Anthropic's late-April drop landed quieter than the version bump suggests. Two things shipped together — Claude Opus 4.7 and Managed Agents in public beta — and one thing got pulled: the 1M-token context window for Sonnet 4.5 and Sonnet 4, retired April 30. If you're running production agents on the Claude stack, this week is a decision point, not a download-and-go.
We've been testing all three since the beta headers went live, building the same workloads we ship for clients — client triage for law firms, internal knowledge agents for consultancies, WhatsApp bots that route to human handlers. Here's the practitioner cut: what we'd migrate this week, what we'd pilot before committing, and what we'd quietly ignore.
What actually shipped — the short version
Opus 4.7 lands at the same $5 / $25 per MTok pricing as Opus 4.6 — no inflation, which is the rare release where "upgrade now" is a near-zero-risk call. The improvements that matter for agent work are stronger software-engineering benchmarks (relevant if your agents write or modify code) and higher-resolution image processing (relevant for OCR-heavy workflows like invoice parsing or contract review).
Claude Managed Agents is the headline. It's a fully managed agent harness: sandboxed execution, built-in tools, server-sent event streaming. You create agents, configure containers, and run sessions through the API, gated behind the `managed-agents-2026-04-01` beta header. Memory for these agents shipped under the same header.
The ant CLI is the quiet feature with the biggest workflow payoff — a command-line client for the Claude API with native Claude Code integration and YAML versioning of API resources.
Bedrock GA expansion: Opus 4.7 and Haiku 4.5 are now self-serve in 27 AWS regions. Useful if you're stuck with VPC-only deploys.
Retirements: Claude Haiku 3 (`claude-3-haiku-20240307`) is dead — requests error. The 1M context beta retired for Sonnet 4.5 and Sonnet 4. If you need 1M context, migrate to Sonnet 4.6 or Opus 4.6.
Why Managed Agents change the calculus (and why you might wait)
The Managed Agents harness is Anthropic running the agent loop for you — execution sandbox, tool dispatch, SSE event stream. For greenfield agents where you don't yet have an opinion on tooling, this is a real time-saver. We piloted it on a client triage agent last week, and the harness handled retries, partial tool failures, and streaming responses without us writing the boilerplate.
But the trade-off is real, and it cuts against teams who've already invested. If you've spent six months tuning a custom orchestrator with retry budgets, tool gating, observability, and bespoke vector search ranking — Managed Agents asks you to give that up for a managed alternative you don't fully control. Cost optimization gets harder when you can't see every token boundary. Custom escape hatches get harder when the harness owns the loop.
Our take: pilot on a low-stakes agent first. Don't migrate production. The harness is good, but the lock-in surface is non-trivial — once your tools, memory schema, and session state live in Anthropic's infrastructure, swapping back to a custom loop is a real project.
The 1M context retirement — what to do this week
This is the urgent action item. If any of your pipelines pass the `context-1m-2025-08-07` header to Sonnet 4.5 or Sonnet 4, they've been hard-failing since April 30. Options are simple: migrate to Sonnet 4.6 or Opus 4.6 (still on 1M), or refactor to chunked retrieval.
Honest take: most agents pretending to need 1M context were actually masking poor retrieval. If the migration cost feels low — meaning you can swap to chunked + Pinecone or pgvector without breaking the workflow — take it as a signal to rebuild the retrieval layer instead of just bumping the model. A well-tuned retrieval pipeline beats a 1M-context dump on cost, latency, and answer quality, in that order.
ant CLI + YAML versioning: the quiet workflow win
The new CLI is the feature we'll be using daily. YAML-versioned API resources mean agent prompts, tool definitions, and system prompts go into git as reviewable, rollbackable artifacts. PR diffs replace screenshots in Slack. Prompt drift between staging and production becomes diffable instead of mysterious.
If you've ever lost a Tuesday hunting down which prompt version is actually running in prod — this is the fix. Pilot it on one agent's config first; it will change how you think about prompt change management.
What to do this week
The clean priority list:
- 1.Today: audit any `context-1m-2025-08-07` header usage. Migrate to Sonnet 4.6 or refactor to retrieval. Do not ship Monday with hard-failing agents.
- 2.This sprint: upgrade to Opus 4.7 wherever you're on 4.6. Same price, free wins.
- 3.This sprint: pilot ant CLI + YAML on one agent's prompt config. Get used to git-diffing prompts.
- 4.This quarter: POC Managed Agents on a greenfield agent. Don't migrate production yet.
At Veya Studio we ship AI Brains for independent experts and boutique professional service firms — law, accounting, consulting, retail strategy. Production loops run on our Trigger.dev + Hono orchestrator, with Pinecone for memory and Supabase for state, after we migrated off n8n earlier this year. If you're scoping where Managed Agents should sit inside your own stack — or stuck on the 1M migration with a hard deadline — that's exactly the architecture call our Brain Discovery is built for.
Related topics we'll cover next:
- 1.Migrating from custom agent loops to Claude Managed Agents: a practitioner's playbook
- 2.Replacing 1M context with proper retrieval: when chunking beats the bigger window
- 3.Versioning agent prompts in git: YAML, ant CLI, and the end of "prompt drift"