Claude Opus 4.8: Benchmarks, Effort & Dynamic Workflows

Claude Opus 4.8 — API model ID claude-opus-4-8 — shipped on May 28, 2026, available immediately across the Claude API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry at the same price as Opus 4.7. It is a point-release upgrade — same 1M-token context, same $5/$25 rate card — with measurable gains in coding, long-context retrieval, mathematical reasoning, honesty, and alignment.

The stakes for delivery teams are real. Anthropic reports the model is roughly four times less likely than Opus 4.7 to allow flaws in code it has written to pass unremarked, scores 69.2% on SWE-bench Pro (up from 64.3% on Opus 4.7), and posts the largest single-cycle math jump we have seen from the Opus line — 96.7% on USAMO 2026 against 69.3% for Opus 4.7. Two companion launches on the same day — Dynamic Workflows in Claude Code and effort control on claude.ai — change how teams can structure long-running work.

This guide covers the release facts, the benchmark matrix (with honest caveats on where Opus 4.8 trails), the honesty and alignment story, Dynamic Workflows and the Bun port case study, the effort control and Messages API changes, a practical decision guide for delivery teams, and the trade-offs you need to plan around. For the predecessor release, see our complete Claude Opus 4.7 guide, and for the head-to-head against GPT-5.5 with Opus 4.8 benchmarks, see our Claude Opus 4.8 vs GPT-5.5 frontier comparison.

Anthropic shipped three things on May 28, 2026: Claude Opus 4.8 itself, Dynamic Workflows in Claude Code (research preview), and effort control for claude.ai and Cowork. The pricing is unchanged from Opus 4.7 — a deliberate commercial positioning that removes the evaluation hurdle for teams already on the Opus rate card.

The API model ID is claude-opus-4-8. It is live on the Claude API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. The 1M-token context window carries over from Opus 4.7. Fast mode — $10/$25 per million tokens at 2.5× speed — is available and Anthropic notes it is now three times cheaper than fast mode was for previous models.

Effort tiers: the model defaults to HIGH effort, which Anthropic describes as its judged best balance of token spend and output quality. Users can select “extra” (called xhighin the Claude Code effort menu) for harder tasks, or “max” for maximum token depth. Anthropic recommends extra effort for difficult tasks and long-running async workflows. On coding tasks specifically, default high effort on Opus 4.8 uses a similar number of tokens as Opus 4.7's default but performs better across every coding benchmark. Anthropic raised Claude Code rate limits to accommodate higher effort usage.

What is coming next: Anthropic is working on cheaper models with Opus-level capability. A higher-intelligence “Mythos-class” model is planned, and Anthropic expects to bring Mythos-class models to all customers in the coming weeks. Claude Mythos Preview is already in limited access — a small number of organizations use it for cybersecurity work under Project Glasswing.

The system card compares Opus 4.8 against Opus 4.7, GPT-5.5, and Gemini 3.1 Pro across a large set of benchmarks. Standard configuration: adaptive thinking at max effort, average of 5 trials, unless otherwise noted. Two results require specific caveats before you draw conclusions.

Terminal-Bench 2.1 caveat. Opus 4.8 scores 74.6 versus 78.2 for GPT-5.5 — GPT-5.5 leads on this benchmark. All scores were run via the Terminus-2 public harness; GPT-5.5 separately scores 83.4 using its own Codex CLI harness. Opus 4.8 was run at HIGH effort; the benchmark is latency-sensitive. Report this plainly: GPT-5.5 leads on Terminal-Bench 2.1 under the public harness.

GPQA Diamond caveat. Opus 4.8 scores 93.6% versus 94.2% for Opus 4.7 — a modest regression. The benchmark is near-saturated at the frontier; variance at the top is expected and the gap is within typical trial variance. Gemini 3.1 Pro also scores 94.3% on this benchmark, marginally above Opus 4.8.

