5 to select Sonnet 5.5. Sonnet 5 remains available under claude-sonnet-5; existing Agents keep their saved model until you update it. Sonnet 5.5 requires paid-model access. Its reasoning.effort accepts low, medium, high, xhigh, and max; omit it for the provider default. The Agent runtime uses a 200,000-token context budget and a 64,000-token output limit, below the provider's maximums.

Rebyte rates per million tokens are $3 input, $15 output, $0.15 cache reads, and $3.75 five-minute cache writes. The Agent manages adaptive thinking and signed history across tool calls, Session continuation and context compaction. Clients consume the normal Agent events and do not need to manage provider thinking blocks.

Claude Opus 5.5

Use claude-opus-5-5 to select Opus 5.5. Opus 5 remains available under claude-opus-5; existing Agents keep their saved model until you update it. Opus 5.5 requires paid-model access. Its reasoning.effort accepts low, medium, high, xhigh, and max; omit it for the provider default. Thinking is always on. The Agent runtime uses a 200,000-token context budget and a 64,000-token output limit, below the provider's maximums.

Rebyte rates per million tokens are $6 input, $30 output, $0.30 cache reads, and $7.50 five-minute cache writes. The Agent manages adaptive thinking and signed history in the same way as Sonnet 5.5.

Diffusion-based LLMs

A diffusion language model does not emit tokens one at a time. It produces a block of text in parallel and refines the block over a few denoising steps, so long answers can finish faster than with an autoregressive model. Streaming still works; deltas arrive in larger chunks because whole blocks are committed at once.

Mercury 3

Use mercury-3 to select Inception Mercury 3. It accepts text input and produces text output only; it is not multimodal. The Agent runtime uses a 200,000-token context budget and a 64,000-token output limit, below the provider's one-million-token context.

Its reasoning.effort accepts low, medium, high, and none. Omit the effort to use the provider default, which is medium.

Rebyte rates per million tokens are $0.15 input, $0.525 output, and $0.015 cache reads.

Diffusion refinement steps are billed as output tokens by the provider, so a short visible answer can bill more output tokens than its length suggests. Compare costs with your real prompts before migrating high-volume workloads.

Decision models

Decision models do not generate text and cannot be selected as an Agent model. They evaluate semantic questions and return probabilities that a Dynamic Workflow branches on. Rebyte exposes TypeSafe Jev through the Decision tool.

Moving from retired model IDs

From October 2, 2026, new Agent API requests must use the current non-GPT IDs. Retired IDs are rejected as unsupported models; they do not silently route to a different model. Update model strings in your application and agent.toml. Existing saved Agents and their executable Session configurations have been migrated as follows:

These changes apply to the Agent model. Claude Code and Codex model settings are separate and remain unchanged. Historical usage and charges retain their original model records.

deepseek-flash uses DeepSeek's explicit Flash route. Rebyte's published Flash rates are $0.45 per million uncached input tokens, $0.009 per million cached input tokens, and $1.80 per million output tokens. These are fixed Rebyte rates, based on DeepSeek's peak list prices with Rebyte's proxy margin; they do not vary with upstream off-peak discounts.

Model access depends on organization credits and subscription policy.

Rebyte Loop (Preview) adds the trainable qwen3.6-35b-a3b base model and organization-owned model families (ftm_) and versions (ftv_). See versions and publishing for private models and controlled rollout.