Claude Opus 4 retires today. Its replacement costs a third as much.

What goes dark at 9am Pacific

Anthropic deprecated both original Claude 4 models on April 14, 2026 and scheduled them for retirement two months later. That date is today. These are the May 2025 flagships - the first models to carry the Claude 4 name - and they are the last of that first wave still answering API calls.

Anthropic gives at least 60 days of notice and emails affected accounts, so this is not a surprise. But the retirement is hard. A request to a retired model returns an error, not a redirect to its successor. If a production job still points at the old string after today, it breaks. The weights themselves don't vanish - Anthropic has committed to preserving retired models long term - they just stop answering API calls.

The unusual part: the upgrade is cheaper

Forced migrations usually cost you. When Haiku 3 retired in April, its replacement was four times the price. This one runs the other way. Opus has had three releases since the original - 4.5, 4.7, 4.8 - and across them Anthropic cut the list price from $15/$75 to $5/$25 and left it there. The model you are being pushed onto is the cheapest Opus there has ever been.

Sonnet is the dull case, in the good way. Sonnet 4.6 holds the same $3/$15 it has carried since the 4.x line began, so the swap changes the model ID and nothing on the invoice. Opus is where the money is. The headline rates are two thirds lower, the context window goes from 200K to a full million, and the output ceiling jumps from 32K to 128K tokens.

What it does to a monthly bill

Two illustrative workloads. The Opus row assumes 20M input and 4M output tokens a month - a typical agentic or analysis pipeline where you reserve the flagship for the hard calls. The Sonnet row assumes higher throughput at 50M input and 10M output, the kind of volume Sonnet usually carries.

The Opus team saves $400 a month on this profile without touching their token usage. One thing the 67% glosses over: Opus 4.7 and later switched to a new tokenizer that can turn the same text into 30 to 35% more tokens, so the real-world saving lands a notch under the headline number rather than on top of it. Caching and batch then widen the gap back out - Opus 4.8 cache hits bill at $0.50/1M, and batch runs at $2.50/$12.50, half the standard rate. If your Opus 4 setup never used those discounts because the sticker price already stung, the move is a good moment to turn them on.

Want the figure for your own traffic? Drop your monthly token counts into the cost calculator and compare the old and new model side by side.

The one line that breaks on the Opus path

Because both old and new models are Claude 4-generation, this is far gentler than a cross-generation jump. The request format, authentication, and response shape are unchanged. There is one exception worth catching before it bites: sampling parameters.

On Claude Opus 4.7 and later - which now includes Opus 4.8 - temperature, top_p, and top_k return a 400 error when you set them to anything other than the default. Opus 4 accepted these happily. Code that tunes temperature will fail outright the moment it talks to 4.8.

// 400 on claude-opus-4-8
client.messages.create({
  model: "claude-opus-4-8",
  temperature: 0.7,   // remove
  top_p: 0.9,         // remove
  ...
});

The fix is to delete the parameters and steer behavior through the prompt instead. Note this is Opus-specific: the Sonnet 4 to 4.6 path is not covered by the restriction, so a Sonnet swap really is just the model string.

Bedrock and Vertex run their own clocks

The June 15 date covers Anthropic-operated surfaces: the Claude API, Claude on AWS, and Microsoft Foundry. Amazon Bedrock and Google Vertex AI are partner-operated and publish their own retirement schedules, which usually lag the first-party date by months. If you call Claude through Bedrock or Vertex, check their model tables rather than assuming today is your deadline - but plan the same migration either way, since you will land on the same Claude 4.x successors.

FAQ

Sources

• Anthropic model deprecations (retirement dates and replacements)
• Anthropic API parameter deprecations (temperature, top_p, top_k)
• Anthropic models overview and pricing
• Anthropic migration guide
• Amazon Bedrock - Claude model availability