Forty-one days. That’s the gap between Opus 4.7 and Opus 4.8, the fastest version of the Opus line has ever run. But the release cadence isn’t the interesting part. What’s interesting is that two specific features in this release, effort controls and dynamic workflows, actually let you change the prompt mid-conversation.
Anthropic AI shipped Claude Opus 4.8 with real benchmark movement on AI coding and math, and two companion features. This article covers the changes and what it means for your AI workflows.
Key Takeaways
- Opus 4.8 was released on May 28, 2026, with no change in price and measurable gains in AI coding, math reasoning, and long-context retrieval over Opus 4.7.
- Effort control gives you a five-level dial over how much the model reasons before answering, with each level trading token cost for output quality.
- Dynamic Workflows enable Agentic AI at scale, letting Anthropic Claude Code spawn up to 1,000 parallel subagents per session for tasks like full codebase migrations.
- Intermediate results in Dynamic Workflows live outside the context window in an orchestration script, which is the technical detail that makes the feature actually different from standard subagents.
- Fast Mode dropped from $30/$150 to $10/$50 per million tokens at roughly 2.5x standard speed, making latency-sensitive AI coding workloads three times cheaper than before.
- Dynamic Workflows is in research preview and only available on Team, Enterprise, and Max plans in Claude Code.
What Is Claude Opus 4.8?
Claude Opus 4.8 is Anthropic AI’s new flagship AI model, which is publicly available. It sits above Sonnet 4.6 and Haiku 4.5 in Anthropic’s lineup, and below the restricted Mythos-class models accessible only through Project Glasswing. The official API model ID is claude-opus-4-8.
Key specs at a glance:
- Context window: 1 million tokens.
- Max output: 128,000 tokens.
- Standard pricing: $5 per million input tokens, $25 per million output tokens (unchanged from Opus 4.7).
- Available on: Claude API, Claude Code, Amazon Bedrock, Google Vertex AI, Microsoft Foundry, Cursor.
According to Anthropic’s migration guide, the API contract is identical to Opus 4.7. You change one string, claude-opus-4-7 to claude-opus-4-8, and you’re on the new model. No new auth flow, no SDK update required.
Two API-level changes are really important to understand. First, system messages can now be passed inside the conversation history after a user turn. This lets agents update their instructions mid-session without invalidating the prompt cache, which matters for long agentic runs where task conditions change mid-flight. Second, the effort enum is now available beyond Claude Code, in the Claude.ai UI and Cowork.
How Does Effort Control Change the Way Opus 4.8 Reasons?
In short, effort control adjusts how many reasoning tokens the model allocates before producing an answer.
On higher settings, Anthropic Claude thinks longer. It breaks the problem into sub-problems, considers alternatives, and checks its own work before responding. On lower settings, it answers faster and cheaper. This is the same underlying concept as the reasoning model controls OpenAI uses in its o-series and Google uses in Gemini Thinking, though Anthropic’s implementation exposes it as a named enum.
Opus 4.8 defaults to “high” effort, which is calibrated to spend a similar number of tokens as Opus 4.7’s default while producing better results. The two levels above the default, extra and max, are where the quality gap becomes noticeable on hard tasks.
Claude effort levels: Low, medium, high, extra, and max
| Effort Level | API Enum | Speed | Token Cost | Best For |
| Low | low | Fastest | Cheapest | Short lookups, quick rewrites, basic chat |
| Medium | medium | Fast | Low | Summaries, simple coding questions, light brainstorming |
| High | high | Balanced | Moderate | Default for everyday tasks, general AI coding, drafting |
| Extra | xhigh | Slower | Higher | Strategy docs, complex debugging, long reports |
| Max | max | Slowest | Highest | High-stakes outputs, autonomous agent runs, legal or financial drafts |
When to use extra or max effort vs. the default
Use extra or max effort when a wrong first step would cost you hours of cleanup, not seconds of re-prompting.
| Task Type | Recommended Effort | Why |
| Short emails, quick summaries | High (default) | Quality gain at higher levels doesn’t justify the token cost |
| Strategy memos, long-form content | Extra (xhigh) | Multi-step reasoning cuts follow-up prompts significantly |
| Architecture planning, code review | Extra (xhigh) | Catching a bad assumption early is worth the added cost |
| Autonomous agent runs | Max | Long sessions where you’re away from the keyboard while it works |
| Legal, financial, or irreversible outputs | Max | Thoroughness matters more than speed when errors compound downstream |
Just create a habit. Every time before typing, ask if Claude gets this wrong, does recovery cost ten seconds or two hours? Match the effort level according to that.
