Claude Opus 4.8: Stronger Agents, Coding, and Vision

Anthropic released Claude Opus 4.8 on May 28, 2026 as its strongest generally available model before the later Fable/Mythos launch. The official announcement emphasizes autonomous software engineering, complex knowledge work, vision, and computer use. The accompanying system card reports 88.6 on SWE-bench Verified, 69.2 on SWE-bench Pro, 84.3 on single-agent BrowseComp, 88.5 with multiple agents, 83.4 on OSWorld-Verified, and 49.8 on Humanity's Last Exam without tools.

Opus 4.8 is important not because every score leads the market, but because it raises the floor across several high-value agent tasks while publishing unusually extensive safety evidence. The benchmark table also demonstrates why model comparisons are fragile: Anthropic's launch page, system card, later competitor pages, and benchmark leaderboards sometimes use different harnesses or updated methodologies. Every value below is tied to the configuration Anthropic reports.

What changed in the Opus line

The release is an incremental name but a substantial agent update. Anthropic says Opus 4.8 follows instructions more reliably across long tasks, explores repositories more effectively, uses tools with less supervision, and produces more polished knowledge-work artifacts. Early customers describe completing end-to-end migrations and hard pull requests that earlier versions left unfinished.

The model remains hosted. It is available through Claude products, Claude Code, and the Claude API rather than as downloadable weights. Deployment therefore includes Anthropic's serving stack, policy layer, tokenizer, effort settings, context management, and product tools.

“Opus 4.8” should not be used as a complete experimental description. The system card's scores can use different effort levels, agent wrappers, compaction, or tools. A production record needs model ID, surface, effort, tool configuration, date, and retry policy.

Capability snapshot from the system card

The first-party evaluation record spans code, search, science, vision, and computer action:

These are strong results, not a common leaderboard. BrowseComp's multi-agent score includes orchestration overhead. HLE with tools is not directly comparable to a no-tool competitor. OSWorld depends on the action interface and episode settings. The full system card is the correct source when a number affects procurement.

Software engineering is the clearest use case

SWE-bench Verified and Pro ask an agent to understand a repository, modify code, and pass tests. The jump from isolated coding to repository resolution is operationally meaningful: the model must navigate files, preserve conventions, diagnose failures, and avoid unrelated changes.

Still, a score does not describe patch quality. Some solutions pass visible tests while adding fragile behavior, weakening types, or skipping edge cases. A local pilot should grade minimality, maintainability, test relevance, security, and reviewer effort in addition to resolution. Run the same tasks multiple times to measure variance.

Claude Code is a substantial part of the experience. Tool permissions, compaction, subagents, and the system prompt affect persistence. Comparing Claude Code with another provider's raw API is a comparison of systems, not just models. Our software-engineering agent guide describes the repository and review controls required around any high-scoring model.

Search and multi-agent orchestration

BrowseComp measures difficult web research whose answers require finding and combining scattered evidence. Opus 4.8 moves from 84.3 in a single-agent setup to 88.5 with multiple agents. That gain is evidence that parallel search and synthesis can improve coverage; it also spends more calls, tokens, and coordination time.

A multi-agent system can amplify correlated errors if all branches rely on the same weak source or repeat the same assumption. It needs source diversity, claim-level citations, deduplication, and an adjudicator that can reject unsupported synthesis. Cost should be reported per accepted research result.

For live news, the date of the source and event matters as much as retrieval. An agent must distinguish a model announcement from a model-card update or an API alias change. That chronology discipline is the reason this series uses a release ledger rather than search snippets alone.

Computer use and visual reasoning

An 83.4 OSWorld-Verified score shows strong ability to operate graphical environments. This can unlock workflows where no reliable API exists. It also exposes the model to hidden state, prompt injection, accidental clicks, and irreversible actions. Success percentage says little about the severity of the remaining failures.

Use computer action inside isolated accounts with least privilege. Require confirmation before sending, publishing, purchasing, changing permissions, or deleting. Capture screenshots and actions so a run can be audited. Prefer structured APIs when they exist; pixels are a fallback, not a governance strategy.

ChartQAPro and other vision results indicate better handling of dense figures. High-stakes extraction should return the cited region and preserve units. Visual fluency does not eliminate OCR, reading-order, or numeric hallucination.

Tools alter the benchmark definition

Humanity's Last Exam rises from 49.8 without tools to 57.9 with tools. The eight-point difference is not a footnote: it reveals what the deployed system gains from retrieval or computation. Tools can reduce memory errors and arithmetic mistakes, but they introduce source quality, tool failure, and authorization risk.

Procurement tables should maintain separate columns for model-only, model-plus-tools, and model-plus-agent. Mixing them rewards whichever provider uses the most capable scaffold. The same principle applies to effort settings, which trade cost and latency for more reasoning.

The frontier model evaluation guide recommends an evaluation manifest with every non-model component. Without it, a later rerun cannot explain why the score changed.

Safety evidence and remaining risk

Anthropic's system card covers misuse, cyber capability, biological risk, alignment, hallucination, and automated behavioral audits. Opus 4.8 received stronger cyber safeguards than ordinary chat models because general capability increases dual-use capacity. The existence of a system card is valuable; it is not a certification for a buyer's deployment.

Risk depends on tools and data. A model with read-only repository access has a different exposure from one with credentials and network reach. Organizations should threat-model indirect prompt injection, secrets, destructive commands, data exfiltration, and compromised tool output. The model's refusal behavior cannot replace sandboxing and least privilege.

Safety evaluations also have coverage limits. Novel workflows and local languages may fall outside published tests. Log interventions and near misses, and rerun the safety suite after provider updates.

Cost, fit, and migration

Opus-class models are intended for work where a better completion can justify higher cost: hard repository changes, long analytical deliverables, complex research, and computer workflows. Routine extraction or classification may be better served by a smaller model. A tiered architecture can use Opus 4.8 for escalations and a cheaper model for predictable cases.

Migration should compare total task cost, not token price alone. Track accepted completions, output tokens, elapsed time, tool calls, reviewer minutes, and incident rate. A model that solves more in one pass can be cheaper despite a higher rate. A model that produces impressive but unverifiable work can be more expensive.

Keep provider-neutral task definitions and full traces. Hosted models change, and a portable acceptance suite is the only reliable way to detect regression or evaluate a successor.

Verdict

Claude Opus 4.8 was a major May release because it joined high repository resolution, strong search, computer use, visual reasoning, and comprehensive safety reporting. Its best evidence is the breadth and detail of the system card, not a single headline.

The correct operational conclusion is a controlled pilot for difficult, high-value agent work. Pin the model and scaffold, separate tool-assisted scores, apply least privilege, and require objective verification. Later Claude releases may exceed its raw capability, but Opus 4.8 remains an important generally available baseline for the summer 2026 frontier.

Source notes — reviewed 2026

• Anthropic's Claude Opus 4.8 announcement provides the date, product positioning, and customer evidence.
• The Claude Opus 4.8 system card supplies the evaluation values and safety methodology.
• Anthropic's model overview documents hosted models and API access.
• Claude Code documentation describes the first-party coding scaffold that forms part of the deployed system.
• SWE-bench's official site documents the repository benchmark family used in the report.