Anthropic released Claude Opus 4.8 on May 28, 2026 — same price as Opus 4.7 ($5/$25 per million tokens), no breaking API changes, but three meaningful upgrades for builders.

Anthropic released Claude Opus 4.8 on May 28, 2026 — same price as Opus 4.7 ($5/$25 per million tokens), no breaking API changes, but three meaningful upgrades for builders. The model is now 4× less likely to ship code with unflagged bugs. Dynamic Workflows in Claude Code lets you run hundreds of parallel subagents on massive tasks — one developer used it to migrate 750,000 lines of code in 11 days. And a new effort control system gives every user direct control over how hard Claude works, which doubles as a cost management tool. If you're running agentic workflows or any serious coding work on Anthropic's platform, this is a no-hesitation upgrade. API model string: claude-opus-4-8.

What's New in Claude Opus 4.8

1. Honesty Is Now an Engineering Feature

The most important change in Opus 4.8 is behavioral, not a benchmark number.

Anthropic trained this model to flag its own uncertainty instead of pushing forward confidently when the evidence is thin. The result: Opus 4.8 is approximately four times less likely than Opus 4.7 to let flaws in code it has written pass without remarking on them. It's also the first Claude model to score 0% on "uncritically reporting flawed results."

If you've spent time debugging agentic workflows, in LangGraph pipelines where one bad tool call cascades silently through five downstream steps — you understand why this matters. The failure mode isn't always dramatic. It's usually an agent that generates plausible-looking output, reports success, and leaves you to discover the problem in production.

An agent that says "I'm not confident about this step" is categorically more useful than one that says "done" and moves on. For MERN developers and API builders running Claude in automated loops, this reliability improvement alone is reason to upgrade.

2. Benchmark Gains — What They Mean in Practice

The numbers are real, and reading them the right way matters:

SWE-bench Pro is the one to anchor on. It uses actively maintained repositories with real multi-file diffs — no public ground-truth leakage, no memorization advantage. A nearly 5-point jump there, at the same price, is a genuine coding improvement. Opus 4.8 beats GPT-5.5 on this benchmark by over 10 points.

One honest caveat: GPT-5.5 still leads on Terminal-Bench 2.1 (78.2% vs 74.6%). For pure CLI workflow automation, the comparison isn't a clean win for Anthropic. Keep that in mind if your pipeline is heavily terminal-based.

GPQA Diamond — graduate-level science reasoning — is basically a three-way tie. Every frontier model has saturated it. Don't use it to make decisions.

3. Fast Mode Is Now 3× Cheaper

Fast mode runs at $10/M input and $50/M output — that sounds steep until you realize that's three times cheaper than what Opus 4.7's fast mode cost, while running at 2.5× the speed of standard mode.

For API builders running high-volume, lower-complexity tasks — batch processing, async document pipelines, intermediate agent steps that don't need deep reasoning — fast mode just became a viable architecture choice rather than a luxury. I'm already rethinking a couple of automation flows where I was using Sonnet purely for cost reasons and getting subpar output quality.

4. Alignment: Better Than 4.7, Close to Mythos

Anthropic's internal evaluation places Opus 4.8 between Opus 4.7 and their Mythos Preview model on the alignment ladder. Rates of misaligned behavior — deception, cooperation with misuse — are substantially lower than Opus 4.7 and approach what they're seeing in Mythos, their restricted frontier model.

One thing to flag clearly for agentic builders: the system card shows that prompt-injection robustness in agent pipelines is slightly weaker than Opus 4.7 — a ~9.6% attack success rate versus 6.0%. If your agent processes untrusted external content (scraped data, user-submitted inputs, third-party API responses), review your safety configurations before you deploy Opus 4.8 in that loop. This isn't a reason to avoid the upgrade — it's a reason to be deliberate.

Dynamic Workflows: The Feature That Changes What You Can Actually Build

Most launch coverage mentions Dynamic Workflows in a paragraph. It deserves a section, because it's not just a feature — it's a new class of task you can give Claude Code.

Here's the plain-language version: Dynamic Workflows lets Claude Code plan a large problem, spin up tens to hundreds of parallel subagents to attack it from independent angles, have those agents verify each other's work, and return a unified result — all in one session. If the workflow is interrupted, it resumes from where it stopped. You don't restart from zero.

The proof point is the Bun rewrite. Jarred Sumner — the developer behind Bun, the JavaScript runtime competing with Node.js — used Dynamic Workflows to port Bun from Zig to Rust. The scope: roughly 750,000 lines of code. The timeline: 11 days. The outcome: 99.8% of the existing test suite stayed green. Hundreds of agents worked in parallel, generating code, cross-verifying outputs, flagging inconsistencies, with human review reserved for final sign-off decisions.