What Are Dynamic Workflows and Why Do They Matter for Agentic AI?
Agentic AI got a genuinely useful new tool with Opus 4.8’s companion launch in Claude Code, Dynamic Workflows. The core mechanism is that Anthropic Claude can now plan a large task, write a JavaScript orchestration script that breaks it into parallel pieces, and execute those pieces across up to 1,000 subagents in a single session.
This is structurally different from standard subagents in Claude Code. They run inside the parent model’s context window. Dynamic Workflows runs through a background runtime, with Claude operating as the orchestrator at a higher level rather than as the executor of each individual task.
Practical use cases include:
- Full codebase migrations across thousands of files.
- Repository-wide bug sweeps validated against a test suite.
- Multi-service refactors where each service runs in parallel.
- Large-scale content audits, document processing, or structured data extraction across hundreds of inputs.
You trigger it by including the word workflow in your prompt to Claude Code. Claude generates an orchestration plan, shows it to you for review, and then executes it in the background.
System memory optimization: Storing variables outside the context window
This is the technical detail that actually makes Dynamic Workflows different from everything that came before it, and it’s easy to skip in the launch coverage.
In standard subagent setups, intermediate results live inside the parent model’s context window. That means the parent’s context limit is the ceiling on how much coordinated parallel work can happen. Once you’re dealing with thousands of files across hundreds of agents, that ceiling becomes the bottleneck.
Dynamic Workflows moves intermediate results into JavaScript variables in the orchestration script, which lives outside Claude’s context window. The background runtime executes the script, stores results externally, and passes only the final consolidated output back to the model.
Here’s the rough architecture split:
| Inside Claude’s Context | Inside the Orchestration Script |
| Final consolidated output | Per-subagent task results |
| High-level task plan | Dependency tracking between sub-tasks |
| Original user prompt | Intermediate validation outputs |
The result: the parent model’s 1 million token context is no longer the ceiling for coordinating parallel work. That’s what makes repository-scale migrations possible in a way standard subagents simply aren’t built for.
How to trigger a workflow in Claude Code
Dynamic Workflows require Claude Code v2.1.154 or later. Here’s how to get started:
- Update Claude Code: npm install -g @anthropic-ai/claude-code@latest
- Verify your version: claude –version
- Include the word “workflow” in your Claude Code prompt. Claude generates an orchestration plan and presents it for your review before executing anything.
- Alternatively, enable “ultracode” mode in Claude Code settings to automatically apply xhigh effort plus workflow orchestration on every prompt.
- Review the orchestration script Anthropic Claude generates. Approve it before the workflow runs.
- Monitor progress in your active session while the workflow executes in the background.
Dynamic Workflows is only available on Claude Code Team, Enterprise, and Max plans. Pro plan users get effort control but not workflow orchestration.
A real-world example: Large-scale code migration
The clearest production proof of what Dynamic Workflows can do came from the creator of the Bun JavaScript runtime, who used the feature to migrate approximately 750,000 lines of code over an 11-day period, using the project’s existing test suite as the correctness bar for each subagent output.
A few things to hold in your head before you get excited:
- The developer had deep familiarity with both the source and target codebases going in.
- Human oversight was the correctness check at every checkpoint. Subagent outputs were reviewed and approved before being accepted.
- Replicating this result requires substantial engineering scaffolding and a robust test suite, not a single well-written prompt.
