Claude Opus 4.8, released by Anthropic on May 28, 2026, centers on three changes. First, Effort Control lets you dial reasoning depth up or down in discrete steps. Second, Dynamic Workflows can run hundreds of parallel subagents in a single session to handle large-scale code work. Third, Fast Mode delivers roughly 2.5x faster output while costing less than the previous model. In short, the depth, scale, and speed of a task are now directly in the user’s hands.
Until now, AI coding models have been closer to “brilliant but hard to control” tools. Given the same prompt, some days they dig in too deep, and other days they answer too shallowly. Claude Opus 4.8 is designed to hand that control back to the user. Here is what changed, why it changed, and how it works, checked against primary sources.
Effort Control — Set Reasoning Depth Yourself
Effort Control lets you specify how deeply the model should think before it answers. According to Anthropic’s official announcement, at higher settings “Claude thinks more often and more deeply,” while at lower settings it “responds faster.” And on Opus 4.8, the default is high.
In terms of the API parameters given in the official docs, three levels are named: high (default), xhigh (extra), and max. Some outlets (MindStudio) describe this with more granular tiers such as Low, Medium, High, Max, and Ultra Code, but those tier names and token ranges (for example, Low at roughly 0–1,000 tokens, Max at roughly 10,000–20,000+ tokens) come from a media interpretation, so it is safer to take them as “effort handled in steps” rather than as absolute figures.
The mechanism is simple. The API’s thinking-related parameter sets an upper bound on the token budget spent on internal reasoning, and the larger that budget, the longer and more thorough the chain of reasoning. The trade-off is that cost and latency rise along with it, because tokens spent on thinking are billed just like output tokens.
While running NABERAL, I tried splitting this by task: low effort for light work like reviewing blog drafts or simple classification, and high effort for code refactoring or multi-step decisions. The biggest change I felt was not output quality but “cost predictability.” Setting the reasoning budget to match the weight of the task pulled my previously erratic token consumption into a controllable range.
Dynamic Workflows — Hundreds of Subagents in One Session
Dynamic Workflows is a feature for handing larger jobs to Claude Code. Anthropic explains that the model can “plan out a task and then run hundreds of parallel subagents in a single session.” It goes beyond simple auto-responses; it is orchestration in which the model breaks the work apart and runs the pieces concurrently on its own.
On the scale limits, media reporting (MarkTechPost) says a single run is capped at 1,000 agents, with up to 16 running concurrently. These specific numbers are not directly confirmed in the body of the official announcement, so I cite them as media reporting. That said, Anthropic’s official page does state it can carry out “a codebase migration on the scale of hundreds of thousands of lines, from kickoff to merge,” so the overall direction toward large-scale work matches the primary source.
For a real-world example, media reports cite the Rust rewrite of the Bun runtime, where “99.8% of the existing test suite passed” and “about 750,000 lines of Rust code” came together in “11 days from first commit to merge.” Reporting also notes that using Dynamic Workflows requires Claude Code v2.1.154 or later.
What this means in practice is clear: the model has started to take on some of the task-splitting and coordination that a person used to do by hand, in the “this goes to A, that goes to B” style. Of course, the 16-concurrent limit also means this is not infinite parallelism. Rather than throwing a big migration over the wall all at once, the operational judgment of breaking work into verifiable units and handing it off in stages still falls to the human.
Fast Mode — Speed and Cost at the Same Time
Fast Mode is an option that raises output speed while keeping quality the same. Anthropic states that Fast Mode runs at “2.5x the speed” and is “3x cheaper than the previous model.” Its pricing is $10 per million input tokens and $50 per million output tokens. Standard pricing is $5 per million input tokens and $25 per million output tokens.
There is an easy point of confusion here. Fast Mode’s per-token price is higher than Standard mode’s, but it is cheaper compared with the Fast Mode of the previous-generation model. In other words, “the premium you pay when you need a faster response” has come down across generations.
For interactive work, or any situation where a user is waiting in front of the screen, speed itself is productivity. Conversely, for something like an overnight batch job where response time matters less, Standard mode is more sensible. Swapping modes to match the nature of the task becomes the key to cost efficiency.
Performance Metrics and Reliability
On benchmarks, Anthropic’s official page cites 84% on Online-Mind2Web, a web-agent evaluation. A SWE-bench Verified score for the coding evaluation is also included in a comparison table, but the specific figure is not spelled out in the body text of the official announcement, so I will not state it as fact here. Direct numerical comparisons with competing models likewise use different measurement conditions from source to source, so it is best to avoid declaring a clear winner.
The New Stack’s coverage headline summarizes the release as effort control, dynamic workflows, a cheaper fast mode, and “better honesty, less deception.” Specific figures behind the honesty and reduced-deception claims are hard to confirm in the text, so it is safer to take them only as a direction of travel.
The Takeaway — Control Has Shifted to the User
Claude Opus 4.8’s three features ultimately tie into a single theme: the user directly controls the depth (Effort Control), scale (Dynamic Workflows), and speed (Fast Mode) of a task. We have moved from an era of hoping the model would just handle things well to one where the user tunes the tool to fit the nature of the work.
But that control also hands over responsibility. Get the effort level wrong and cost leaks; throw a workflow at the model without verification and you will have to unwind large-scale results one by one. In the end, good outcomes depend not only on model performance but on how you design and split the work. The more powerful the tool, the more the human’s planning of where and how to use it matters.
The figures and feature descriptions in this article were cross-checked against Anthropic’s official announcement (anthropic.com/news/claude-opus-4-8) and reporting from MarkTechPost, MindStudio, and The New Stack, as of June 4, 2026. Some figures are noted in the body as coming from media reporting.
Sources: Anthropic — Claude Opus 4.8, MarkTechPost, MindStudio, The New Stack