That's the kind of project that normally needs a sprint plan, a team of engineers, and a quarter of runway. It finished in under two weeks.

What This Unlocks for Agencies, SaaS Teams, and Solo Builders

You don't need to be rewriting a JavaScript runtime to use this effectively. Here's how Dynamic Workflows maps to real work:

Legacy codebase migrations:-

Upgrading a MERN app from an older Express pattern, migrating a monolith to a microservices structure, updating a dependency across hundreds of files — these used to mean a project plan, sprint allocation, and human reviewers at every step. Dynamic Workflows makes it a Claude Code session with built-in verification.

Parallel bug hunts:- Instead of one agent scanning your codebase sequentially, Claude deploys agents across the entire service at once, then runs independent verification on every finding before reporting back. You get a list of real issues, not false positives.

Multi-file architectural refactors:- Structural changes that touch dozens of interdependent files, with consistency enforced across every single change before anything comes back to you.

Automated test generation at scale:- Writing test suites for an existing codebase is tedious and slow one-file-at-a-time. With parallel subagents working across the whole codebase simultaneously, this becomes a session, not a week.

API integration audits:- For automation builders managing complex API layers — something I deal with regularly — having agents scan every endpoint, check for deprecated calls, verify authentication patterns, and cross-reference documentation in parallel is now feasible.

Before You Run It: What You Need to Know

Dynamic Workflows consumes significantly more tokens than a standard session. Anthropic recommends starting with a scoped task — pick something bounded enough that you can see how the workflow decomposes and how many subagents it spawns. That gives you a calibration point before you throw your full codebase at it.

Plan requirements: Max, Team, and Enterprise only. Not available on Pro. For Enterprise, admins need to explicitly enable it — it ships disabled by default. For Max and Team, it's on by default. Access is through Claude Code CLI, Desktop, and VS Code extension.

Effort Control: The Feature You'll Use Every Single Day

Dynamic Workflows is the headline. Effort control is the thing that actually changes your daily workflow.

Available on all claude.ai plans — including Free — the new effort control sits alongside the model selector. You choose how hard Claude works on any given task:

• Low effort — Faster responses, lower token burn, slower rate limit drain. Use this for quick lookups, simple code edits, high-throughput tasks where depth isn't the priority.

• High effort (default) — Anthropic's recommended balance. Similar token usage to Opus 4.7 on most coding tasks, with better output quality.

• Extra / Max effort — Claude thinks more deeply and revisits its reasoning more frequently. Noticeably better on genuinely hard problems. Costs more tokens. Best for complex debugging sessions, hard architectural decisions, or long-running async workflows where quality matters more than speed.

Who Should Upgrade — and When

Upgrade now if:

You're running agentic loops — LangGraph pipelines, Claude Code sessions, or any workflow where Claude operates autonomously across multiple steps. The honesty and code reliability improvements reduce one of the most frustrating failure modes of agentic systems. Same price, same API surface, zero breaking changes. Just swap claude-opus-4-7 to claude-opus-4-8.

You're on Max, Team, or Enterprise and want Dynamic Workflows. It's live now in research preview.

You have fast mode workflows running on Opus 4.7. The 3× cost reduction is significant at any meaningful volume — this is worth recalculating immediately.

Test before switching if:

You have a heavily optimized production pipeline built around Opus 4.7's specific behavior. The API migration is just a model string change, but effort level defaults have shifted and token consumption patterns may differ from what you've tuned for. Run your evaluation suite before switching over.

You're running agents on untrusted external content. Double-check your safety layer against the prompt-injection change first.

Consider alternatives if:

You're cost-constrained and your workload doesn't require frontier reasoning capability. Claude Sonnet 4.6 at $3/$15 per million tokens is the right call for most standard tasks. DeepSeek V4 is the value play for teams where cost-per-token is the dominant constraint — about 22× cheaper than Opus 4.8, with meaningful but not catastrophic quality tradeoffs on hardest tasks.

You're on a Pro plan and want Dynamic Workflows specifically. That's a Max, Team, or Enterprise feature. Either scope your workflow to standard Claude Code or evaluate whether the plan upgrade makes sense for your volume.

Pricing and Availability

1. API model string: claude-opus-4-8

2. Available across: Claude API, Amazon Bedrock, Google Cloud Vertex AI, Microsoft Foundry, claude.ai, Claude Code (CLI, Desktop, VS Code).

