Claude Opus 4.8: New Sub-Agent Workflows and Cheaper Fast Mode

Anthropic describes Claude Opus 4.8 as a “modest but tangible improvement” — and that framing is accurate. On SWE-Bench Pro, the agentic coding benchmark, Opus 4.8 reaches 69.2% against 64.3% for Opus 4.7 and 58.6% for GPT-5.5. The benchmark gains are real. But the operationally significant update is a Fast Mode that costs one-third what previous fast model access cost.

What Changed in 4.8

The headline benchmark improvement carries through on reasoning efficiency as well. On GDPval-AA — a productivity benchmark measuring token and step efficiency — Opus 4.8 requires 15% fewer passes and produces 35% fewer output tokens than Opus 4.7 while achieving equivalent output quality. That efficiency compounding matters for high-volume workflows: the cost reduction isn’t just in per-token pricing, it’s in how much the model does with each token.

Dynamic Workflows is the most architecturally significant new capability. Opus 4.8 can spawn hundreds of parallel sub-agents within a single session, enabling codebase-wide migrations, large-scale analysis tasks, and distributed execution patterns that weren’t viable in single-threaded session models. For teams already running agentic CI and repository automation workflows, this changes what’s achievable in a single coordinated session.

Anthropic also reports that Opus 4.8 is “more likely to flag uncertainties about its work and less likely to make unfounded assertions” — with measurably fewer undetected bugs than its predecessor. In agentic coding contexts, an agent that surfaces its uncertainty is substantially less risky than one that silently generates plausible-but-wrong outputs.

The Fast Mode Pricing Shift

Standard API pricing is unchanged: $5 per million input tokens, $25 per million output. Fast Mode now costs $10 per million input and $50 per million output — one-third the per-token cost of the previous fast offering.

For agentic workflows, the routing decision matters. Fast Mode trades some reasoning depth for throughput; for most tool-use steps in a longer agent loop, throughput matters more than marginal reasoning quality. At one-third the previous cost, routing routine agentic steps to Fast Mode rather than full-depth Opus is now an economically compelling choice, not just a compromise.

Effort Controls add a third dimension: adjustable work levels from standard to max, where higher settings consume more tokens but produce more thorough outputs. Combined with Fast Mode routing, this creates a more granular cost-quality tradeoff than the binary “full model vs. cheap model” decision of previous generations.

Source: The Decoder

Content created with AI assistance and reviewed for accuracy.