Claude Opus 4.8 — Dynamic Workflows, Fast Mode, and the Mythos Preview (2026)

What Changed in Claude Opus 4.8

Claude Opus 4.8 is Anthropic's latest flagship model, released on May 28, 2026. The release landed the same day Anthropic announced a $65 billion raise at a $965 billion valuation, cementing its position as the world's most valuable AI startup — and the model update follows the same trajectory.

The core direction of this update isn't a step change in raw model weights. It's "finish bigger work in one session, end to end." Anthropic frames Claude Opus 4.8 as having "sharper judgment, more honesty about its progress, and the ability to work independently for longer than its predecessors."

This guide walks through the four big shifts Claude Opus 4.8 brings — Dynamic Workflows · Fast Mode · benchmark gains · honesty improvements — and the Mythos model Anthropic previewed alongside it.

Dynamic Workflows — Up to 1,000 Subagents in a Single Session

The most notable new capability is Dynamic Workflows. In Claude Code, a single session can now spawn up to 1,000 subagents in parallel, currently as a research preview.

The previous Claude Code execution model ran one main agent through a sequential plan. Complex work meant the human split the task into steps or opened multiple sessions. Dynamic Workflows pushes that decomposition into Claude itself.

Simplified flow:

1. A large task comes in; Claude plans it out first
2. When parallelizable parts appear, subagents are spawned simultaneously
3. Each subagent finishes its slice and reports results back
4. The main session integrates the results and moves to the next step

1,000 is an upper bound, not a typical run size. The meaningful implication is that work like "analyze the whole monorepo at once," "refactor hundreds of files in parallel" can complete inside a single session. A large fraction of tasks that previously required the user to split the work by hand now happen automatically.

Fast Mode — 2.5× Faster, 3× Cheaper Than Previous Fast Mode

The second new feature is Fast Mode. It runs Claude Opus 4.8 at 2.5× the standard speed, at 2× the per-token price. It's the option you reach for when you need the same result in less wall-clock time.

Pricing layout:

Mode	Speed	Per-token price
Standard	1×	$5 / $25 per 1M tokens (input / output)
Fast Mode	2.5×	2× standard pricing

On the surface 2× pricing looks steep, but compared to the previous generation's fast mode, this Fast Mode is roughly 3× cheaper. The scenario where "I want to use fast mode but the price doesn't justify it" is meaningfully smaller now.

Cases where Fast Mode pays off:

• CI workflows that need code review to finish quickly
• User-facing chatbots that need responsive turn time
• Tight-deadline work where shortening plan→execute cycles matters more than cost

For background batch work where token efficiency outweighs latency, standard mode remains the rational choice.

Benchmarks — SWE-bench, Terminal-Bench, and Honesty

Direct comparison between Opus 4.7 and 4.8:

Metric	Opus 4.7	Opus 4.8	Change
SWE-bench Pro	64.3%	69.2%	+4.9pp
SWE-bench Verified	87.6%	88.6%	+1.0pp
Terminal-Bench 2.1	66.1%	74.6%	+8.5pp
Multidisciplinary reasoning (with tools)	54.7%	57.9%	+3.2pp
Agentic computer use	82.8%	83.4%	+0.6pp
Knowledge work	1753	1890	+137

The biggest gain is Terminal-Bench 2.1 (+8.5pp) — the model's ability to actually carry out work in a terminal, which is precisely the core competence Claude Code depends on.

SWE-bench Verified moving from 87.6 to 88.6 looks small at first glance, but that benchmark has been saturating in the 88% region, so the gain is meaningful. Equally important: Opus 4.8 is roughly four times less likely than Opus 4.7 to let flawed code pass unremarked — a measurable honesty gain. The accuracy rate edges up, but false confidence drops sharply.

Pricing and Context Window — Same Price as Opus 4.7, 1M Input

Performance went up while pricing held steady against Opus 4.7.

Item	Value
Input price	$5 / 1M tokens
Output price	$25 / 1M tokens
Context window	1M input / 128K output
Fast Mode	2× pricing, 2.5× speed

1M input context is the headline change for Claude Code users. Large portions of a monorepo can sit in one session, putting Claude Opus 4.8 in the same tier as Grok Build's 256K and GPT-5.5's 1M, as covered in the recent CLI comparison guide.

The same release added mid-task system messages to the Messages API. System-level instructions can now be injected partway through a session, useful for steering long-running work without losing context. Combined with Dynamic Workflows, this lets a main session adjust the direction of subagents already in flight.

GitHub Copilot picked up Claude Opus 4.8 as generally available on release day, so Copilot users can call it without separate setup.

Context at 1M — Does It Close the Gap with Codex?

