Claude Opus 4.8 releases effort control and dynamic workflows for Claude Code

• Anthropic's launch summary says Claude Opus 4.8 keeps Opus pricing flat at $5 per million input tokens and $25 per million output tokens, while adding claude.ai effort control, a cheaper fast mode, and dynamic workflows for Claude Code.
• In Anthropic's launch post, the company frames dynamic workflows as a way to run hundreds of parallel subagents for codebase-scale jobs, and the dedicated workflows post shows that ultracode is the switch that pairs xhigh effort with automatic workflow orchestration.
• According to the HN discussion summary, early users were already testing that stack on long runs: one said Opus 4.8 in ultracode mode built a one-file RTS cleanly, while another said it was the first model to generate a crossword with a good layout.
• The main HN thread also carried the first real caveat: one commenter reported worse performance on trivia and domain-specific knowledge tasks, even as agentic coding and structured generation looked stronger.

You can read Anthropic's launch post, the fuller dynamic workflows explainer, and Anthropic's follow-up on workflow harness design. Simon Willison's notes pull out the honesty angle, while the HN launch thread is where the useful weirdness showed up fast: ultracode, mid-conversation system messages, and people stress-testing layout generation instead of yet another todo app.

Anthropic bundled three separate product changes into this release, and the bundle matters more than the version bump.

• Claude Opus 4.8 replaces Opus 4.7 at the same standard API price, $5 per million input tokens and $25 per million output tokens, per Anthropic's launch post.
• Fast mode now runs at up to 2.5x the speed and costs $10 per million input tokens and $50 per million output tokens, which Anthropic says is three times cheaper than fast mode on previous models, per Anthropic's launch post.
• Effort control is now exposed in claude.ai on all plans, letting users trade speed and rate-limit burn for deeper reasoning, according to Anthropic's launch post.
• Dynamic workflows arrived for Claude Code, where Anthropic first described them as a research preview for Enterprise, Team, and Max in the launch post, then described them as generally available across Claude Code surfaces in the workflows post.
• The model slug is claude-opus-4-8, available via the Claude API, per Anthropic's launch post.

Anthropic also used the release to push an honesty claim. In the same announcement, the company says Opus 4.8 is around four times less likely than Opus 4.7 to leave flaws in its own code unmentioned.

The sharpest new capability is not the base model alone. It is the new harness Anthropic wrapped around it.

In Anthropic's workflows post, the system works like this:

• Claude plans the task from the prompt.
• It breaks the job into subtasks.
• It fans those subtasks out to tens or hundreds of parallel subagents.
• It runs verification before folding results back into one answer.
• It saves progress so interrupted long runs can resume.

The activation path is also more concrete than the launch headline suggests. In the workflows post, Anthropic says users can either ask Claude to create a workflow directly or flip on ultracode, a Claude Code setting that sets effort to xhigh and lets Claude decide when to spin up a workflow.

Anthropic's later harness write-up adds a second useful detail: workflows can choose different models per agent and decide whether subagents run in isolated worktrees. That turns the feature into a task-specific orchestration layer, not just a bigger reasoning knob.

Effort control is the user-facing change creative and product teams will feel first, because it turns model behavior into a visible interface setting.

In Anthropic's launch post, the company says higher effort makes Claude think more often and more deeply, while lower effort responds faster and burns rate limits more slowly. Opus 4.8 defaults to high effort. Anthropic recommends extra, called xhigh in Claude Code, for harder tasks and longer asynchronous runs.

That matters because the same model can now present very different personalities depending on the setting. One HN commenter in the main thread complained that six effort levels, including ultracode, make comparisons messier because people can report opposite experiences while technically using the same model.

Anthropic also says it raised Claude Code rate limits to absorb the extra token spend from higher effort modes. Flat list pricing, in other words, does not mean flat usage.

The first outside reports were less about benchmark tables and more about whether Opus 4.8 can hold a shape across longer runs.

According to the HN discussion summary, one early tester used ultracode mode to build a simple RTS game in one file and said it nailed the task. In the same thread, another user said a crossword-layout smoke test, their personal check for structured creative generation, was the first time a new model produced a good layout.

Those reports line up with the way Anthropic positioned the release. In the launch post, the company emphasizes long-running agentic work, coding reliability, and a better ability to flag uncertainty instead of bluffing through it.

A broader Ask HN thread about "oh shit" GenAI moments added more production-style examples around the same week. That discussion summary includes log-file analysis done in seconds, scientific equation implementation from a reference manual, and step-by-step reverse engineering of an old USB audio driver. The throughline is less "look what it wrote" and more "look what it stayed on top of."

The early consensus was not cleanly positive.

One top-level HN commenter in the main thread said Opus 4.8 did a bit worse on their own tests and seemed weaker on trivia or domain-specific knowledge tasks. That fits a recognizable split in this launch: Anthropic is selling a model tuned for agentic execution, while some users are still probing it like a general-intelligence quiz machine.

The official framing is more restrained than the surrounding excitement. Anthropic calls Opus 4.8 a "modest but tangible improvement" in the launch post, and Simon Willison's write-up highlights the same line. For a model that shipped with a new harness, new effort controls, and cheaper fast mode, that understatement is probably the honest read.

The buried feature is an API change that arrived in the same release and quietly explains why the new harness can keep adapting mid-run.

In Anthropic's launch post, the Messages API now accepts system entries inside the messages array, not just at the top. Anthropic says that lets developers update Claude's instructions mid-task without breaking the prompt cache or routing the update through a user turn.

That is the detail the HN summary picked up through Simon Willison's comment: mid-conversation system messages let a long-running session append fresh instructions later, instead of restating the whole system prompt. Anthropic says the mechanism can update permissions, token budgets, or environment context while an agent is already running, which makes it a plumbing change with visible product consequences.