3. Dynamic Workflows: Max, Team, Enterprise — research preview. Admin opt-in for Enterprise.

4. Effort control: All plans including Free. API: xhigh and max via Claude Code settings.

5. One API change worth noting for developers: The Messages API now accepts system entries inside the messages array. You can update Claude's instructions mid-task without breaking the prompt cache or routing it through a user turn. In practice, this means you can adjust permissions, token budgets, or environment context on the fly as an agent runs — something I've wanted for multi-step API pipelines where the context legitimately needs to change between steps.

Try Claude Opus 4.8: claude.ai · API docs · Claude Code

What's Coming Next

Anthropic has been explicit: Opus 4.8 is not the ceiling. It sits between Opus 4.7 and their Mythos Preview model on the internal capability ladder, and Anthropic has signaled that Mythos-class models should reach general availability within the coming weeks. Currently, Mythos is restricted to organizations participating in Project Glasswing for cybersecurity work while safety safeguards are finalized.

Opus 4.8 is a strong, reliable increment with the right improvements in the right places. The infrastructure being built around it — Claude Code, Dynamic Workflows, effort control, the new API behaviors — is the foundation of something larger. The ceiling is moving fast.

AIWerse Opinion:-

Opus 4.8 is a clear upgrade recommendation for anyone already on Anthropic's platform launched on on May 28, 2026. Not a revolution — Anthropic knows that, and they said it themselves. But every change lands where production AI systems actually break: code reliability, agentic orchestration, and cost control.

The migration is genuinely low-risk. Same price, same context window, same API surface. For solo developers, agencies, and SaaS founders running real workloads — not demos — the improved reliability of an agent that catches its own errors is worth the five-minute config change.

AIWerse — breakdowns on the tools, models, and platform moves that matter for builders.

Frequently Asked Questions

What is Claude Opus 4.8 and when was it released?

Claude Opus 4.8 is Anthropic's latest flagship AI model, released on May 28, 2026. It is an upgrade to Opus 4.7 with improvements in code reliability, agentic capabilities through Dynamic Workflows, a cheaper and faster fast mode, and a new effort control system for users. It is available on the Claude API, Amazon Bedrock, Google Cloud Vertex AI, Microsoft Foundry, and claude.ai.

What is the price of Claude Opus 4.8?

The pricing is unchanged from Opus 4.7: $5 per million input tokens and $25 per million output tokens for standard mode. Fast mode — which runs at 2.5× the speed — is $10 per million input tokens and $50 per million output tokens. Fast mode is now three times cheaper than what Opus 4.7's fast mode cost.

What are Dynamic Workflows in Claude Code?

Dynamic Workflows is a new feature in Claude Code (currently in research preview) that lets Claude plan a complex task, spin up hundreds of parallel subagents to work on it simultaneously, verify their outputs against each other, and return a unified result. It's designed for large-scale problems: codebase migrations, full-service bug hunts, multi-file refactors, and similar work that used to require a full engineering sprint. Available on Max, Team, and Enterprise plans only.

Is Claude Opus 4.8 better than GPT-5.5?

yes — Opus 4.8 leads GPT-5.5 on SWE-bench Pro by over 10 points (69.2% vs 58.6%), the hardest real-world coding benchmark available. GPT-5.5 still leads on Terminal-Bench 2.1 (78.2% vs 74.6%), so for pure CLI workflow automation, GPT-5.5 remains competitive. For agentic coding, multi-file work, and complex knowledge tasks, Opus 4.8 has the edge. GPT-5.5 is also cheaper at $3/$15 per million tokens, so the decision depends on your specific workload.

How do I access Claude Opus 4.8 via the API?

Use the model string claude-opus-4-8 in your Anthropic API calls. Pricing and context window (1M tokens) are unchanged from Opus 4.7. There are no breaking API changes — it's a drop-in model upgrade. Claude Opus 4.8 is also available on Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry.

What is effort control in Claude Opus 4.8?

Effort control is a new feature on claude.ai that lets you choose how much reasoning and computation Claude applies to any task. Options range from low effort (faster, fewer tokens, slower rate limit drain) to max effort (deeper thinking, better output on hard problems, higher token cost). It is available on all plans including Free. In Claude Code and the API, the settings are xhigh and max.

Should I upgrade from Claude Opus 4.7 to Opus 4.8?

For most developers and teams running agentic workflows or coding tasks, yes — it is the same price, same context window, no breaking changes, and the reliability improvements in code quality are meaningful. The only cases where you should test before switching: heavily tuned production pipelines (run your evals first) and agent systems processing untrusted external content (the prompt-injection robustness is slightly weaker in Opus 4.8 than 4.7). If cost is the primary constraint, Sonnet 4.6 or DeepSeek V4 are better alternatives.