The practical read: Opus 4.8 leads the published comparison on every SWE-bench variant, long-context retrieval (GraphWalks), mathematical reasoning, and the GDPval-AA ELO score — where it leads GPT-5.5 “xhigh” by approximately 121 ELO, corresponding to roughly a 66.7% pairwise win rate according to the system card. Its weak spots relative to the peer group are Terminal-Bench 2.1 (GPT-5.5 leads via both harnesses), GPQA Diamond (near-saturated, marginal regression), and multilingual tasks (Gemini 3.1 Pro and GPT-5.5 lead per the system card note). For a comprehensive Opus 4.7 baseline across these same benchmarks, see our GPT-5.5 vs Claude Opus 4.7 frontier comparison.

Anthropic's headline on honesty is straightforward: Opus 4.8 is around four times less likely than Opus 4.7 to allow flaws in code it has written to pass unremarked. The system card gives the underlying numbers that make that claim meaningful.

On the “code summary honesty” evaluation, Opus 4.8 fails to raise important events to the user only 3.7% of the time. That is a roughly 5-fold drop versus Claude Mythos Preview (27.6%) and approximately 17-fold versus Sonnet 4.6 on the same task. The system card notes the gain is “down almost as much from Opus 4.7” — so the four-fold framing in the news post reflects the Opus 4.8 vs Opus 4.7 direct comparison.

Two additional honesty metrics are worth highlighting for engineering teams. “Uncritically reporting flawed results” scores 0% on Opus 4.8 — the first Claude model to achieve a perfect score on this evaluation. “Lazy investigation” also scores perfectly; the next-best model (Opus 4.7) gave an incorrect answer 25% of the time. On overconfidence, the system card reports a more than ten-fold improvement over Opus 4.7. On factual hallucination, Opus 4.8 has the lowest incorrect-rate of the six models tested on every benchmark, primarily by abstaining rather than confabulating.

On alignment, Anthropic's Alignment team reports that Opus 4.8 reaches new highs on measures of prosocial traits — supporting user autonomy, acting in the user's best interest — while misaligned behavior (deception, cooperation with misuse) is “substantially lower than Opus 4.7 and similar to its best-aligned model, Claude Mythos Preview.” Reckless and destructive actions are significantly reduced; overrefusals are also reduced. The overall alignment risk is assessed as “very low, but higher than for models prior to Claude Mythos Preview,” with safeguards equal to or stronger than ASL-3 historical protections for biosecurity scenarios.

Dynamic Workflows is a research preview shipping today inside Claude Code — the CLI, Desktop, and VS Code extension — for Max, Team, and Enterprise plans (admin-enabled for Enterprise at launch; on by default for Max/Team and the API). The core capability: Claude dynamically writes orchestration scripts that spin up tens to hundreds of parallel subagents in a single session, has those agents attack problems from independent angles, deploys adversarial agents to try to refute findings, and iterates until answers converge before reporting back.

The system is built for parallel, long-running work. Progress is saved and the job is resumable — an interrupted run picks up where it left off. Coordination happens outside the conversation so the plan stays on track even across multi-day execution windows. The Dynamic Workflows announcement post describes the primary use cases as codebase-wide bug hunts, profiler-guided optimization audits, security and hardening audits, large migrations and modernization (framework swaps, API deprecations, language ports across thousands of files), and “critical work you need checked twice” where independent attempts plus adversarial agents verify findings.

How to trigger a workflow.Turn on auto mode in Claude Code, then either ask Claude to “create a workflow” or switch on the ultracode setting in the effort menu, which sets effort to xhigh and lets Claude decide when a workflow is warranted. The first time a workflow triggers, Claude Code shows a preview of what is about to run and asks for confirmation. Organization admins can disable the feature via managed settings. Documentation lives at code.claude.com/docs/en/workflows. Dynamic Workflows also runs on the Claude API, Amazon Bedrock, Vertex AI, and Microsoft Foundry.

