Anthropic adds server-side compaction to Messages API (beta)

Anthropic just put conversation "compression" on the wire: its Messages API beta now offers server-side compaction that summarizes prior messages so you can feed a short context to Claude without losing semantic state. This is an API-level primitive for trimming long-lived chats — not a client-side trick — and it changes what platform engineers need to log, audit, and secure.

You should be glad — and immediately suspicious. Putting compaction into the API is the right call: it prevents every team from reinventing fragile, inconsistent summarizers that leak system instructions, tokenize badly, or bias prompts. But it's also a new trust boundary. If your compliance posture treats the messages array as the immutable canonical transcript, compaction breaks that assumption.

How it works (brief)

Anthropic lets you request compaction as part of the Messages call. In the published examples the client supplies a compaction block in the request payload that instructs the model which semantics to preserve (user preferences, outstanding tasks, safety constraints) and the API returns a compacted representation you can pass back as conversation history for subsequent calls. The compaction is model-aware server-side summarization rather than a blind truncation.

A simplified example:

curl -X POST "https://api.anthropic.com/v1/messages" \
  -H "Authorization: Bearer $API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "messages": [ /* full conversation turns */ ],
    "compaction": {
      "mode": "summary",
      "preserve": ["user_preferences", "outstanding_tasks", "safety_constraints"],
      "target_tokens": 300
    }
  }'

The actual schema is more expressive; Anthropic's docs expose options for target brevity and which semantic classes to preserve. The important point is that compaction runs server-side with model-aware logic, not as a dumb truncation.

Why platform teams must change their pipelines

Audit logs: If you accept compacted blobs as the canonical history, you've just removed raw evidence. For forensics, incident response, and regulators that will be unacceptable in many environments. Keep raw transcripts or enable an append-only archive before compaction runs.
Metrics and billing: Token counts drop, latency improves, and cost models shift. Teams using tokens as a signal for user behavior, throttling, or billing must decide whether to report pre- or post-compaction usage. Reporting post-compaction without storing pre-compaction artifacts will obscure abuse signals.
Safety telemetry: Some vendors and safety teams have warned that server-side transformations can obscure the provenance of hazardous instructions. Compaction can accidentally scrub cues that detection systems rely on. Your content-moderation pipelines should ingest both the compacted context and the raw turns (or metadata) so checks don't silently degrade.
Determinism and retrievability: Summaries are lossy. If your agent orchestration or retrieval-augmented pipeline expects deterministic state (a canonical list of tasks, confirmations, or system directives), replace the assumption with explicit state objects (CRDs, database rows) rather than relying on an ephemeral summary string.

This isn't just an Anthropic problem

Vendors are increasingly baking operational primitives into model APIs (latency-tuned variants, vision upgrades, and server-side transformations like compaction). Compaction is the logical next step — everyone with long-session LLM workflows has been doing local summarization hacks. Making it an API primitive is overdue, but it shifts responsibility for provenance and observability onto platform teams.

Opinion: Do compaction in the API, but own the raw data

Anthropic did the right thing by offering compaction server-side — it's safer and more consistent than ad-hoc client summarizers. But platform teams that treat compacted output as authoritative will get burned. Store raw transcripts, emit compaction metadata (timestamp, policy used, fields preserved), and rewire safety telemetry to operate on both pre- and post-compaction text. If you don't, you'll find your forensics and compliance posture hollow the first time an incident requires precise, uncompressed context.

Expect copycats. Vendors will soon add tunable compaction, selective redaction, and compaction-aware safety labels. The practical change for platform engineers: treat compaction as another transformation in your observability pipeline, with mandatory provenance attached. If you don't, you'll be left chasing a future audit where the model's summary is all that's left — and that's a very expensive illusion.

Anthropic adds server-side compaction to Messages API (beta)

Sources

Qwen-3.8 2.4T weights released — Qwen-3.8-Max snapshot and Qwen Code v0.23.x updates

Alibaba Qwen 3.8 Max, Qwen-Image 3.0, and Qwen AgentWorld — what platform engineers must change

Anthropic internal Claude test build accessed external systems — an agent safety gap for platform teams