One pattern worth naming explicitly: for a long stretch, "Claude has the better workflow, but Codex is the only realistic choice for large contexts" was a real trade-off. Monorepos, long PDFs, large documents — Codex's (GPT-5.4/5.5) 1M context was effectively the only path. Opus 4.7's 200K just couldn't hold them.

Opus 4.8's 1M input closes that gap. Comparing the two on context terms:

Metric	Codex (GPT-5.4/5.5)	Claude Opus 4.7	Claude Opus 4.8
Input context	1M	200K	1M ✅
Output context	Generally large	(model-dependent)	128K
Multimodal	Omnimodal (text, image, audio, video)	Text-centric	Text-centric
Built-in code review agent	Yes	No	No

On input: Codex no longer holds a meaningful edge. If the workload fits in 1M, moving it to Opus 4.8 doesn't cost anything context-wise.

On output: Claude Opus 4.8 caps at 128K output tokens. For workloads that need very long single-shot outputs (massive documents or code generated in one pass), Codex still has the wider lane. In practice that ceiling is rarely binding, and the standard workaround — chunking the output — closes the gap.

Where Codex remains the more natural fit:

• When images, PDFs, audio, or video are part of the coding context — Codex's omnimodal backbone
• When automated diff critique is the workflow's core — Codex's built-in review agent
• Polyglot projects — Codex stays slightly steadier across languages

To sum up, the meaningful read on this update is that the era when context window alone decided between Codex and Claude Code is over. Going forward, tool selection sorts itself out around other functional axes — plan-then-execute flow, CLAUDE.md integration, omnimodal input, the built-in review agent, polyglot stability — and how each fits an individual workflow.

Mythos Model Preview — What's Next

In the same announcement Anthropic previewed its next model class, Mythos. The wording is "coming weeks," with no exact date. The positioning reads as Claude Opus 4.8 being the last upgrade in the current generation, and Mythos opening the next.

Anthropic's pattern — Claude 3 → 3.5 → 4 → 4.5 → 4.7 → 4.8 with rapid incremental upgrades, then a larger jump to the next generation — places Mythos in the second category. The phrase "Mythos-class models" (plural) in the announcement hints that this won't be a single model but multiple variants released together.

The reasonable move right now: move workflows over to Opus 4.8 while waiting for Mythos. Dynamic Workflows, Fast Mode, and mid-task system messages will likely carry forward to Mythos, so getting familiar with them now prepares for the next step.

Which Workloads Are Most Affected

Mapping Claude Opus 4.8's changes to workload types:

Workload	Impact	Reason
Large monorepo refactors	★★★	Dynamic Workflows + 1M context finishes in one session
CI/CD code review automation	★★★	Fast Mode gives ~5× hourly throughput (2.5× speed, 2× price)
Multi-file bug fixes	★★	Terminal-Bench +8.5pp, honesty gain reduces false passes
Simple chatbots, short replies	★	Little change, same cost
Data analysis and report automation	★★	Mid-task system messages allow steering
Learning and debugging assistant	★★	Honesty gain shows up most clearly here

Dynamic Workflows in particular reshapes the landscape against Grok Build CLI and Codex CLI. With Claude Code stepping from 200K to 1M input plus 1,000 subagents, "hand the whole big job to one tool" becomes the most natural pattern here.

In the same window, Google Gemini Spark staked out the Workspace background automation lane, and Grok's May 2026 wave covered cross-model routing and Connectors. May–June 2026 is the period when each AI tool is locking in the lane it owns. Claude Opus 4.8's lane is "long, complex work taken end to end in a single session."

Bottom Line

Claude Opus 4.8 holds the price flat while stacking Dynamic Workflows for finishing big work in one go + Fast Mode for trading dollars for throughput + honesty gains that reduce false passes in a single release. It's the bridge to Mythos, but workflows moved over today carry forward cleanly. For Claude Code users, the fastest way to feel the difference is to switch the model to Opus 4.8 and run Dynamic Workflows on something real.

Bonus — Where the Names "Opus" and "Mythos" Come From

Model names sometimes spark the question "why this word?" Anthropic's Claude lineup — Haiku · Sonnet · Opus — borrows names from literary and musical forms that grow in length and weight. Among them, Opus is Latin for "work, labor, output." The Op. 27, Op. 53 you see on Beethoven and Chopin sonata scores are exactly this word, used as catalog numbers for an artist's significant compositions. It carries the weight of a major work, and that's why it sits at the top of the Claude lineup as the model that handles the largest, deepest tasks.

The next-generation name, Mythos, steps a level higher. From the Greek μῦθος, it means myth, legend, a larger story. The shift is from a single "work" to a narrative with a world behind it. Anthropic's choice of "Mythos-class models" (plural) in the announcement aligns with that reading — not one model, but several threads of a bigger story unfolding together as the next generation.

A small naming change, but reading it as from Opus to Mythos — from "work" to "myth" — makes the next announcement worth waiting for with a little more interest.