Parsec vs Woz

Woz (WOZCODE) is the strongest in-bench rival — a measured 35% token cut and, unlike the rest, an agentic quality benchmark behind it. In the same public benchmark, on cost per solved task, Parsec still comes out ahead — $1.45 to $2.33, solving 62 to 55. Compare · Dasein · 2026-07-15. Scatter of the six benchmark arms plotting tasks solved against cost per solved task. Parsec sits top-left at 62 solved and $1.45. Woz sits lower and to the right at 55 solved and $2.33. Caveman ($2.05, 58), the no-compression baseline ($2.58, 57), RTK ($3.07, 54), and Headroom ($3.66, 58) trail further right. All points are measured on the Code-Compression Bench. Key takeaways: Woz cuts input tokens by a measured 35% (203.1M vs 312.2M baseline), and its 80.2% Terminal-Bench 2.0 score is an agentic benchmark, not single-shot QA. Parsec runs at $1.45 per solved task, against $2.58 uncompressed — 39% off the total bill on 100 graded SWE-bench Verified tasks, while solving more (62/100 vs 57/100). Both ran the same benchmark, head to head. Same scaffold, same model, same grader. Parsec solved 62 to Woz's 55, at $1.45 vs $2.33 per solved task, in 10.8 hours to Woz's 17.8. Measured, Woz was the slowest arm. Woz markets "30–40% faster" (its own self-labeled estimate); on the bench it ran 17.8 hours — 23% slower than the baseline. What Woz is: A paid Claude Code plugin (YC W25) that replaces Claude Code's built-in file tools with AST-aware tools — function bodies stubbed, types and exports kept — and routes read-only exploration to a Haiku sub-agent. It cut input tokens to the main model 35% in our run. Its flagship quality benchmark is an 89-task agentic harness, and the 80.2% Terminal-Bench 2.0 figure is corroborated on the official leaderboard. One trade-off from that same independent source: the 80.2% is measured on Claude Opus 4.7, a stronger model than the Opus 4.6 of the "Claude Code" line it is compared against — the comparison moves the model, not just the layer. What Parsec is: Dasein's context-compression layer for coding agents. It sits in the request path and curates the working context that grows every turn — tool results, file reads, transcript — so the model reads what it needs and no more, without churning the provider's prompt cache. It is provider-agnostic: frontier APIs and open models alike, nothing else in the stack changes. The measured result: $1.45 per solved task · 62/100 solved vs 57 baseline · −39% total cost · −25% wall-clock. One fixed agent (headless Claude Code on claude-sonnet-4-6), 100 SWE-bench Verified tasks, official Docker grader, cache-aware pricing. In the same benchmark, side by side: Parsec — 62/100 solved — $1.45 per solved task — $89.65 total cost (−39% vs base) — −54% input vs base — 10.8 hours wall-clock — measured, Code-Compression Bench. Caveman — 58/100 solved — $2.05 per solved task — $118.99 total cost (−19%) — −19% input — 12.0 hours — measured, Code-Compression Bench. Woz (WOZCODE) — 55/100 solved — $2.33 per solved task — $128.28 total cost (−13%) — −35% input — 17.8 hours — measured, Code-Compression Bench. Baseline (no compression) — 57/100 solved — $2.58 per solved task — $147.30 total cost — baseline input — 14.4 hours — measured, Code-Compression Bench. RTK — 54/100 solved — $3.07 per solved task — $165.77 total cost (+13%) — +16% input — 16.2 hours — measured, Code-Compression Bench. Headroom — 58/100 solved — $3.66 per solved task — $212.14 total cost (+44%) — +6% input — 16.3 hours — measured, Code-Compression Bench. Woz posts the second-largest cost cut (−13%) and a −35% input reduction — yet solves two fewer than the baseline and is the slowest arm. Diagram of Woz's request path. The coding agent hands off to Woz, which replaces the built-in file tools with AST-aware reads and routes read-only exploration to a Haiku sub-agent; a loop between Woz and the sub-agent marks the extra round-trips, before the main call reaches the Anthropic API on claude-sonnet-4-6. A footer notes the measured result: 35% fewer input tokens, but 17.8 hours wall-clock — the slowest arm, 23% slower than baseline, at 14.3 seconds mean latency. When to choose what. Choose Parsec when: Cost per solved task is the number that matters — Parsec's $1.45 beats Woz's $2.33 on the same 100 graded tasks, at a measured −39% total cost. Wall-clock time is part of the bill — 10.8 hours against Woz's 17.8, the fastest arm versus the slowest. The working context is what's growing — Parsec cuts input tokens 54% (vs Woz's 35%) by curating the transcript, not just stubbing file reads. The model must stay fixed and the quality delta must be clean — held on claude-sonnet-4-6, Parsec solved 62/100 to Woz's 55. Stay with Woz alone when: The bottleneck is oversized file reads — AST-aware truncation that stubs function bodies while keeping types and exports is a good fit. Read-only exploration dominates the workload — offloading it to a Haiku sub-agent is a measured −35% input-token cut. You want an agentic quality signal, not a QA proxy — Woz's 80.2% Terminal-Bench 2.0 is an 89-task agentic benchmark. You are already on the Claude Code plugin path and latency on long tasks is not the constraint — the 35% token reduction still holds. FAQ. Can I run Parsec and Woz together? They are two different layers on the same request path, and the benchmark ran each on its own — a head-to-head, not a stack. Woz swaps Claude Code's file tools and adds a Haiku sub-agent; Parsec curates the growing working context and is not tied to the Claude Code plugin model. Where does the headline-vs-measured gap come from? Woz's marketing numbers and its bench result diverge for three mechanical reasons. First, "Cut costs in half" is a favorable-case ceiling ("up to 50%"); under cache-aware pricing a cached prefix is billed far below fresh input, so a real −35% token cut becomes a −13% bill cut. Second, "30–40% faster" is, in Woz's own words, "the only metric we estimate" — measured, the Haiku sub-agent's extra round-trips made it the slowest arm at 17.8 hours (+23%, 14.3 seconds mean latency). Third, the 80.2% quality figure is real but runs on Opus 4.7, stronger than its Opus 4.6 baseline; with the model held fixed here, Woz solved 55 to the baseline's 57. How were these numbers measured? Every arm ran in one public benchmark — a single fixed scaffold on claude-sonnet-4-6, 100 SWE-bench Verified tasks, the official Docker grader, and one cache-aware price table. Only the compression layer changed from arm to arm. Woz's and Parsec's figures in the table, and the four other arms', all come from that one run. Run the numbers yourself: The benchmark, methodology, and per-task data are open at https://github.com/daseinlabs/code-compression-bench Wiring Parsec into an existing request path takes one change, and the number that matters is the one measured on your own traffic at https://daseinlabs.ai Sources: Code-Compression Bench, run 2026-07-04 — sponsored and operated by Dasein, identical methodology for every arm including Parsec's — https://github.com/daseinlabs/code-compression-bench Code-Compression Bench, per-task data (paired.csv) and full per-arm rollup, run 2026-07-04 — https://github.com/daseinlabs/code-compression-bench WOZCODE, How it works (cost, speed, and token claims), retrieved 2026-06-24 — https://www.wozcode.com/how-it-works Terminal-Bench 2.0 official leaderboard (WOZCODE 80.2% on Opus 4.7; plain Claude Code 58.0% on Opus 4.6), retrieved 2026-06-24 — https://www.tbench.ai/leaderboard/terminal-bench/2.0 Anthropic, Claude Sonnet 4.6 announcement and list pricing — https://www.anthropic.com/news/claude-sonnet-4-6