Opus 4.8 in Claude Code: What’s New, Pricing, and How to Switch in 2026

Ailoot reference documentation: ailoot.ai/docs/.

What was added alongside Claude Opus 4.8 on May 28, 2026?#

On May 28, Anthropic rolled out Opus 4.8 — 41 days after Opus 4.7. Claude Code v2.1.154 and the blog post “Introducing dynamic workflows in Claude Code” came out the same day. The next day, v2.1.156 fixed a bug with thinking blocks that broke the API on Opus 4.8.

Four changes that affect everyone who codes in Claude Code:

1. The new Opus 4.8 model — the same prices ($5/$25 per million tokens), but higher benchmarks and it misses bugs in its own code 4 times less often.
2. Dynamic Workflows — Claude writes a JS script that can launch up to 1000 subagents over a run, with no more than 16 active concurrently. The trigger is the word workflow in the prompt.
3. xhigh effort — a new reasoning level. The default on Opus 4.8 is now high; for demanding tasks, you can raise it to xhigh.
4. Fast Mode became 3 times cheaper — it was 6× the standard price; now it is 2× for 2.5× the speed.

Anthropic itself speaks about this release cautiously:

In this guide, I’ll show exactly what changed, what it costs in practice, and whether you should switch now or wait.

How is Opus 4.8 different from 4.7?#

What improved in the numbers

Anthropic published benchmarks in its official announcement and in the Claude API Docs. Summary table:

⚠️ These are vendor-published figures — numbers from Anthropic, not an independent measurement.

The main promise: “4× less likely to miss bugs”

Anthropic claims:

That sounds great, but there’s a nuance under the hood: according to Anthropic’s System Card, Claude Opus 4.8 had the lowest rate of incorrect answers of six models on every benchmark because it abstains when it is uncertain.

Translated from System Card language into vibe-coder language: the model did not become 4 times smarter. It became more likely to say “I’m not sure, check this” instead of producing a plausible hallucination. Honesty instead of an IQ increase — that’s how it evolved.

In practice, this means that on the same tasks where Claude used to lie confidently, it will now say, “I’m not sure about this.” That lowers risk but does not increase coverage.

What stayed the same

• Standard price ($5 per million input tokens, $25 per million output tokens) — unchanged.
• Knowledge cutoff — January 2026, as with 4.7.
• Default context — 1 million tokens on the Anthropic API, AWS Bedrock, and GCP Vertex AI. On Microsoft Foundry — 200K.
• Max output — 128K tokens.
• Sampling parameters (temperature, top_p, top_k) — still unsupported; they return a 400 error.

Where Opus 4.8 falls short

Honestly: on Terminal-Bench 2.1, Opus 4.8 scores 74.6% — not bad, but it is not the category leader. If most of your work happens in the terminal, keep this in mind: Codex is often a little stronger on purely terminal tasks. Opus 4.8 is ahead on the other benchmarks, but this difference is real.

The second candid point: on complex agentic pipelines with unverified input (website parsing, RSS reading), protection against prompt injection in Opus 4.8 is not guaranteed. If you are building automation that reads other people’s documents, keep a separate input sanitizer.

How much does Opus 4.8 cost in 2026?#

Pricing tiers

The standard price stayed the same as for 4.7 and 4.6. That is rare — new generations usually cost more. Fast Mode became 3 times cheaper: on 4.6 and 4.7, fast mode was 6× the standard price. On 4.8, it is 2×.

What that means per hour of work

One hour of active coding in Claude Code on Opus 4.8 is about 1–2 million input tokens and 100–300 thousand output tokens. At standard prices, that is $5–10 for input and $2.5–7.5 for output. Rounded, it’s $8–15 per hour on the API alone.

On subscription plans:

• Pro ($20/month) — enough for 1–2 hours of active work a day. A single demanding prompt with a Dynamic Workflow burns through the limit.
• Max 5x ($100/month) — 4–6 hours a day consistently. The sweet spot for most vibe coders.
• Max 20x ($200/month) — a full workday without worrying about the limit. It makes sense if you launch many workflows or work with the Agent SDK.

More details in the guides how not to burn through Claude Code limits and where tokens go.

