The second table on the DeepSeek-V4.1-Flash model card is titled "Performance across agent scaffolds." Same model, same parameters — N=8, temperature 1.0, 1M context, max_steps=500 — eight different harnesses. On DeepSWE, the score runs from 65.5 on OpenCode to 74.2 on mini-SWE, an 8.7-point spread. On Terminal-Bench 2.1, it runs from 84.1 on Codex to 90.6 on DSH Minimal.
The table above it is the one that got quoted. There, V4.1-Flash beats Opus-5.0 by 0.2 points on DeepSWE and 1.5 on TB 2.1. Both leads fit inside the harness spread with room to spare. Fix the harness at Claude Code and the same model scores 69.8 and 88.0 — behind on both.
This is not a gotcha. The vendor printed the denominator itself, one section below the numerator, and that is the story: the thing deciding what a model costs is no longer the model.
The vendor printed the denominator itself, one section below the numerator, and that is the story: the thing deciding what a model costs is no longer the model.
A Berkeley study out in September put a number on it. Running GPT-5.6 Sol on the Pi harness instead of Claude Code cut cost per resolved task from $1.540 to $0.441 — a 71% reduction, same model, same result. Across all 42 within-model harness comparisons in the study, exactly one result hit p < 0.05, where chance alone would produce about two, and none survived a Holm-Bonferroni correction. The caveats matter: published rollouts were sorted by cost, which rules out a paired test, and at 90 rollouts per cell the study can only detect swings of roughly fifteen points. The quality gap is undemonstrated at this sample size, not proven absent. The cost gap is measured.
Writer ran the cleaner experiment. Twenty-two locked enterprise tasks, six foundation models, and only the orchestration layer swapped — a conventional production agent loop against the Writer Agent Harness. Blended cost per task fell 41%, from $0.21 to $0.12. Median wall clock fell 44%, from 48 seconds to 27. Tokens per task fell 38%, from 14.2k to 8.8k. Quality held: 0.78 to 0.81.
Two findings in that paper deserve to be read twice. First, the efficiency gains were model-invariant — every one of the six models got cheaper, by 33% to 61%. Second, the quality gain was not: how much a model improved under the better harness correlated almost perfectly with how strong it already was (r = 0.99, n = 6), a pattern the authors call harness leverage. On that workload, changing the orchestration layer moved cost per task more than switching between the cheapest and the most expensive model would have.
Tomasz Tunguz turned that into a P&L. Two startups bid the same $250k contract. One calls a frontier model on every step and burns $131k on inference. The other coalesces the work into deterministic code and reserves the expensive model for the few steps that need it, burning $37k. Identical revenue, identical non-inference costs. That's 38% gross margin against 75%, and a payback period on a $150k sales cost of nineteen months against ten. Past roughly 8,600 evaluations, the first company loses money on every additional one and has to ration usage — which makes it the worse product exactly when the customer finds it most valuable. And the cheaper company can't be copied by simply routing to a smaller model: knowing which tasks a cheap model clears comes from having watched ten thousand versions of the same work. The cost advantage and the moat are the same asset.
Which brings in the router, the layer above the harness. In one four-day stretch in July, Ramp open-sourced the router it built for its own products and credited it with a 30% cut in LLM costs; Cursor launched its own router, reporting 30% to 50% savings for early enterprise customers versus sending everything to one frontier model; and the Journal reported Stripe in talks to buy OpenRouter for roughly $10 billion, nearly eight times its May valuation.
Routers are best understood as procurement decisions made in milliseconds — how much intelligence to buy for this request. But routing is a decision under uncertainty, because the router cannot know in advance whether a cheap model clears the bar. The cost of being wrong is not the token price; it's the retry, the escalation, the lost cache, the human review, the incorrect business action. And as one practitioner framing puts it, if the harness is the workshop, the router is the dispatcher deciding which worker enters it. A frontier model in a poor harness can lose to a smaller model in a good one.
Perplexity is now shipping that view to users. Its Computer product added effort presets that "combine the orchestrator model and reasoning depth," with the stated goal of "maximum intelligence at minimum cost through multi-model orchestration," and an explicit commitment to keep optimizing a model-agnostic harness as frontier models advance. The pushback in the replies was sharp and worth keeping: minimum cost is not minimum total work, extra model calls add latency, and without showing cost, latency, and model mix per run, the tradeoff is unauditable. A separate reply made the point that retrieval quality is the same tax — a subagent burning three extra tool calls to re-resolve a messy product page costs what picking the wrong model costs.
The spread can be extreme. One harness×model benchmark recorded a 475x difference in credits per task between configurations on the same task, with the cheap configuration succeeding on all observed runs at 0.47 credits against 223. Databricks found one harness cost more than twice as much per completed task as another at equivalent quality inside its own codebase. Vercel replaced eighteen specialized tools with one general-purpose command-line tool and saw reliability and speed both improve. The cautionary counterexample is NVIDIA's perfect 100 on a hard reasoning benchmark against a separately reported 30% for the same model alone — NVIDIA itself warns against drawing the obvious conclusion, because too much else differed. The year's largest apparent harness win is the one nobody can cleanly attribute.
Some of the gain isn't even orchestration. SGLang v0.5.20 shipped with four sets of acceleration numbers carrying model names, batch sizes, and PR numbers. The sharpest: DeepSeek-V4-Flash on 4× RTX PRO 6000, decode TPOT falling from 36.1 ms to 10.5 ms at batch 1 after enabling DeepGEMM paged-MQA and FP4 MoE. A 3.4x speedup with zero bytes of the model changed. "Model performance" now has a quotable answer for how much of it is the model.
The protocol layer is bending the same way. MCP's Skills extension, merged September 13, solves a pure context-economics problem: server instructions are a bounded field, and the extension's own example is an 875-line skill that doesn't fit. So skills become resources fetched on demand via skill:// URIs rather than instructions pushed on every call. The main spec hasn't shipped since July 28; the extension track is where the design work is happening, and it's happening because context is scarce.
All of this makes the closed-to-open migration playbook look different than it did a year ago. Together's five-stage version — discover, evaluate, adapt, decide, production — leans on the fact that harnesses, gateways, and tooling are already built to adopt a variety of models, so the blast radius of swapping one is small. Its shortlisting advice is the tell: go past raw benchmark scores and look at cost per task, tokens burned per attempt, steps required, end-to-end runtime, and how quickly a correct answer can be verified. A model that scores a few points lower but finishes in half the tokens and half the time may be the better fit.
Two closing implications. First, developers demonstrably form loyalty to harnesses and none to models — they'll swap the backend behind Claude Code the moment a cheaper option clears the bar, and the feature gap between leading harnesses is reportedly closing in about eleven days. That asymmetry is a product lesson: own the contract with users, treat harness and model as parameters.
Second, the harness is now a place where value can leak out the other side. Two industrial distributors both bought the same integrator, the same frontier model, the same orchestration, and both cut quote turnaround from two days to two hours. Responsiveness stopped being a reason customers chose either one, buyers moved to price, and the margin both expected went to the customer. The program worked. The metric moved. The money left.