Claude Opus 4.8 Is Here: The New Ultracode Effort Level, Dynamic Workflows, and a More Honest Model

Opus 4.8: An Upgrade That Chose Honesty Over Hype

Claude Opus 4.8 arrived on May 28, 2026, only 41 days after Opus 4.7 — a sharp acceleration from Anthropic's usual pace. It is live on claude.ai, in the API as claude-opus-4-8, and on the major cloud platforms, with pricing untouched at $5 per million input tokens and $25 per million output. The benchmarks moved in the right direction: agentic coding from 64.3% to 69.2%, multidisciplinary reasoning with tools from 54.7% to 57.9%, and an 84% score on Online-Mind2Web that, by Anthropic's testing, makes it the strongest computer-use and browser-agent model available, ahead of Opus 4.7 and GPT-5.5. Yet the real story is not the climb in capability — it is honesty. Anthropic frames 4.8 as its most candid model so far: testers say it is about four times less likely to let a coding mistake go by without comment, faster to admit when it is unsure, and much less prone to confident claims it cannot support. For real work, that matters more than a few benchmark points: a model that tells you when to double-check it is safer than one that is slightly sharper but quietly overconfident.

The Effort Slider: From Low All the Way to Ultracode

The most practical change for developers is the Effort control in Claude Code: a single slider that runs from Faster on one end to Smarter on the other, with six steps along it — low, medium, high, xhigh, max, and a brand-new top setting called ultracode. You move it with the arrow keys, confirm with Enter, cancel with Esc. The lower you sit on the slider, the faster Claude answers and the more slowly it eats into your usage limits; the higher you climb, the more deeply it reasons and the more it spends. The default sits at high, which is genuinely enough for the bulk of everyday coding. xhigh is for hard, focused problems and long-running asynchronous work. max pours the most tokens into producing the single best answer. And at the very top sits ultracode — the headline addition of this release, and a meaningfully different kind of setting from the ones below it. The smart habit is not to live at the top of the slider. It is to match the level to the job: low and medium for quick edits, high for normal work, and the top tiers only when the task genuinely earns the extra spend.

What Ultracode Really Is: xhigh Plus Dynamic Workflows

Ultracode is not simply one notch above max — it is a different kind of setting, defined as xhigh plus workflows. It fuses two capabilities into a single choice. The first is the highest practical reasoning effort, so Claude deliberates as carefully as it can. The second is Dynamic Workflows, the research-preview feature Anthropic shipped alongside Opus 4.8, in which Claude takes a large task, plans it, then spins up hundreds of parallel subagents within one session to execute the pieces simultaneously, and finally verifies its own outputs before declaring the work done. Put together, this is what allows Claude Code to take on jobs that used to be measured in engineer-weeks — codebase-scale migrations spanning hundreds of thousands of lines of code, carried from the first command all the way to a merge, with the project's existing test suite serving as the objective bar for whether the result is correct. It is the difference between asking an assistant to edit a few files and handing an autonomous team a large, well-defined mission and letting it report back when it is finished and self-checked.

The Bigger Picture — and Mythos on the Horizon

Step back, and two things define this moment. The first is pace: Opus 4.8 landed just 41 days after 4.7, in the middle of a sharpening race — the same stretch saw major moves from OpenAI's Codex and Google's Gemini, and both Anthropic and OpenAI are reportedly steering toward public market debuts later this year. The second is the emphasis. Rather than chasing capability for its own sake, this release leans into trust and control: a more honest model, an effort slider that lets you decide exactly how much to spend, and a top tier in ultracode that bundles deep reasoning with self-verifying parallel workflows. That direction reflects where adoption actually stalls — not on whether the model is clever enough, but on whether you can rely on it and govern its cost. Looking just ahead, Anthropic confirmed that its most powerful tier, the Mythos-class models held back under tight restriction for their advanced cybersecurity capabilities, are expected to reach all customers within the coming weeks, once the final safeguards under Project Glasswing are in place. With Opus 4.8's alignment scores already brushing Mythos Preview levels, the distance between what the public can use and what Anthropic holds internally is set to shrink fast.

Prompt

# CLAUDE OPUS 4.8 + THE NEW EFFORT SLIDER — MAY 28, 2026

# ─── THE MODEL ───
# API string: claude-opus-4-8
# Available: claude.ai, API, major cloud platforms
# Price: unchanged — $5 / M input, $25 / M output
# Released 41 days after Opus 4.7

# Benchmark gains over 4.7:
#   Agentic coding              64.3% → 69.2%
#   Multidisciplinary reasoning  54.7% → 57.9%
#   Computer use (Online-Mind2Web): 84% — beats 4.7 and GPT-5.5
# Headline trait: honesty — ~4x less likely to let a coding error pass

# ─── THE EFFORT SLIDER (Claude Code) ───
# A single dial from Faster ←→ Smarter, six levels:
#
#   low      → fastest, lightest reasoning, slowest limit consumption
#   medium   → quick everyday tasks
#   high     → the default — deep reasoning, balanced spend
#   xhigh    → hard tasks + long-running async workflows
#   max      → spend the most tokens for the best single answer
#   ultracode → THE TOP TIER = xhigh + workflows
#               highest reasoning effort PLUS Dynamic Workflows
#               (parallel subagents + self-verification)
#
# Adjust with ←/→ , Enter to confirm, Esc to cancel.
# Effort controls are also on claude.ai and Cowork (without the slider UI).

# ─── WHAT ULTRACODE ACTUALLY DOES ───
# It fuses two things into one setting:
#   1. xhigh effort  → Claude reasons as deeply as possible
#   2. Dynamic Workflows → Claude plans a big task, spawns HUNDREDS of
#      parallel subagents in one session, then verifies its own output
# Result: codebase-scale migrations across hundreds of thousands of lines,
#         from kickoff to merge, with your existing test suite as the bar.

# ─── HOW TO CHOOSE — QUICK RULE ───
# Simple edits / fast iteration   → low or medium
# Normal daily coding              → high (default)
# Hard, focused problems           → xhigh
# One critical answer, max quality → max
# Huge multi-file jobs / migrations → ultracode

# ─── PRO TIPS ───
# Do NOT default to ultracode — it spends the most and burns limits fastest
# Reserve ultracode for large, parallelizable jobs with a clear test bar
# For everyday work, high is genuinely enough most of the time
# Anthropic raised Claude Code rate limits to fit the higher-effort tiers

# ─── COMING NEXT ───
# Mythos-class models expected for all customers in the coming weeks,
# once Project Glasswing cybersecurity safeguards are complete.