Opus 4.8 migration, the first Opus update with no breaking changes

Opus 4.8 went GA on May 28, 2026, barely six weeks after 4.7. If you still have the 4.7 migration in your head, with the three hard 400 errors because temperature, top_p and budget_tokens got kicked out, then the good news here is short: that does not happen this time. Opus 4.8 is a genuine drop-in. Your 4.7 code runs unchanged, you just swap claude-opus-4-7 for claude-opus-4-8 in the model field.

Still, it is not a migration you wave through blind. The model behaves differently, and two new levers arrive that touch your cost and architecture decisions. This playbook is for you if you have API code or Claude Code as a daily driver and want 4.8 to move in cleanly.

Step 1, the one-line upgrade

Search your codebase for the model string and swap it.

grep -rn "claude-opus-4-7" src/
# switch every match to claude-opus-4-8

The 1-million-context variant is still called claude-opus-4-8[1m], same spelling as before. The model is available immediately on the Claude API, Amazon Bedrock, Google Vertex AI, Microsoft Foundry and GitHub Copilot. No beta header, no new SDK pin, no changed request fields. That is it technically.

Step 2, the pricing stays, but check your own docs

Standard pricing is unchanged from 4.7: 5 dollars per million input tokens, 25 dollars per million output tokens. No surcharge for the new model.

A concrete stumbling block from practice: I had 15 dollars per 75 dollars wrongly written in my internal docs, a transposed number from an early note. If you have pricing numbers hardcoded somewhere (cost dashboards, internal wikis, customer quotes), check them against the official Anthropic pricing page before you calculate on the basis of wrong numbers. The real source is platform.claude.com/docs/en/about-claude/pricing, not your memory.

Step 3, the behaviour changes, even without an API change

That is the actual reason 4.8 is not a trivial migration. The core improvement is honesty and alignment. Anthropic states that 4.8 waves through a code error without flagging it about four times less often than 4.7. The model flags uncertainties about its own work more often and makes unsupported claims less often.

In practice that means: if you have an agent that used to run through confidently, under 4.8 it will more often say "I am uncertain here" or "I did not verify this assumption". That is intended and usually better. But if your pipeline builds on a particular output form (e.g. a parser that expects a clear yes/no), the extra hedging can surprise your downstream code. Run through your most important agent paths once and see whether the more cautious answering behaviour breaks a hard assumption anywhere.

The benchmark numbers for it, if you need them: agentic coding rises from 64.3 to 69.2 percent, multidisciplinary reasoning with tools from 54.7 to 57.9, agentic computer use from 82.8 to 83.4. The coding improvement is the most noticeable jump in the daily driver.

Step 4, effort defaults to high

Opus 4.8 uses the effort level high by default. Anthropic states that on coding tasks this consumes roughly the same number of tokens as 4.7, at better performance. You can crank it up to extra (in Claude Code that is called xhigh) or max, then the model thinks longer and consumes more tokens.

If you come from the 4.7 world where Claude Code partly defaulted to xhigh, then high as the new default is more of a slight cost relief at equal or better quality. For cost-sensitive workflows that is good news. For critical reasoning tasks where you want the last few percent, deliberately set max instead of relying on the default.

Step 5, Fast Mode as a new cost-speed lever

New with 4.8 is Fast Mode. The model runs roughly 2.5 times faster, but costs 10 dollars per 50 dollars per million tokens, so double the standard. Anthropic describes this Fast Mode as three times cheaper than the Fast Mode of earlier models.

The trade-off is clear: double the token price for 2.5 times the speed. That only pays off where latency costs real money or real UX, e.g. an interactive chat flow a user waits on live, or a bulk job with a hard time window. For background cron agents, research pipelines or anything that runs asynchronously, stay on standard. Otherwise you pay double for speed nobody needs.

Step 6, Dynamic Workflows is a research preview, not a production feature

Together with 4.8, Anthropic launched Dynamic Workflows as a research preview, a system to coordinate complex tasks across hundreds of parallel subagents in Claude Code. That sounds like exactly the CEO-worker pattern you might already be building by hand.

Important for the migration: research preview does not mean production-ready. Try it in a toy project before you rebuild your existing subagent architecture on top of it. If you already work with your own orchestrator that runs today, that is no reason to switch immediately. Watch the feature until it is out of preview.

Migration checklist, five points

One, switch the model field from claude-opus-4-7 to claude-opus-4-8, a single grep-and-replace.

Two, check hardcoded pricing numbers against the official pricing page, do not calculate from memory.

Three, run through your most important agent paths once and see whether the more cautious answering behaviour (more hedging, more uncertainty flags) breaks a hard downstream assumption anywhere.

Four, set the effort level deliberately where it counts, instead of relying on the new high default.

Five, run eval tests against 4.8 once anyway. Even without breaking changes, a prompt that was optimized for 4.7 behaviour can react differently under 4.8, precisely because of the honesty change.

Should you migrate

Yes, and quickly. There are no breaking changes, the coding performance is measurably better, and the honesty improvement reduces exactly the class of errors that hurts you most in agent operation, namely the confidently wrong output. Opus 4.7 stays addressable as claude-opus-4-7, but this time there is little reason to stay on it.

What comes next

If you still have the jump from 4.6 to 4.7 ahead of you, do the Opus 4.7 migration first, that one has the real breaking changes. If you want to set up your Claude Code update ritual cleanly, Claude Code update strategy explains which releases are stable. And if Fast Mode plus effort level move your bills, Claude Code cost controls for the daily driver has the workflow to measure it.

Source

• Anthropic Opus 4.8: https://www.anthropic.com/claude/opus
• Claude API Pricing: https://platform.claude.com/docs/en/about-claude/pricing
• What's new in Claude Opus 4.8: https://platform.claude.com/docs/en/about-claude/models/whats-new-claude-4-8
• TechCrunch on the Dynamic Workflows tool: https://techcrunch.com/2026/05/28/anthropic-releases-opus-4-8-with-new-dynamic-workflow-tool