- Nvidia‘s SoL-Pi system cuts coding-agent token usage nearly in half.
- The gains come from optimizing the agent harness, not the underlying model.
- It validates the finding that the harness rivals the model as a cost lever.
- Token efficiency is becoming the competitive front in agentic coding.
What Happened
Nvidia‘s SoL-Pi system cuts coding-agent token usage nearly in half by optimizing the harness — the scaffolding that manages context, tools, and steps around the model — The Decoder‘s Jonathan Kemper reported on September 26, 2026.
Why It Matters
The result confirms what Databricks’ internal benchmarking suggested last year and what we argued in our harness-versus-model analysis: the wrapper around a model moves total cost as much as the model choice itself. A near-50% token reduction is equivalent to a price cut no provider has to offer — and it comes from engineering, not negotiation. That Nvidia, a hardware company, is shipping harness software shows where the efficiency frontier has moved.
Technical Details
Agent harnesses resend context, tool definitions, and intermediate state on every step, so waste compounds across long sessions; SoL-Pi’s optimization targets that overhead rather than the model’s reasoning. Halving tokens roughly halves both cost and the context pressure that degrades long-running agents. Combined with provider-side moves like GPT-6’s improved prompt caching, agent economics are improving from both ends of the stack.
Who’s Affected
Teams running coding agents at scale see the largest absolute savings. Harness maintainers — Claude Code, Codex, open-source frameworks — get a public benchmark to match. Model providers face a subtle shift: efficiency gains captured in the harness don’t show up as their revenue.
What’s Next
Independent replication across different models and harnesses will show whether the ~50% figure generalizes. Expect harness efficiency to join model price and context length in cost-per-task comparisons.
