Claude Opus 4.8 — What’s New in Anthropic’s Model: Dynamic Workflows, Effort Control, Pricing

On May 28, 2026, Anthropic released Claude Opus 4.8 — its newest and, by the company’s own account, most capable generally available model, just 41 days after Opus 4.7. It’s a “modest but tangible” improvement on its predecessor, available at the same price. Per CNBC, the new model outperforms GPT-5.5 and Gemini 3.1 Pro on agentic-coding, financial-analysis and computer-use benchmarks.

But the launch is about more than scores. Alongside the model come Dynamic Workflows in Claude Code, effort control in claude.ai and Cowork, and a cheaper fast mode. Below, what actually changes — and who should care.

What Claude Opus 4.8 Is and What Changed

Opus 4.8 builds directly on Opus 4.7 with improvements across benchmarks and, per early testers, sharper “judgement” on agentic tasks. The API model ID is claude-opus-4-8. Core specs match 4.7: a 1M-token context window (default on the Claude API, Amazon Bedrock and Vertex AI; 200k on Microsoft Foundry), up to 128k output tokens, and adaptive thinking (the model decides when to “think” before answering).

Anthropic focused improvements on three behavioral areas: long-horizon agentic coding (better long-context handling and post-compaction recovery), reasoning-effort calibration, and tool triggering (fewer skipped tool calls a task required). It’s a response to Opus 4.7 complaints.

Dynamic Workflows — Hundreds of Subagents in One Session

The headline new feature alongside the model is Dynamic Workflows (research preview) in Claude Code. It lets Claude plan large work and run hundreds of parallel subagents in a single session, then verify outputs before it presents them. With Opus 4.8, the subagents can run even longer.

The flagship example: Claude Code with Opus 4.8 can carry out a codebase-scale migration — across hundreds of thousands of lines — from kickoff to merge, using the existing test suite as its bar. It’s available on Claude Code for Enterprise, Team and Max plans.

Effort Control — From “High” to “Max”

claude.ai and Cowork now have an effort control next to the model selector. On higher settings Claude thinks more often and more deeply (better answers); on lower settings it responds faster and uses rate limits more slowly. The control is available on all plans.

Opus 4.8 defaults to “high” — what Anthropic judges the best balance of quality and experience. For hard tasks and long async workflows, you can pick “extra” (“xhigh” in Claude Code) or “max,” where the model uses more tokens for better results. Claude Code rate limits were raised to accommodate the higher token usage.

Fast Mode — 2.5x Faster and Three Times Cheaper Than Before

Fast mode for Opus 4.8 delivers up to 2.5x higher throughput (output tokens per second) from the same model, at premium pricing. Notably, for Opus 4.8 it’s three times cheaper than it was for previous models. On the Claude API, fast mode is a research preview (speed: "fast").

New API Features for Developers

The biggest change for those building on the Claude API is mid-conversation system messages. Opus 4.8 accepts role: "system" entries inside the messages array (after a user turn). That lets you update instructions in a long session without restating the full system prompt. Crucially, it doesn’t break prompt-cache hits on earlier turns. In practice: cheaper agentic loops where you change permissions, token budgets or environment context on the fly.

Other changes:

• Lower cache minimum — the minimum cacheable prompt drops to 1,024 tokens. Short prompts that couldn’t cache on 4.7 now create entries with no code changes.
• Refusal stop details — on a refusal, the API returns a category, making it easier to route the user appropriately.
• Unchanged from 4.7 — no temperature, top_p or top_k (non-default values return a 400 error), and adaptive thinking only (no manual thinking budgets).

Less Guessing — the Model Fakes Confidence Less Often

The most emphasized improvement is what Anthropic calls “honesty.” AI models tend to jump to conclusions and confidently claim progress on thin evidence. Per Anthropic, Opus 4.8 more often flags uncertainty and is less likely to make unsupported claims. In internal evaluations it is roughly four times less likely than its predecessor to leave a flaw in its own code without flagging it.

Testers echo this: the biggest difference is often Opus 4.8’s “tendency to proactively flag issues with the inputs and outputs of an analysis — something other models routinely missed and left to the user to catch.”

Alignment and Safety — Close to Anthropic’s Best-Aligned Model

Before release, Anthropic’s alignment team ran a full assessment. The conclusion: Opus 4.8 reaches new highs on prosocial traits — supporting user autonomy and acting in the user’s best interest. Rates of misaligned behavior (deception or cooperation with misuse) are substantially lower than Opus 4.7 and similar to the company’s best-aligned model, Claude Mythos Preview. The full assessment and a suite of safety tests are in the Opus 4.8 System Card.

Benchmarks and Comparison — Where Opus 4.8 Does Best

Anthropic positions Opus 4.8 as its strongest generally available model for complex reasoning and high-autonomy work. From launch data and partner evaluations:

• Computer use: per one tester, Opus 4.8 scored 84% on Online-Mind2Web — a meaningful jump over both Opus 4.7 and GPT-5.5.
• Agentic coding, financial analysis, computer use: per CNBC, the model outperforms GPT-5.5 and Gemini 3.1 Pro.
• Tool efficiency: on CursorBench, Opus 4.8 exceeds prior Opus models at every effort level, using fewer steps.
• Unstructured data: in Databricks’ Genie agent, Opus 4.8 reasons over PDFs and diagrams at 61% cheaper token cost than Opus 4.7.
• Terminal-Bench 2.1: scores reported on the public Terminus-2 harness (for reference, GPT-5.5 scores 83.4% on the Codex CLI harness).

Pricing and Availability

Opus 4.8 is available across the main channels from launch day. Standard pricing is unchanged from Opus 4.7:

• Standard mode: $5 per million input tokens, $25 per million output tokens.
• Fast mode: $10 per million input tokens, $50 per million output tokens.

The model runs in claude.ai, Claude Code and Cowork, and developers use claude-opus-4-8 via the Claude API, Amazon Bedrock, Google Vertex AI and Microsoft Foundry.

Who Opus 4.8 Is For and How to Start

If you use claude.ai or Cowork, Opus 4.8 is already the default — just set an effort level appropriate to the task (consider “extra” for hard ones). For engineering teams, the biggest value is Dynamic Workflows in Claude Code (Enterprise, Team, Max) — especially for migrations and codebase-scale work. Developers building agentic loops gain most from mid-conversation system messages and the lower cache minimum, which cut the cost of long sessions.

What’s Next — Mythos-Class Models “in the Coming Weeks”

Anthropic flags two directions. First, models with Opus-like capability at lower cost. Second, a new class of models more intelligent than Opus: under Project Glasswing, a small number of organizations use Claude Mythos Preview today for cybersecurity work. Models of that class need stronger safeguards, but the company says it expects to bring them to all customers “in the coming weeks.”

Źródła / Sources (wspólne dla obu wersji):

• Anthropic — oficjalne: Introducing Claude Opus 4.8 — Anthropic, What’s new in Claude Opus 4.8 — Claude API Docs, Introducing dynamic workflows in Claude Code — Claude Blog
• Prasa: Anthropic releases Opus 4.8 with new ‘dynamic workflow’ tool — TechCrunch, Anthropic tops OpenAI, releases Opus 4.8 — CNBC, Anthropic upgrades Claude with new Opus 4.8 model — 9to5Mac, Anthropic launches Claude Opus 4.8, prepares Mythos-class models — Help Net Security