⚠️ An important detail from Anthropic: a workflow with 1000 agents burns “substantially more tokens than a typical session.” A large run can consume substantial allowance even though no more than 16 agents run at once. Check the available usage and set a run budget before launching; concurrency alone does not determine total cost.

⚠️ Fast Mode is not included in Pro / Max / Team / Enterprise plans. It requires usage credits to be enabled and is billed separately at $10/$50 per million tokens on top of the subscription. If you are on Max and enable Fast Mode without credits, you will get a message saying you need to top up your balance. This is often missed when reading the documentation.

What are Dynamic Workflows in Claude Code?#

How a workflow differs from a subagent and a skill

Previously, everything Claude Code did autonomously used turn-by-turn orchestration. On each turn, Claude decided which tool to call, and intermediate results sat in its context. This works for up to 4–5 subagents; beyond that, the context overflows.

Dynamic Workflow changes the pattern. Claude writes a JavaScript script that orchestrates the agents itself. Intermediate results live in that script’s variables, not in Claude’s context. Only the final result returns to the context. This has three effects:

1. You can launch up to 16 agents in parallel, up to 1000 in total per run.
2. Claude’s context does not overflow — each subagent works in its own sandbox.
3. The script is reusable — after a successful run, press s in /workflows, and it is saved as your slash command.

Anthropic’s workflow docs put it this way:

Runtime limits

Three ways to activate a workflow

1. The word workflow in the prompt. “Run a workflow to audit every API endpoint under src/routes/ for missing auth checks”. Claude sees the trigger, writes a script, and launches it.
2. /effort ultracode — combines xhigh reasoning with automatic workflow orchestration. It applies to all substantive tasks in the session. The docs say:

3. /deep-research [question] — a built-in workflow for research with web search and cross-checking sources.

Example: Bun on 750,000 lines of Rust in 11 days

The Bun team used Dynamic Workflows for an experiment: porting Bun from Zig to Rust. According to the Anthropic blog:

• ~750,000 lines of Rust generated.
• 11 days from the first commit to merge.
• 99.8% of tests in the existing suite passed.
• ~300 rules for translating Zig → Rust.

The technical pattern was three workflows in a row:

1. Workflow #1 maps the correct Rust lifetime for every struct field in the Zig codebase.
2. Workflow #2 writes each .rs file as a behavior-identical port of its .zig counterpart. Hundreds of file tasks distributed across batches, with at most 16 concurrent agents and two reviewers per file.
3. Workflow #3 is a fix loop: runs build + test until everything passes.

The main trap: workflows will burn through your limit on small tasks

Practical takeaway: if a task can be handled by 1–2 simple subagents, do not call a workflow; use subagents manually. Workflows pay off for 50+ files, not 5.

Dynamic Workflow is the next step in context engineering. A workflow can distribute work across up to 1000 agent launches in total, scheduling no more than 16 concurrently. Without shared project context, errors can accumulate across those batches because each agent sees only its own piece.

When should you enable xhigh effort?#

In the Claude Code v2.1.154 release notes, Anthropic explicitly says: Opus 4.8 runs at effort=high by default — this is the new standard after the release. The standard used to be lower. The new xhigh level is available on top of that for the most demanding tasks, where the model spends more thinking tokens on every turn.

When to use xhigh:

• Migrations (new architectural pattern → 50+ files).
• Security audits (security review of 200+ endpoints).
• Planned refactors with invariant checks.
• Complex, long debugging sessions where the model needs to remember the whole path.

When NOT to use it:

• Routine edits in one file (write a test, rename a method).
• Simple prompts with a clear structure.
• Documentation / comments — xhigh costs extra there.

In Claude Code v2.1.154, the effort slider was renamed from “Speed/Intelligence” to “Faster/Smarter.” The default /effort high is now applied under the hood; you do not need to switch it explicitly.

When should you use Fast Mode?#

Prices:

When to use it:

• Interactive tasks where response time matters more than accuracy (live code review on a stream).
• Quick edits and testing hypotheses.
• A demo / pitch to a client.

