Claude Dreams Now Supports Fable 5 and Sonnet 5: The Model-Tier Routing Decision for Agent Memory

When Anthropic launched the Dreams API in April 2026, operators had no routing choice: agent memory consolidation ran on claude-opus-4-8, period. As of July 10, that constraint is gone. Dreams now supports Claude Fable 5 and Claude Sonnet 5, and the decision of which model to use for your memory jobs is now an operator policy — with real cost implications attached.

What changed

Dreams is the asynchronous background job in Anthropic's Managed Agents platform that reads an existing memory store alongside past session transcripts, then produces a reorganized output store: duplicates merged, stale entries replaced, new insights surfaced. It does not modify the input store, so the output is reviewable before you swap it in.

Until July 10, every POST /v1/dreams request had to pass model: "claude-opus-4-8". The July 10 platform release notes confirm that Fable 5 and Sonnet 5 are now accepted in the model field of a dream creation request, at the same API shape.

Nothing else changed in the Dreams API. The required beta headers remain managed-agents-2026-04-01,dreaming-2026-04-21. Sessions and memory store IDs work the same way.

Why this matters for AI engineering teams

On paper, more model options is always good. In practice, Dreams just became the first Managed Agents primitive where teams have to make an active routing decision about cost versus quality — and the cost difference is not trivial.

Consider a team running 10 consolidation jobs per day on a mid-size agent deployment, each job reading 50 sessions. At Opus 4.8 pricing (the previous only option), the token cost per dream is dominated by session transcript ingestion at the input side. Switching to Sonnet 5 at its introductory pricing can reduce that cost by 70–80% per job. Fable 5 sits in between on price but offers stronger reasoning about what is stale versus genuinely updated knowledge.

The choice depends on what your memory store contains:

• Simple fact-and-preference stores (user name, preferred language, active project slugs): Sonnet 5 handles these well and the cost savings are large.
• Domain knowledge stores (architecture decisions, code patterns, nuanced long-term reasoning outcomes): Fable 5 is worth the premium — it is less likely to incorrectly merge contradictory entries or collapse nuanced distinctions.
• High-stakes production agent memory where stale decisions can cause incidents: Opus 4.8 remains the safe default. Its superior instruction-following reduces the risk of the dream incorrectly discarding a critical constraint.

The router/operator angle

This upgrade introduces a class of routing decision that did not exist before in the Managed Agents stack: model selection for background, non-interactive AI jobs.

Most API gateway routing focuses on interactive, latency-sensitive calls. Dreams are async and non-interactive — the user is not waiting for a response. That changes the tradeoff. Latency matters almost not at all; token cost and consolidation quality are the dominant variables.

For teams running multi-tenant agent deployments, this creates an opportunity to tier your memory maintenance by account type:

• Free/low-tier accounts: Sonnet 5 dreams to keep infrastructure cost down.
• Pro/paid accounts: Fable 5 dreams to maintain higher memory coherence over time.
• Enterprise accounts: Opus 4.8 for guaranteed consolidation quality SLA.

Your gateway or agent orchestration layer should treat model in dream creation requests as a routing policy field, not a static constant. Teams using TheRouter to manage their Anthropic API traffic can apply the same per-account model-tier rules to dream jobs as they already apply to interactive chat sessions.

What to do now

If you are using Dreams with a hardcoded claude-opus-4-8: audit whether your consolidation jobs actually need Opus-class reasoning. Most fact-and-preference memory patterns do not. A one-time test comparing a Sonnet 5 dream output against an Opus 4.8 dream output on the same memory store will tell you quickly.

If you are building Dreams into a product for the first time: do not default to Opus 4.8. Start with Fable 5, validate output quality against your memory store type, and use Sonnet 5 for lower-complexity stores. Opus 4.8 should be the escalation path, not the default.

If you are building a multi-tenant platform on Managed Agents: add dream_model to your per-account configuration schema now, before it becomes a painful refactor.

Check the Dreams documentation for the full supported model list and the API shape for creating and polling dreams. See TheRouter's model documentation for current Fable 5, Sonnet 5, and Opus 4.8 pricing and capability profiles before making the tier decision for your deployment.