Claude Opus 4.8

Learning Objectives

• Understand what changed from Opus 4.7 to Opus 4.8 — honesty, agentic judgment, and the new Dynamic Workflows coordination layer
• Evaluate when Dynamic Workflows actually helps versus when single-agent Opus is the better fit
• Pick the right pricing mode (standard versus Fast) for your workload

What Is Claude Opus 4.8?

Claude Opus 4.8 is a flagship-tier Anthropic model released on May 28, 2026 — landing forty-one days after Opus 4.7 and coinciding with Anthropic's $65 billion Series H close. It replaced Opus 4.7 as the general flagship and, since the June 9, 2026 launch of Claude Fable 5, now sits just below Fable 5 as the economical default for routine work — and the model Fable 5 falls back to on sensitive cybersecurity and biology requests. Opus 4.8 brings sharper uncertainty handling, improved agentic judgment, and a coordination layer called Dynamic Workflows that scales single-session work to hundreds of parallel subagents.

Opus 4.8 is part of the Claude model family alongside Sonnet 4.6 (the balanced workhorse) and Haiku 4.5 (speed and efficiency). Standard pricing matches Opus 4.7 exactly, so migration is a model-ID swap rather than a cost rebudget.

What Changed from Opus 4.7

Honesty and Uncertainty Flagging

Opus 4.8's headline behavioral improvement is sharper handling of uncertain or low-quality input. Anthropic reports the model is four times less likely to overlook code flaws compared to Opus 4.7, and Bridgewater's evaluation noted that Opus 4.8 proactively raises concerns about analytical inputs and outputs rather than asserting unsupported claims. Alignment metrics on prosocial traits also improved, and misaligned-behavior rates dropped relative to 4.7.

For analyst, legal, and research workflows where confident-but-wrong answers are the worst failure mode, this is the most consequential change in the release.

Dynamic Workflows (Research Preview)

The standout new capability is Dynamic Workflows — a coordination layer that lets a single Opus session orchestrate hundreds of parallel subagents. Anthropic positions this for codebase-scale work: multi-hundred-thousand-line migrations from kickoff to merge, using the existing test suite as the success bar.

In practice, Dynamic Workflows fans the work out across subagents, gathers their results, and synthesizes a coherent final output — without the developer having to script the fan-out by hand. It is available in research preview on Enterprise, Team, and Max plans.

When NOT to reach for Dynamic Workflows: short single-file edits, prototype work where the fan-out overhead dwarfs the task, and any workload where serial reasoning is what produces value (an architecture review, for example).

Agentic Benchmarks

Anthropic published several agentic-benchmark numbers alongside the release. Specific scores are emerging, but the headline framings the company highlighted are:

• Online-Mind2Web (computer-use, browser-agent): 84%
• Legal Agent Benchmark: first model to break 10% on the all-pass standard
• Super-Agent benchmark: only model completing every case end-to-end
• CursorBench: exceeds prior Opus models across every effort level

These are agentic benchmarks more than they are coding benchmarks — they measure the model's ability to drive long-running tasks to completion rather than win a single-shot reasoning evaluation.

Pricing

API Pricing

Standard mode (unchanged from Opus 4.7):

• Input: $5 per million tokens
• Output: $25 per million tokens
• Prompt caching: Up to 90% savings on cached input tokens
• Batch processing: 50% savings on eligible requests

Fast mode (new with Opus 4.8):

• Input: $10 per million tokens
• Output: $50 per million tokens
• Speed: 2.5-times faster than standard mode
• Three times cheaper than Fast mode was for previous Claude models

Fast mode is for latency-sensitive production workloads where response time directly affects user experience or downstream throughput. Standard mode remains the right default for everything else.

Claude.ai Subscription Tiers

The Claude Model Family

Choosing between models:

• Start with Sonnet for nearly all tasks — it handles most professional work well at lower cost
• Escalate to Opus 4.8 when Sonnet's output is materially insufficient — complex architecture decisions, multi-file refactoring, long-horizon agentic tasks, anything where the consequences of confidently-wrong answers are large
• Use Haiku for high-volume production workloads where per-request cost dominates (chatbots, classification, customer service)

Opus 4.8 vs. Competing Frontier Models

Opus 4.8's positioning at launch is agentic durability — it does the long-horizon work where confident-but-wrong answers compound across hours of automated reasoning. GPT-5.5 retains the largest consumer and third-party-tool ecosystem. Gemini 3.1 Pro is the most accessible flagship through the Google Workspace bundle.

Products Built on Opus 4.8

Several Anthropic products are powered by Opus 4.8:

• Claude Code — Anthropic's terminal-based coding agent. Dynamic Workflows lands here first for the codebase-scale migration use case.
• Claude Design — AI-powered visual creation for prototypes, decks, one-pagers, and UI mockups
• Claude Cowork — Desktop agent for autonomous file and document management
• Claude Managed Agents — Cloud-hosted agents at scale through Anthropic's infrastructure

Each product brings Opus 4.8's capabilities to a different surface — terminal (Code), visual design (Design), desktop (Cowork), and cloud deployment (Managed Agents).

Strengths

• Sharper uncertainty handling — proactively flags concerns about inputs and outputs; four times less likely to miss code flaws versus Opus 4.7
• Dynamic Workflows — hundreds of parallel subagents in a single session for codebase-scale work
• Standard pricing unchanged — drop-in replacement for Opus 4.7 at the same per-token cost
• Fast mode — 2.5-times speed at twice the per-token cost; three times cheaper than Fast was on previous Claude models
• Best-in-class agentic benchmarks — leads Online-Mind2Web, Legal Agent Benchmark, and Super-Agent at launch
• 1 million token context — full-codebase and full-document analysis with no per-token price penalty
• Safety-focused design — Constitutional AI training; improved alignment metrics over Opus 4.7

Limitations and Considerations

• Dynamic Workflows is research preview — Enterprise, Team, and Max plans only; expect behavioral changes as the feature stabilizes
• Closed model — API-only; no downloadable weights or self-hosting option
• Fast mode costs twice as much per token — only worth it for latency-sensitive workloads; standard mode is the right default
• Pro plan required on claude.ai — $20/month minimum for Opus access (Free plan gets Sonnet only)
• Conservative refusals persist — safety training occasionally produces overly cautious refusals on legitimate requests
• Smaller ecosystem than GPT — fewer third-party tutorials and integrations compared to OpenAI's ecosystem
• No native image generation — Anthropic does not currently offer image generation capabilities

Key Takeaways

• Claude Opus 4.8 is a flagship-tier Anthropic model — released May 28, 2026 with sharper uncertainty handling, improved agentic judgment, and the new Dynamic Workflows coordination layer; since June 9, 2026 it sits just below Claude Fable 5, the new public flagship, and is the model Fable 5 falls back to on sensitive requests
• Dynamic Workflows scales single-session work to hundreds of parallel subagents for codebase-scale migrations; available in research preview on Enterprise, Team, and Max
• Standard pricing matches Opus 4.7 exactly ($5 input / $25 output per million tokens), so migration is a model-ID swap; Fast mode adds a 2.5-times-speed option at twice the per-token cost
• Use Opus 4.8 for long-horizon agentic work where confident-but-wrong answers compound; Sonnet remains the default for general professional work; Haiku for high-volume production