On May 28, 2026, Anthropic officially released Claude Opus 4.8, with the model API identifier claude-opus-4-8. Unlike previous versions that emphasized single-model capability improvements, Opus 4.8's core change is a shift from "model capability enhancement" to a "configurable agent-workflow infrastructure." This strategic pivot directly addresses a key pain point in enterprise AI deployments: how to keep large models reliable and controllable in long-running, multi-step complex tasks.

Standard pricing remains the same as Opus 4.7: $5 per million input tokens and $25 per million output tokens. But Fast Mode is priced at one-third of the previous generation (input $10 per million tokens, output $50 per million tokens) while running 2.5x faster.

Databricks, in tests within its Genie AI Agent, found that for the same workload Opus 4.8 reduced token costs by 61% compared with Opus 4.7. This data comes from production environments rather than labs, making it more relevant for real deployments.

Benchmark tests: multiple comparisons show dominant results

At launch Anthropic published comparisons with GPT-5.5 (released April 23, 2026) and Gemini 3.1 Pro. Opus 4.8 led on six key benchmarks:

The only benchmark where GPT-5.5 retained an edge was Terminal-Bench 2.1 (a terminal coding benchmark), where GPT-5.5 scored 78.2% versus Opus 4.8's 74.6%. This suggests GPT-5.5 remains competitive for developer workflows centered on command-line/terminal tasks.

GDPval-AA is an authoritative benchmark developed by Artificial Analysis using the open-source Stirrup testing framework to simulate economically valuable enterprise tasks. Opus 4.8 scored 1890 Elo on that benchmark, corresponding to roughly a 67% win rate, which was characterized as a "meaningful lead rather than a rounding error."

Core innovations: from "more capable" to "more honest"

The technical upgrades in Opus 4.8 are reflected not only in benchmark scores but also in a deep restructuring of agent workflows:

Configurable Effort Modes: Users can set the "depth of thought" and "execution duration" Claude invests in tasks, addressing the polar extremes of "overthinking" and "underthinking" in agent tasks.

Dynamic Workflows — research preview: Claude Code adds the ability to launch hundreds of parallel sub-agents within a single session, autonomously plan multi-step workflows and adjust dynamically, no longer relying on manually preconfigured prompt chains. Anthropic positions this as a core capability for enterprise code-migration scenarios — for example, large-scale refactors involving thousands of files.