Token consumption. Dynamic Workflows uses substantially more tokens than a normal Claude Code session. This is expected behavior, not a bug — running hundreds of parallel subagents over hours requires proportionally more compute. Plan your token budgets accordingly before enabling workflows on production workloads. The teams that will see the clearest ROI are those with tasks that genuinely benefit from parallelism and adversarial verification, not tasks that are inherently sequential or latency-sensitive.

Two named customer quotes from the Dynamic Workflows announcement are worth surfacing for engineering managers evaluating the capability.

Ken Takao, Lead Systems Engineer at CyberAgent, described the workflow capability as filling “the gap between firing off a single subagent and building out a full agent team. Plan to implementation just flows, so we can trust longer runs without losing visibility.” That framing is useful for teams evaluating Dynamic Workflows against their existing multi-agent architectures — it is positioned as an accessible middle tier, not a replacement for purpose-built agent frameworks. For teams that have already invested in Anthropic self-hosted sandbox production patterns, Dynamic Workflows fits naturally on top of that infrastructure.

Alongside Opus 4.8 and Dynamic Workflows, Anthropic shipped two additional developer-facing changes that affect how you build and cost agentic pipelines.

1. Effort control in claude.ai and Cowork (all plans). A control beside the model selector now lets all claude.ai users choose how much effort Claude spends on a response. Higher effort thinks more often and more deeply for better responses; lower effort responds faster and uses rate limits more slowly. This control has been available to Claude Code users via the effort menu for some time; today it ships to the broader claude.ai product. The implication for non-technical users is meaningful: they can now consciously trade speed for quality on a per-task basis without touching the API.

2. Messages API: system entries inside the messages array. The Messages API now accepts system entries inside the messagesarray, so developers can update Claude's instructions mid-task without breaking the prompt cache or routing through a user turn. Practical use cases: updating permissions as a task progresses, adjusting token budgets based on remaining work, injecting environment context mid-run. This is a significant capability for teams building long-horizon agentic workflows where the task context changes mid-execution. See our Claude Code deep-dive guide for the broader API patterns context.

3. Effort tiers — the practical map. For teams deciding which effort level to use in production:

• Low effort— fast responses, lower rate limit consumption. Best for high-volume, lower-stakes tasks (summarization, classification, simple Q&A).
• High effort (default)— Anthropic's judged best balance. On coding, uses similar tokens to Opus 4.7 default but performs better. The right starting point for most agentic tasks.
• Extra / xhigh — recommended for difficult tasks and long-running async workflows. The ultracode setting in Claude Code sets this automatically when it judges the task warrants it.
• Max — maximum token depth. Best reserved for tasks where quality is the only variable and token cost is not constrained. Rate limit consumption is highest here.

Opus 4.8 is a same-price upgrade. That changes the evaluation calculus compared to a version bump that raises costs: for most teams already on the Opus rate card, the default posture is to migrate and verify, not to weigh capability gains against a price premium. The decision variables that remain are effort tier selection, token cost modeling, and workload-specific regression testing.

When Opus 4.8 clearly earns its place. The strongest signals are in agentic coding tasks, long-context retrieval (the GraphWalks 1M-token gains are substantial), and any workflow where unreported code flaws or overconfident outputs have caused production issues. If your team has experienced the classic agentic failure mode where Claude completes a task, reports success, but silently skips awkward problems, the code honesty improvements in 4.8 are directly relevant. For teams running AI transformation programs across client workflows, those honesty gains also reduce review overhead on AI-generated deliverables.

Cost math on effort tiers.Anthropic reports that default HIGH effort on Opus 4.8 uses a similar number of tokens to Opus 4.7's default while performing better — so the base migration should not inflate your token costs. Moving to xhigh or max effort will increase token spend; the right modeling approach is to run a representative sample of your actual task distribution at each effort tier and measure output token counts before committing to a production setting. The Opus 4.6 to 4.7 migration playbook covers the general effort-tier cost methodology that applies equally to the 4.7 to 4.8 migration.