This isn’t a prompt once and leave it for 11 days thing. It’s a story about what becomes possible when a developer who understands the problem uses Dynamic Workflows as a force multiplier. The token cost of running hundreds of parallel Opus 4.8 subagents is real and compounds fast, which I cover in the limitations section below.
What Changed in the Opus 4.8 Benchmarks?
For AI coding specifically, here’s how Opus 4.8 compares to its predecessor:
| Benchmark | Opus 4.7 | Opus 4.8 | Delta |
| SWE-Bench Verified | 87.6% | 88.6% | +1.0 pt |
| SWE-Bench Pro | 64.3% | 69.2% | +4.9 pts |
| Terminal-Bench 2.1 | 66.1% | 74.6% | +8.5 pts |
| USAMO 2026 (Math) | 69.3% | 96.7% | +27.4 pts |
| GraphWalks F1 (1M context) | 40.3% | 68.1% | +27.8 pts |
For AI coding, I’d focus on the SWE-Bench Pro and Terminal-Bench 2.1 rather than SWE-Bench Verified. Verified is saturating near the top, where single-point movements carry less meaning. SWE-Bench Pro draws from real GitHub issues in production projects, which is much closer to what engineering teams actually face day-to-day.
Where the gains are real vs. where they’re marginal
The math and long-context jumps are large. The AI coding gain on SWE-Bench Verified is incremental. Both are true at the same time, and it’s worth being specific about which is which.
A few things to flag for anyone making decisions based on these numbers:
- The 4x less likely to leave code defects unremarked figure is self-reported by Anthropic’s Alignment team and has not yet been independently verified. Read it as a directional signal, not a third-party benchmark.
- Cursor’s internal evaluation, CursorBench, independently found Opus 4.8 outperforms Opus 4.7 across every effort level. That’s the closest available independent production endorsement.
- The Legal Agent Benchmark result, the first AI model to break 10% on the all-pass standard, matters specifically for teams building in regulated industries doing contract review or compliance work.
How Does Opus 4.8 Compare to GPT-5.5 and Gemini?
| Capability | Opus 4.8 | GPT-5.5 | Gemini 3.1 Pro | Qwen3.7-Max |
| AI Coding (SWE-Bench Verified) | 88.6% | ~82% | Not disclosed | Not disclosed |
| Raw Reasoning (AA Intelligence Index) | ~58 | 60.2 | Competitive | Competitive |
| Long-Context Retrieval (1M tokens) | Strong | Limited | Strong | Moderate |
| Standard Input/Output Pricing | $5/$25 per M | Higher | Moderate | Lower |
| Best Use Case | Engineering depth, long agent runs | Structured reasoning, academic tasks | Long-context analysis, cost-sensitive | High-volume Agentic AI loops |
No single AI model wins across every category, and anyone telling you otherwise is selling something. GPT-5.5 still leads on raw reasoning benchmarks. Gemini 3.1 Pro holds its own on long-context analysis with better pricing for that specific use case. Qwen3.7-Max is the practical cost winner for high-volume Agentic AI tasks.
I’d choose Opus 4.8 for production engineering workflows where reliability over long sessions and AI coding accuracy matter more than per-token unit economics.
What Does Opus 4.8 Cost, and Is It Worth Upgrading?
| Tier | Input | Output | Speed |
| Opus 4.8 Standard | $5/M tokens | $25/M tokens | Standard |
| Opus 4.8 Fast Mode | $10/M tokens | $50/M tokens | ~2.5x faster |
| Opus 4.7 Fast Mode (for comparison) | $30/M tokens | $150/M tokens | ~2.5x faster |
The Fast Mode price drop is the most commercially significant number in this release. Going from $30/$150 to $10/$50 makes latency-sensitive AI coding agents and customer-facing products viable on Opus-class quality where they weren’t before.
- Suppose you’re on Opus 4.7 via the API: Upgrade today. It’s one line change, identical standard pricing, and the SWE-Bench Pro and Terminal-Bench gains are real. There’s no scenario where staying on 4.7 makes sense.
