Anthropic shipped Claude Opus 4.8 today. Benchmark scores are up across coding, computer-use, and agentic tasks. Pricing is unchanged. Three capabilities co-launching with the model deserve more attention than the numbers: Dynamic Workflows, a repriced fast mode, and mid-conversation system messages that change how production agentic systems are built. Together they signal where Anthropic is placing its infrastructure bets for enterprise workloads.

Opus 4.8: Benchmarks and What They Reflect

Opus 4.8 is positioned as the best generally available model for coding and agentic enterprise workflows. Pricing holds at $5 per million input tokens and $25 per million output tokens. Benchmark gains are real, especially on OSWorld-Verified at 83.4%, where early access partners describe a measurable improvement over both Opus 4.7 and competing frontier models on computer-use and browser-agent tasks.

Honesty calibration is the result worth examining most carefully. Anthropic's alignment team reports Opus 4.8 is approximately four times less likely than its predecessor to let flaws in its own code pass without comment. The model flags uncertainties, pushes back on unsound plans, and does not claim progress it has not verified. Early access partners at Shopify and Harvey both surface this independently. For enterprise agentic deployments, silent failure is one of the most common reasons pilots stall before production. Opus 4.8 addresses that directly.

Anthropic's alignment assessment also notes Opus 4.8 reaches new highs on prosocial traits including supporting user autonomy and acting in the user's best interest, with rates of misaligned behavior substantially lower than Opus 4.7.

Dynamic Workflows: Model-Native Orchestration at Scale

Dynamic Workflows, in research preview inside Claude Code, allows a single session to plan a large task end-to-end, spawn tens to hundreds of parallel subagents, and verify outputs before delivering results. This is not a chatbot operating on a larger context window. It is closer to an orchestration layer with a model-native planner at the center, capable of running for hours or days with progress saved incrementally so an interrupted job resumes rather than restarts.

Anthropic's reference case: the Bun codebase, ported from Zig to Rust across approximately 750,000 lines, with 99.8% of the existing test suite passing, in eleven days from first commit to merge. Parallel subagents handled file-by-file porting with two reviewers per file. A fix loop ran until build and test came back clean. A follow-on overnight workflow identified unnecessary data copies and opened a pull request per finding for final review.

For enterprise architects evaluating legacy modernization timelines, 750,000 lines in eleven days is a number worth pressure-testing against your own migration backlog before setting it aside.

Enterprise governance note: Dynamic Workflows are off by default for Enterprise plan customers at launch and require admin enablement. That is the appropriate default for regulated environments, but it puts adoption on the IT and security review track rather than the user discovery track. Administrators should plan for this rather than assume it surfaces organically.

Fast Mode Repriced, Mid-Conversation System Messages Added

Fast mode for Opus 4.8 runs at $10 per million input tokens and $50 per million output tokens: 2.5 times the output speed at a 66.7% price reduction versus fast mode for Opus 4.7. For organizations running high-volume agentic workloads where latency is a constraint, this changes the economic model for parallel subagent sessions at production scale.

The Messages API now accepts system entries inside the messages array. Developers can update Claude's instructions during an active conversation or agent turn without breaking the prompt cache or routing the update through a user turn. Permissions, token budgets, and environment context can be modified as an agent runs. Teams building production agentic systems on the API will recognize this as a meaningful reduction in infrastructure complexity.

Benchmark Reference

Note: Anthropic revised its OSWorld-Verified methodology to better reflect real-world performance. Opus 4.7's score is restated to 82.3%. Buyers comparing against previously published numbers should account for the methodology change before drawing conclusions about the improvement margin.

Project Glasswing and the Capability Horizon Above Opus

Anthropic disclosed that a model class operating above the Opus tier already exists. Mythos Preview is available to a small number of organizations for cybersecurity work under Project Glasswing. The constraint on broader release is cyber safeguards, not capability. Anthropic expects to bring Mythos-class models to all customers within weeks.

For enterprise buyers considering multi-year AI infrastructure commitments, this is relevant timing information. If Mythos ships on the stated schedule, Opus 4.8's position as the frontier ceiling will shift before most enterprise evaluation cycles conclude. The capability roadmap above Opus belongs in the planning conversation alongside the model available today.

Availability

Claude Opus 4.8 is available today across all Claude products and the API, including Google Cloud Vertex AI, Amazon Bedrock, and Microsoft Foundry. The API model string is claude-opus-4-8. Effort control is available to claude.ai and Cowork users on all plans. Dynamic Workflows are in research preview on Claude Code for Enterprise (admin-enabled), Team, and Max plans.