Claude Sonnet 4.8: Dynamic Workflows and the June 15 Deadline

Anthropic will retire Claude Sonnet 4 and Opus 4 model IDs on June 15, requiring users to migrate to newer versions. Claude Sonnet 4.8 launches June 16-18 with dynamic workflows and effort control, features previously introduced in Opus 4.8, offering parallel subagent orchestration and cost management at a lower price point.

Two things are happening in the next 72 hours. First, claude-sonnet-4-20250514 and claude-opus-4-20250514 go offline tomorrow, June 15 — if your pipelines still reference those IDs, they break at midnight. Second, Claude Sonnet 4.8 drops June 16-18 with dynamic workflows and effort control, the same headline features that shipped with Opus 4.8 two weeks ago. The timing is not a coincidence. Anthropic engineered a clean handoff: patch the old models out, and you land one step from the most significant Sonnet upgrade in months. Tomorrow’s Deadline: What’s Going Offline The deprecated model IDs are claude-sonnet-4-20250514 , its alias claude-sonnet-4-0 , claude-opus-4-20250514 , and its alias claude-opus-4-0 . These were the original Claude 4 launch models from May 2025. They had a good run. They’re done. Migration targets are straightforward: Sonnet users move to claude-sonnet-4-6 , Opus users to claude-opus-4-8 . The places to look are less straightforward. Run this sweep before you do anything else: grep -r "claude-sonnet-4-20250514\|claude-opus-4-20250514\|claude-sonnet-4-0\|claude-opus-4-0" . Then check everywhere that grep does not reach: .env files, Kubernetes secrets, CI/CD pipeline configs, Terraform and Pulumi stacks, and any third-party integrations that might be forwarding a model ID on your behalf. The usual suspects miss infrastructure-as-code and secrets management systems — that’s where midnight surprises come from. What Claude Sonnet 4.8 Brings And Why the Timing Matters The reason to migrate now rather than next week is Sonnet 4.8. It is expected to land June 16-18 with the two features that defined the Opus 4.8 release: dynamic workflows and effort control. If you wait until the retirement deadline breaks something before acting, you will be patching while Sonnet 4.8 is already available — a worse position than handling it today. Dynamic Workflows: Parallel Subagents at the Sonnet Price Point Dynamic workflows let Claude Code orchestrate parallel subagents within a single session. Anthropic describes the flagship use case as “codebase-scale migrations across hundreds of thousands of lines of code from kickoff to merge” — with the test suite as the validation bar. The model plans the work, fans it out to subagents running in parallel, runs adversarial verification passes, and delivers a result the test suite has already validated. That’s the pitch. Here are the caveats worth knowing before you get excited. Dynamic workflows is a research preview, available on Enterprise, Team, and Max plans only. More critically, it runs in the Claude Code UI — not the Messages API. If you were planning to trigger parallel subagent orchestration programmatically from your application, you’re waiting. There’s no API surface for it yet. The token cost is also real: roughly 3x baseline consumption, around 110 million tokens for a complex session versus the typical 35 million. Set max tokens explicitly and profile your workloads before scaling. Even with those constraints, this is a meaningful capability shift. The Sonnet price point — expected to hold at $3/$15 per million tokens — makes parallel agentic sessions accessible at a cost level where Opus 4.8 $5/$25 starts to add up quickly on high-volume workloads. Effort Control: A Direct Cost Lever The output config.effort parameter gives you five tiers of computational intensity: low , medium , high , xhigh , and max . high is the default when you omit the parameter — meaning existing code that does not set it will not change behavior. The interesting tiers are the extremes. low is for high-volume, latency-critical work: bulk classification, log parsing, tasks where you’re running thousands of calls and shaving latency matters more than reasoning depth. xhigh is what Anthropic recommends for long-horizon agentic coding sessions — the kind that run 30-plus minutes and benefit from the model thinking harder before acting. max exists for frontier-tier problems where cost is secondary to answer quality. response = client.messages.create model="claude-opus-4-8", max tokens=4096, messages= {"role": "user", "content": "Refactor this 50,000-line codebase..."} , output config={"effort": "xhigh"}, Breaking Changes: The Ones Migration Guides Skip Two breaking changes from Opus 4.8 will likely apply to Sonnet 4.8 as well, and both are subtle enough to quietly break production code. The first is adaptive thinking. The budget tokens parameter is gone on Opus 4.8 — passing it returns a 400 error. The replacement is thinking={"type": "adaptive"} , which lets the model determine reasoning depth based on effort level and task complexity. Whether budget tokens survives on Sonnet 4.8 is unconfirmed; Sonnet 4.6 still supports it. Test on day one if your pipeline depends on it. The second is sampling parameters. temperature , top p , and top k now return errors on Opus 4.x. If you are steering output style with sampling parameters, switch to prompt engineering. There is no gradual deprecation here — they are gone. One positive API change arrives alongside: you can now insert system entries within the messages array, enabling mid-task instruction updates without disrupting prompt caching, which is useful for long agentic sessions where context evolves. What to Do Right Now Three steps, in order. Run the grep sweep against your full codebase, including config files and infrastructure-as-code. Update every reference to the retired model IDs before June 15 — to claude-sonnet-4-6 or claude-opus-4-8 depending on your use case. Then, when Sonnet 4.8 lands June 16-18, stage it before promoting to production — the adaptive thinking change and potential sampling parameter removal are worth validating even if the retirement migration went cleanly. The official Claude Opus 4.8 announcement https://www.anthropic.com/news/claude-opus-4-8 covers the new features in detail. The models overview page https://platform.claude.com/docs/en/about-claude/models/overview has the current API ID reference table. The deprecation schedule https://platform.claude.com/docs/en/about-claude/model-deprecations is the authoritative source for retirement dates. The breaking changes are subtle enough that they will not always surface in basic integration tests — do not skip the staging step.