Anthropic shipped Claude Opus 4.8 on Thursday, less than two months after Opus 4.7 landed in April. The new model is faster, more accurate on coding benchmarks, and — by Anthropic's own accounting — less prone to stating things it cannot support. It also arrives alongside a research preview of Dynamic Workflows, a feature in Claude Code that allows the model to spawn and coordinate hundreds of parallel subagents to tackle jobs that would otherwise take days of human project management. The update lands as Anthropic prepares a broader release of its Mythos-class security model and, according to multiple reports, draws closer to a public offering later this year.
The Benchmark Numbers
On SWE-Bench Pro, the leading benchmark for AI-powered coding agents, Opus 4.8 scored 69.2 percent. That is up from 64.3 percent for Opus 4.7 and well ahead of OpenAI's GPT-5.5, which sits at 58.6 percent. On agentic computer use (OSWorld-Verified), Opus 4.8 reached 83.4 percent, nudging past its predecessor's 82.8 percent and comfortably ahead of GPT-5.5 at 78.7 percent and Gemini 3.1 Pro at 76.2 percent.
The gains on reasoning are equally notable. On Humanity's Last Exam, a multidisciplinary benchmark designed to be difficult enough that no single human expert can consistently pass it, Opus 4.8 scored 49.8 percent without external tools and 57.9 percent with tools. Anthropic says the model also shows improved performance on financial analysis and longer-horizon knowledge work, two domains where enterprise customers have historically reported the most friction with earlier versions.
Claude Opus 4.8 at a Glance
- SWE-Bench Pro (agentic coding)69.2% (vs. 64.3% for Opus 4.7)
- Humanity's Last Exam (with tools)57.9%
- OSWorld-Verified (computer use)83.4%
- Code flaw miss rate vs. Opus 4.74× lower
- Standard API pricing (input / output)$5 / $25 per million tokens
- Fast Mode pricing (2.5× speed)$10 / $50 per million tokens
Dynamic Workflows and the Scalability Leap
The headline capability addition is Dynamic Workflows, currently available as a research preview in Claude Code for Enterprise, Team, and Max plan subscribers. The feature lets Claude plan a large task, then spin up hundreds of parallel subagents within a single session to execute it, before verifying outputs and reporting back. Earlier Opus models could handle agentic work in sequence; Dynamic Workflows puts that work in parallel.
The practical demonstration Anthropic points to is codebase-scale migration: taking a large codebase — hundreds of thousands of lines — from one framework or language version to another, from kickoff through merge, using the project's existing test suite as the acceptance bar. Subagents handle individual modules simultaneously while a coordinating layer tracks dependencies and resolves conflicts. The same architecture applies to anything with a well-defined output that can be verified: generating comprehensive test suites, auditing code for security patterns, or refactoring across a monorepo.
Opus 4.8 also extends how long individual subagents can run compared to 4.7, which was the main practical limit on the kinds of work Dynamic Workflows could tackle. Users now control how much thinking effort the model applies — from Low (faster, uses fewer rate-limit credits) through High (the new default) to Max — a dial that determines the depth of reasoning Claude brings before acting.
"Sharper judgement, more honesty about its progress, and the ability to work independently for longer than its predecessors." Anthropic, announcing Claude Opus 4.8, May 28, 2026
Pricing, Fast Mode, and the Honesty Promise
Standard pricing for Opus 4.8 is unchanged from 4.7: $5 per million input tokens and $25 per million output tokens, with up to 90 percent savings available through prompt caching and 50 percent through batch processing. Fast Mode, which runs at 2.5 times standard speed, is priced at $10 per million input tokens and $50 per million output tokens — Anthropic says that is three times cheaper than the previous fast mode tier, which had been available on earlier models.
The "most honest" framing Anthropic is attaching to Opus 4.8 maps to two concrete improvements. The model is four times less likely than Opus 4.7 to miss flaws in the code it produces, and it shows a measurable reduction in unsupported claims — instances where the model states something with apparent confidence that it cannot actually verify. Those changes matter most in long-running agentic sessions, where a single overconfident intermediate step can cascade into incorrect downstream decisions. The extended session length in Dynamic Workflows makes that honesty property especially relevant: a model running hundreds of subagents over a large codebase needs to know when it is uncertain, and to say so.
Anthropic noted in the release that Mythos-class models — the company's more advanced security-focused architecture, currently in limited preview through Project Glasswing — are expected to reach broader availability "in the coming weeks." Opus 4.8 does not replace Mythos in capability or intent; it advances the general-purpose flagship line. The Opus 4.7 release in April introduced better vision and a higher coding ceiling; 4.8 pushes that ceiling further and adds the orchestration layer that enterprise customers have been requesting since Claude Code's rapid adoption curve began earlier this year.
Claude.ai and Cowork subscribers can access Opus 4.8 directly by selecting it from the model picker. Dynamic Workflows requires an Enterprise, Team, or Max plan and is flagged as a research preview, meaning the feature may change before full release.
