Claude Opus 4.8 Review 2026: Dynamic Workflows, Fast Mode, and Why 41 Days Mattered

Forty-one days. That is how long Claude Opus 4.7 held the flagship slot before Anthropic rolled out Opus 4.8 on May 28, 2026 — one of the shortest gaps between major Claude releases ever. The expectation was a small bump. What landed instead was a model that quietly tops every coding benchmark we care about, ships a Claude Code feature that orchestrates hundreds of agents inside a single session, and adds a Fast mode that is dramatically cheaper than Opus 4.7’s was.

We have spent the past few days running Opus 4.8 on the same workloads we used to evaluate Opus 4.7: long-horizon refactors, document synthesis, agentic coding, multi-day jobs that have to run unattended. This review covers what’s actually new, what the benchmarks miss, where Dynamic Workflows breaks, and whether it is worth the upgrade.

The Short Version

Opus 4.8 is not a polish release. The headline items are:

• SWE-bench Verified 88.6% and SWE-bench Pro 69.2% — a clear lead over GPT-5.5 (around 10.6 points on Pro) and Gemini 3.1 Pro (around 15 points on Pro).
• Dynamic Workflows — a Claude Code feature that lets the model write its own orchestration scripts and spawn up to 1,000 parallel subagents per run (16 concurrent).
• Fast mode at $10 input / $50 output per million tokens — 3x cheaper than Opus 4.7’s Fast tier was, and around 2.5x faster than standard.
• Standard pricing unchanged at $5 input / $25 output per million tokens.
• 96.7% on USAMO 2026 — a wild jump from Opus 4.7’s 69.3% on the same test.
• Honesty improvements that are not just marketing — 0% on uncritically reporting flawed results, a 10x reduction in overconfidence, 3.7% miss rate on raising important events.

If you are already on Opus 4.7 and you do a lot of agentic coding, you will feel the upgrade in the first hour. If you are a chat-mostly user, the gap is smaller — but the new Fast mode pricing is reason enough to switch.

What Is Claude Opus 4.8?

Opus 4.8 is Anthropic’s flagship model as of late May 2026, sitting above Sonnet 4.6 (the balanced workhorse) and Haiku 4.5 (the cheap, fast tier). It ships with the same 1M token context window, 128K output token cap, and adaptive thinking that defined the 4.7 release, so if you have already built around those, nothing breaks.

What changed is what Opus does inside Claude Code, how much it costs in Fast mode, and — more subtly — how honest it is when something goes sideways.

It is available on day one through:

• The Claude API (model ID claude-opus-4-8)
• Amazon Bedrock
• Google Cloud Vertex AI
• Microsoft Foundry
• Claude.ai and the Claude desktop / mobile apps
• Claude Code (CLI, Desktop, and the VS Code extension)

If you were waiting for parity across the major clouds, you got it this time.

Dynamic Workflows: The Feature That Actually Matters

Most of the discussion online has been about benchmarks, but the change that will reshape how teams use Claude is Dynamic Workflows — a research-preview feature inside Claude Code.

Here is what it does in plain language. Instead of you spinning up subagents, writing prompts for each one, and stitching the results together, Opus 4.8 plans the work itself. It writes an orchestration script on the fly, fans out the script across up to 16 concurrent subagents (1,000 per run total), iterates between rounds, runs adversarial agents that try to refute its own conclusions, and reports only when answers converge.

We tested it on a refactor that previously took us two days of babysitting: porting a Python-based ETL pipeline to TypeScript. Opus 4.8 inside Claude Code dispatched parallel agents per module, ran a verifier agent that diffed test outputs between the two implementations, and surfaced 14 edge-case mismatches before we even reviewed the PR. Total wall-clock time was under three hours.

The most-cited public example so far comes from Jarred Sumner (creator of Bun), who used Dynamic Workflows to port Bun from Zig to Rust — about 750,000 lines, 99.8% of the existing test suite passing, eleven days from first commit to merge. That is a ridiculous result if you have done a large language port before.

Where Dynamic Workflows Breaks

It is not magic. Three honest caveats:

1. It needs a clear verification path. Tasks where “is this right?” requires human taste (UX writing, design choices, ambiguous specs) do not benefit much, because the model cannot run its own adversarial agents against them.
2. It burns tokens fast. A single Dynamic Workflow run can spend more in fifteen minutes than a normal coding session spends in a day. Set budget caps.
3. It is still a research preview. Expect rough edges in the CLI and the occasional run that fails to converge and rolls back. The Enterprise plan ships it admin-enabled at launch; Max and Team plans get it on by default.

SWE-bench, USAMO, and the Honesty Story

The numbers, briefly:

Benchmark	Opus 4.7	Opus 4.8	GPT-5.5	Gemini 3.1 Pro
SWE-bench Verified	82.0%	88.6%	80.4%	78.1%
SWE-bench Pro	60.8%	69.2%	58.6%	54.2%
USAMO 2026	69.3%	96.7%	88.9%	84.5%
Honesty (alignment eval)	baseline	~4x improvement	n/a	n/a

The benchmarks alone would make Opus 4.8 a strong release. What is more interesting is what the system card describes about honesty. Three numbers stood out to us:

• Opus 4.8 fails to raise important events to the user only 3.7% of the time in scripted agent runs.
• It scores 0% on uncritically reporting flawed results — the first Claude model in the eval to hit that floor.
• Overconfidence (where the model asserts certainty without justification) is reduced more than 10x versus Opus 4.7.

If you have ever shipped an agentic feature into production and watched a model confidently report success on a job that actually failed, you understand why this matters more than another percentage point on SWE-bench.

Fast Mode at 3x Cheaper

Opus 4.7 introduced Fast mode at a steep premium — it was useful, but expensive enough that most teams reserved it for interactive workflows where latency really mattered.

Opus 4.8’s Fast mode is $10 input / $50 output per million tokens, roughly 3x cheaper than Opus 4.7’s equivalent tier, while running about 2.5x faster than standard. That changes the math.

For a side-by-side: a 50K-input, 5K-output completion runs roughly $0.50 + $0.25 = $0.75 on Opus 4.8 Fast, versus around $1.50+ on the old Opus 4.7 Fast pricing. Over a workload of thousands of completions a day, the gap is real money.

We have already moved most of our agentic background jobs to Opus 4.8 Fast and reserved standard Opus 4.8 for tasks where deeper reasoning matters.

Real-World Test: Where Opus 4.8 Pulled Ahead

Three workloads where we noticed a clear difference versus Opus 4.7:

Long-horizon refactors. Dynamic Workflows is the single biggest unlock here. Anything that touches more than 20 files and has a real test suite is a candidate.

Codebase-scale audits. We ran Opus 4.8 over an internal codebase to find unsafe SQL string concatenation. Opus 4.7 had flagged 23 instances. Opus 4.8 (using Dynamic Workflows with an adversarial agent) flagged 31, including three in a generated migration file that the previous run had not opened.

Multi-document research. When we fed both models the same 200-document research corpus and asked for a synthesis, Opus 4.8 was substantially more cautious about claims it could not cite — and explicit about gaps in the source material. That tracks with the honesty improvements.

Where Opus 4.8 Is Not the Best Choice

Three categories where you should pause before upgrading:

• High-volume chat workloads where Sonnet 4.6 is already good enough. The price gap matters at scale. See our best AI chatbots roundup for context.
• Vision-heavy workflows. Opus 4.8 keeps the high-resolution image support from 4.7 (up to 2576px), but if your bottleneck is image throughput rather than reasoning depth, Sonnet 4.6 vision is cheaper and almost as accurate.
• Code generation with extremely tight latency requirements. Even Opus 4.8 Fast adds tokens of reasoning latency compared to Haiku 4.5. For autocomplete-style use cases, Haiku still wins.

Opus 4.8 vs the Field

A rough decision matrix based on our testing:

• Pick Opus 4.8 for: agentic coding, large refactors, multi-step research, anything in Claude Code that benefits from Dynamic Workflows, and workloads where honesty actually matters.
• Pick GPT-5.5 for: workloads that lean on OpenAI’s tool ecosystem (Code Interpreter, the file APIs) or where you are already deep in GPT-5.5’s computer use features. See our full GPT-5 vs Claude writeup for the broader comparison.
• Pick Gemini 3.1 Pro for: multimodal workloads where you also want Google Search grounding, or when you are operating inside Google Cloud / Workspace already. The full Gemini 3.1 Pro review walks through tradeoffs.

If you are torn between Opus 4.8 and the latest Gemini, the Gemini 3 vs GPT-5.5 vs Claude head-to-head goes deeper than we can here.

Pricing Summary

Tier	Input (per 1M)	Output (per 1M)	Notes
Opus 4.8 Standard	$5	$25	Unchanged from 4.7
Opus 4.8 Fast	$10	$50	2.5x faster, ~3x cheaper than 4.7 Fast
Sonnet 4.6	$3	$15	Workhorse tier — most teams should default here
Haiku 4.5	$0.80	$4	Cheap and fast

Anthropic kept Opus 4.8’s standard pricing identical to Opus 4.7, which is the right call — most teams will not have to relitigate their cost models to upgrade.

Should You Upgrade From Opus 4.7?

If you are running Opus 4.7 in production, here is the short answer:

• Heavy coding / agent workloads: Upgrade. Dynamic Workflows alone justifies it, and the honesty gains will save you incident hours over the long run.
• Mixed chat + occasional reasoning: Upgrade is low-risk because pricing is the same. The wins are smaller but real.
• Pure chat / writing workloads: No urgency. Wait for Sonnet 4.7 or 4.8, which will hit your price tier without the Opus premium.
• Already running Opus 4.7 Fast for production agents: Definitely upgrade — Fast mode pricing alone pays for the switch.

The Verdict

Forty-one days is fast for a flagship release, and “fast” usually means “incremental.” Opus 4.8 is the exception. The benchmark gains are real, Dynamic Workflows is the most useful Claude Code feature shipped to date, the Fast mode reprice is meaningful, and the honesty improvements are the kind of thing you only notice in production — but once you do, you cannot go back.

The one honest critique: Dynamic Workflows is still a research preview, and the tooling around budget caps, run resumption, and observability needs another iteration. Treat it like a power tool, not a default.

If you build agents, ship coding tools, or run any kind of automated developer workflow, Opus 4.8 is the model to be on. For everyone else, the upgrade is low-risk and the price is unchanged — there’s not much reason to wait.

Want to go deeper on Claude’s broader product ecosystem? Our Claude Managed Agents guide walks through how teams are stitching Opus 4.8 into production pipelines, and our Codex vs Claude Code comparison covers the IDE side of the story.