When NOT to use it:

• Long-horizon agentic work — 2× the price on a 1M context hurts the economics.
• A large Dynamic Workflow — Fast Mode doubles the applicable token rates, and usage accumulates across all launched agents. The 1000 total-launch limit is not a fixed 1000× task-cost multiplier; budget from actual aggregate input and output usage.
• Background sessions where speed is not critical.

How do you switch to Opus 4.8 without losing settings?#

Step 1. Update Claude Code

In the terminal:

claude upgrade

Or reinstall it using Anthropic’s documentation.

⚠️ The target version is v2.1.156 or later. v2.1.154 had a bug: when using Opus 4.8, thinking blocks were modified, which caused API errors. The bug was fixed in v2.1.156.

Check the version:

claude --version

Step 2. Switch the model in /config

In Claude Code:

/model claude-opus-4-8

Or open /config → Model section → select Opus 4.8 from the list. The model ID on the API endpoint is claude-opus-4-8.

Step 3. Enable Dynamic Workflows (Pro plan)

Dynamic Workflows are disabled by default on the Pro plan. Open /config → Workflows section → enable them. They are enabled from the start on Max, Team, and personal Enterprise. On Enterprise with an admin, they may be disabled by the admin; ask them to enable it.

Step 4. Run a test

In Claude Code, ask for a simple task with a clear result to make sure the model responds correctly:

Write a TypeScript function that splits an array into chunks of a fixed length.
Give me tests in vitest.

If the response is appropriate and the phrase “using claude-opus-4-8” appears at the beginning, the switch worked. If there is a 400 error, check your Claude Code version (go back to step 1).

What you do NOT need to migrate

• CLAUDE.md works with any model without changes. More details in the guide how to configure CLAUDE.md.
• Skills are automatically picked up by Opus 4.8. More details in the guide Claude Code Skills.
• MCP servers work identically. The guide to MCP is 7 ready-made combinations.
• Subagents and Plan Mode are unchanged. A summary of the choice between Skills, Subagents, MCP, and Plugins is in the relevant pillar.

Should you switch from Sonnet 4.6 or Haiku 4.5?#

I already covered choosing between Sonnet and Opus in an older guide, but that was about 4.7. The rules shift a little with Opus 4.8.

Decision matrix for 2026:

For most vibe coders, the sweet spot is Sonnet 4.6 as the default model, plus Opus 4.8 for demanding tasks through the /model switcher. That way, you pay less and still handle demanding tasks properly.

What does NOT work in Opus 4.8?#

Three areas where Opus 4.8 falls short according to practical reports:

1. Data-heavy strategy and roadmaps. Practical reports say 4.7 is still preferred for analysis. If your main use case is to build a metrics table and propose a roadmap, try 4.8 first; if it fails, go back to 4.7 using /model claude-opus-4-7.

2. “The last 10%” in existing codebases. The model does greenfield work (a new project from scratch) very well, but struggles when maturing old code — edge cases and the specific quirks of the architecture. This is a typical weakness of reasoning models: they generate well but are poor at “finishing off” something that already exists.

3. Hallucinations remain. “4× less likely to miss bugs” means that the model says “I don’t know” more often, but when context is insufficient, it can still produce a plausible fabrication. The protection is project context: the more complete CLAUDE.md is, the less room there is for hallucinations.

This does not mean “don’t update.” It means: understand the limitations and keep Sonnet 4.6 and Opus 4.7 in your back pocket as fallbacks.

What comes after Opus 4.8?#

In Anthropic’s phrasing, “coming weeks” usually means 4–8 weeks. In other words, a Mythos-class model will most likely come out in June–July 2026. This changes the strategy:

• If your current projects use Sonnet 4.6 or Opus 4.7 — update to 4.8 and try xhigh and workflows on specific tasks. This is a free upgrade in terms of price.
• If you planned a large migration using a workflow — do not rush. A model that can do it even better may come out in a month or two. You can start with an experiment on one module to understand how it works.
• If you were going to rebuild your infrastructure around Dynamic Workflows — wait. Mythos may change the underlying patterns.

That uncertainty is the main advice: stay flexible. Opus 4.8 is a good incremental upgrade. Mythos may be a leap. Do not invest in major rework a month before a possible new model.

Continue exploring Ailoot at ailoot.ai.