- If you’re a Claude Code user: Update to v2.1.154, switch to Claude Opus 4.8 as your default, and try Dynamic Workflows on one real, well-defined task. Treat it as a research preview with genuine token cost implications.
- If you’re evaluating Opus vs. Sonnet 4.6: Opus costs about 3x more than Sonnet. For general tasks that don’t require sustained multi-step reasoning over many turns, Sonnet 4.6 still comes forward. Opus earns its price differential specifically on long agentic runs and complex reasoning chains.
What Are the Real Limitations of Dynamic Workflows?
Token cost at scale is non-trivial: Running a workflow with 200 subagents on Opus 4.8 at max effort can use roughly 100 times the tokens of a single complex prompt. For one-off migrations that would otherwise take a multi-engineer team months, the economics work. For repeated weekly AI workflows, you need to compare cost-per-final-result before committing.
Research preview means instability: Anthropic Claude explicitly notes that Dynamic Workflows pricing and architecture may change. Building production pipelines that depend on current behavior without accounting for that risk is a mistake.
Plan restrictions are real: Dynamic Workflows is locked to Team, Enterprise, and Max plans. Pro users get effort control but not workflow orchestration.
Subagent drift on vague task definitions: When the unit of work is loosely defined, subagents produce inconsistent quality across parallel outputs.
How Do You Start Using Effort Control and Workflows in Your AI Workflows?
Here’s a practical three-step plan you can act on this week without rewriting anything from scratch.
Step 1: Assess task difficulty before picking an effort level.
Before you type, ask: if Claude gets the direction wrong here, does recovery cost ten seconds or two hours? That single pre-prompt question changes your results more than adding more words to your prompts.
Step 2: Identify which recurring tasks have parallel structure.
Look at your last month of Claude sessions. Are there tasks where you’re processing many similar units? Document reviews, content audits, structured data extraction, outreach personalization? Those are workflow candidates. The test is whether the task can be broken into N copies of the same job applied to N different inputs. If yes, Dynamic Workflows can handle it.
Step 3: Write a workflow-shaped prompt.
For non-coding AI workflows, here’s a template structure to adapt:
“Run a workflow to review each article in [folder] against these five criteria: [list criteria here]. For each article, return a JSON object with the filename, a 0-5 score per criterion, and a one-sentence fix suggestion for any score below 3. Save all results to [output file].”
The key ingredients are a defined unit of work, a finite list of those units, and a testable success criterion per unit. This helps Claude split the job cleanly across subagents without drift.
Final Thoughts
Opus 4.8 is the right AI model for production engineering teams, practitioners running long Agentic AI sessions, and anyone who has hit the ceiling of what Sonnet 4.6 can reliably handle over many turns. It’s not a reason to switch if Sonnet is working for your workloads, and the price difference is real.
One thing worth knowing before you make long-term infrastructure decisions: Anthropic Claude has confirmed that Mythos-class models are coming to all customers in the near term. Once they ship, Opus becomes the mid-range option in Anthropic AI’s public lineup, not the flagship. Plan for that transition when you’re thinking about how your AI workflows are structured.
FAQs
Anthropic AI released Opus 4.8 on May 28, 2026, exactly 41 days after Opus 4.7, making it the fastest Opus version cadence to date.
The five effort levels are low, medium, high (default), extra (xhigh in Claude Code), and max. Higher levels use more reasoning tokens and cost more per request.
No. Dynamic Workflows is a research preview available only on Claude Code Team, Enterprise, and Max plans. Pro users get effort control but not workflow orchestration.
Opus 4.8 improves on Opus 4.7 on AI coding benchmarks, math reasoning, and long-context retrieval at identical standard pricing with no breaking API changes.
No. Anthropic’s Mythos-class models are more capable but restricted to Project Glasswing partners. Opus 4.8 is the strongest publicly available Anthropic AI model right now.
Fast Mode costs $10/$50 per million input/output tokens at roughly 2.5x standard speed, down from $30/$150 on Opus 4.7, three times cheaper than the previous tier.