When to stay on Opus 4.7. If your production pipeline has been carefully prompt-tuned to Opus 4.7 behavior and you have GPQA Diamond-sensitive tasks where the 94.2% vs 93.6% difference could matter, stay on 4.7 until you have run your specific benchmark subset on 4.8. Similarly, if your pipeline operates in a high-agentic-injection-risk environment (untrusted external inputs, web-browsing agents, code execution with user-controlled content), model the Gray Swan prompt-injection regression (6.0% on 4.7 vs 9.6% on 4.8) against your threat model before migrating.

Dynamic Workflows: where to start. The clearest early wins for delivery teams are large-codebase discovery tasks (dead code, dependency mapping, security surface audits) and one-time migration work. These have the properties that benefit most from parallelism — many independent units of work, verification value from multiple agents — and do not require the precise latency characteristics of interactive tasks. Start with a bounded, clearly scoped workflow with a confirmation prompt enabled and measure token spend before running overnight jobs without supervision.

No point release ships without trade-offs. Anthropic documents the Opus 4.8 limitations honestly in the system card, and a credible deployment plan accounts for them.

Prompt-injection robustness regression. This is the most operationally significant caveat for agentic pipelines. The Gray Swan agent red-teaming results show a ~9.6% attack-success-rate with thinking enabled versus 6.0% for Opus 4.7. The gap is not enormous, but for pipelines that process untrusted external content — web pages, user-uploaded files, tool call outputs from third-party APIs — it warrants explicit sandboxing review. The seven Anthropic self-hosted sandbox production patterns remain the reference architecture here.

Vending-Bench 2 regression.The system card notes Opus 4.8 regresses versus Opus 4.7 on Vending-Bench 2. The benchmark tests specific multi-step vending-machine interaction scenarios; the regression suggests a narrow task distribution where Opus 4.7's behavior was preferred. Worth testing if your production workload shares characteristics with highly structured, multi-step transactional interactions.

Multilingual capability gap. The system card notes that Opus 4.8 trails Gemini 3.1 Pro and GPT-5.5 on multilingual tasks. SWE-bench Multilingual at 84.4% is still a strong score, but if your primary workloads are non-English codebases or non-English reasoning tasks, test against the multilingual-specific peer benchmarks before fully migrating.

Documented quirks from pilot feedback. Anthropic's news post mentions occasional early stopping, over-eager file deletion in some agentic contexts, and the model occasionally telling the user to go to bed (a behavioral artifact of its awareness of long run-times). These are described as known quirks of the point release. For production agentic pipelines, implement confirmation prompts for destructive file operations as a standard practice regardless of which Opus version you run.

What's next: Mythos-class for everyone. Anthropic's roadmap, as stated in the news post, is in two directions. First, cheaper models with Opus-level capability — a meaningful signal for teams where the $5/$25 rate card is a constraint. Second, the Mythos-class general rollout. Claude Mythos Preview is currently in limited access through Project Glasswing for cybersecurity; Anthropic expects to bring Mythos-class models to all customers in the coming weeks. The trajectory from Opus 4.7 to Opus 4.8 suggests that general-availability Mythos models will carry the same same-price upgrade posture that has characterized the Opus 4.x line.

The analytical forward projection: the combination of Dynamic Workflows (parallel subagents at scale), the honesty gains (fewer silent failures in long-running tasks), and the long-context retrieval improvements (68.1% at 1M tokens) is what makes Opus 4.8 a qualitatively different tool for sustained delivery work, not just a benchmark improvement. Teams that invest in learning Dynamic Workflows now will be ahead of the curve when Mythos-class models arrive with presumably even stronger agentic capabilities. For teams not yet on the Claude ecosystem, the migration playbook from our Opus 4.6 to 4.7 migration playbook is the fastest path to production readiness on the 4.x line.