The April 30 Cliff: Migrating Off Sonnet 4.5's 1M Context Beta

If you are running Claude Sonnet 4.5 or Sonnet 4 in production with the 1M context beta header, you have 17 days before it stops working.

On April 30, 2026, Anthropic retires the context-1m-2025-08-07 beta header. After that date, requests to Sonnet 4.5 and Sonnet 4 that exceed the standard 200k-token context will be rejected. The header will have no effect.

This is the kind of change that hits production systems quietly. The beta has been in place since August 2025, long enough that most teams stopped thinking about it. The retirement announcement landed in the changelog on March 30. If it got lost in the noise, your workload has two weeks to keep running before it breaks.

What Actually Breaks

If your requests to claude-sonnet-4-5 or claude-sonnet-4 include the beta header and send more than 200k tokens, they will fail starting May 1. Common scenarios where this matters:

• Document analysis pipelines processing long PDFs, contracts, or legal filings
• Codebase-level reasoning where you are feeding entire repositories into context
• Multi-document synthesis combining reports, transcripts, and research notes
• Conversational agents with long session histories or large system prompts

If you built any of these on Sonnet 4.5 in the last eight months, you almost certainly opted into the beta for the context headroom. You need to check.

How to Check

Run this against your codebase, logs, or API gateway configuration:

grep -r "context-1m-2025-08-07" .

Then cross-reference with your model field. If you see the header alongside claude-sonnet-4-5 or claude-sonnet-4, you have a migration to execute.

If you are using Bedrock or Vertex, check your inference profile configurations for the same header. Both platforms honor it and will see the same retirement.

The Migration Path

Anthropic's guidance is to move to Sonnet 4.6 or Opus 4.6. Both support 1M context natively at standard pricing, no beta header required.

Sonnet 4.6

Recommended for most workloads that were running on Sonnet 4.5 with the beta header. A drop-in replacement with better quality and no pricing surprise.

• 1M context window, GA
• 64k max output
• $3 / $15 per million input/output tokens
• Supports adaptive thinking for better quality at lower token cost

Opus 4.6

The right choice if quality was marginal on Sonnet or you need complex reasoning over long context.

• 1M context window, GA
• 128k max output
• $5 / $25 per million input/output tokens
• Best for complex reasoning, synthesis, and high-stakes outputs

What to Change in Your Code

Three things:

• Remove the beta header: delete context-1m-2025-08-07 from your request configuration.
• Update the model identifier: change claude-sonnet-4-5 to claude-sonnet-4-6 (or claude-opus-4-6).
• Retest your prompts: the 4.6 generation behaves differently. Adaptive thinking is the new default. Your prompt regression suite should run before production cutover.

If you do not have a prompt regression suite, build one before the cutover. This is a model migration, not a header swap. Quality and latency will shift. You need to know how.

What This Signals

This is a predictable pattern. Anthropic ships a feature as a beta, iterates, then either promotes it to GA in a newer model or retires it. The context-1m-2025-08-07 beta did its job: it proved demand for longer context and gave customers a path. Now the capability is native in 4.6, and the beta is wound down.

Expect more of this. The interleaved-thinking-2025-05-14 header is deprecated on Sonnet 4.6 and will likely see a similar retirement timeline. Manual extended thinking with budget_tokens is deprecated on 4.6 in favor of adaptive thinking. If your infrastructure is tied to beta headers, you are on a treadmill.

The teams that handle these transitions well have two things: monitoring against the changelog as a first-class engineering practice, and a migration playbook that treats model upgrades with the same rigor as database migrations. The teams that get caught flat-footed are the ones who treat the beta is working fine as a stable state.

It is not a stable state.

Getting Ahead of the Next One

At Riptide, we monitor the Anthropic changelog daily and maintain migration playbooks for every active beta feature our clients depend on. If you want a hand with the April 30 migration, or want to stop discovering these changes the hard way, reach out.

The cliff is in 17 days. Start now.