A team from NVIDIA, NTU and MIT released SoL-Pi on September 21, 2026, an MIT-licensed extension that cuts token traffic for the open-source Pi coding agent by 44.7% to 49.0% on the 51-task EdgeBench evaluation. API costs dropped roughly 33% while scores stayed within about 6% of the original Pi harness on both GPT-5.6 Sol and Opus 5 backends. The work targets a growing cost problem: coding agents now run for hours, and every edit, test run, and log read feeds back into the model's context window.
The extension ships on GitHub under NVlabs and runs on an unmodified Pi 0.85.1 release with Node.js 22.19 or newer. Rather than lowering per-token cost through faster kernels or quantization, SoL-Pi reduces how many tokens a task consumes by modifying the harness layer - the code that handles tool calls, context management, observations, and delegation.
How an AI found the efficiency mechanisms
A research AI observed execution traces from a separate agent running base Pi, then proposed harness changes and tested them. The search covered 152 proposed directions across six families - context, progress, tools, delegation, prompt and policy, and improvement and evaluation - using 535 executable environments built from GitHub issue-pull request pairs and synthetic tasks. More than 3,000 runs and 60,000 agent-environment interactions fed the loop.
Each search ran as a disposable, isolated loop following the autoresearch cycle with a Ralph Loop implementation step and an independent reviewer. Acceptance rules were fixed before the search started. Every capability metric had to stay within a predeclared tolerance, and candidates had to improve at least one efficiency metric. EdgeBench remained held out: 11 of its 51 tasks were used for one-way acceptance of frozen candidates, and the other 40 for final evaluation. Held-out results never fed back into the search.
The researchers said the approach avoids a known pitfall. "A recent study found that evolved harnesses can overfit their search tasks and give only marginal gains on unseen ones." By freezing acceptance criteria and holding out tasks, the team aimed for mechanisms that generalize.
The four mechanisms that survived
Action Fusion merges a file edit and its follow-up command - a test, build, or run - into a single tool request, returning both outcomes in one observation. This removes a model round trip that base Pi would otherwise make.
Online Context Compact tracks plan steps through update_plan. When a step completes, the harness estimates remaining requests and compares projected input savings against the cost of rewriting the prompt cache. It invokes Pi's native compaction only when the gate passes or when context nears the window limit.
ObservationPack archives tool outputs above 10 KiB locally and sends them in full for the next two provider requests. From the third request onward, the model sees a stable handle, the original size, and a short excerpt of head and tail lines. Exact pages stay retrievable through the handle.
Evidence-Preserving Reducer sends build and test logs of at least 4 KiB to a cheaper model, GPT-5.6 Luna at high, which writes a compact receipt. A deterministic verifier checks the receipt's schema, source hash, exit status, exact quotes, and size. The harness falls back to the original log if verification fails, credentials are suspected, or the receipt is not smaller.
Results on EdgeBench and beyond
On GPT-5.6 Sol, SoL-Pi's efficiency configuration used 1.10 billion tokens versus Pi's 2.15 billion - a 49.0% reduction - while API cost fell from $1,339 to $894. The average score held at 42.0, or 93.7% of Pi's 44.8. On Opus 5, the stack was deployed without further search. It kept 94.3% of Pi's score, cutting token traffic by 44.7% and API cost by 33.5%.
A separate Performance configuration used the single best mechanism per backend - ObservationPack on GPT-5.6 Sol and Action Fusion on Opus 5 - lifting scores 5.3% and 12.8% above Pi respectively. The paper estimates hourly savings of $4.36 to $5.71 versus Pi and $8.75 to $13.50 versus the native Codex and Claude Code harnesses.
On Terminal-Bench 4, a set of 63 CPU-only tasks, SoL-Pi solved 15 tasks versus 18 for both Codex and Pi while lowering total cost by 26.3% versus Pi. On the IMO 2026 benchmark with Lean 4 verification, SoL-Pi passed 3 of 6 problems, matching Pi, at the lowest cost per passed problem. The team also tested an agent swarm with a Codex coordinator and 20 SoL-Pi workers, reaching 1,127 cycles at $60.11 versus 1,366 cycles at $82.12 with Pi workers. A single Codex agent remained the cheapest option at $39.20.
The cross-model transfer was described as preliminary. Mechanisms triggered less often on Opus 5, which the team attributed to the search using only GPT-5.6 Sol trajectories.
Why this matters for development and research teams
SoL-Pi shows that meaningful cost reduction in coding agents does not require switching models or rebuilding infrastructure. The mechanisms sit at the harness layer - the code your team already controls - and ship as an opt-in extension under a permissive license. For teams running long-horizon coding tasks, a 33% drop in API spend with single-digit score trade-offs changes the calculus on when to deploy agents in production pipelines. The approach also demonstrates that automated search over harness configurations can find portable optimizations, though the cross-model transfer still needs tuning when the search backend differs from the deployment backend.
Your membership also unlocks:
