Development Approach

Custom Eval Harness vs Published Benchmark Scores: Which Should Decide Your Model?

GPT-5.6 Sol scored 7.8% and 38.3% on the same ARC-AGI-3 tasks. Only the harness changed. When to trust leaderboards, when to build your own eval.

4
Custom Eval Harness
vs
4
Published Benchmark Scores
Quick Verdict

Use published benchmarks to shortlist, use your own harness to decide, and never let the two swap roles. The July 2026 ARC-AGI-3 dispute is the cleanest evidence anyone has produced for this. OpenAI's position — "benchmarks never measure just the model but also the technical setup around it" — is correct. ARC Prize's position — that official scores must use a standardised approach without provider-specific settings so that cross-vendor comparisons stay honest — is also correct. François Chollet drew the line that actually resolves it: harnesses built to solve the benchmark, or that encode knowledge of its format, are off limits; general-purpose API features available to every customer are fair game, provided the settings and the cost are reported. That is a rule you can apply to your own evaluations tomorrow. For a buying decision, the practical consequence is a sequence, not a choice. Leaderboards eliminate candidates quickly and cost nothing: if a model is not in the conversation on public reasoning or coding benchmarks, no amount of scaffolding will rescue it. But once two or three candidates sit within a few points of each other, the published number stops discriminating and your harness starts. A roughly fivefold score swing from two API flags is larger than almost any gap between frontier models on a leaderboard, which means the scaffolding decision dominates the model decision at exactly the moment teams stop thinking about it. Cursor's engineering data points the same way from a completely different direction: cloud agents went from roughly one in ten merged pull requests in December to more than half by July, and the company credits the environment — giving agents their own machines and letting them repair their own setup — rather than a better model. Two independent 2026 data points, one conclusion: the system around the model is the variable that moved. There is one case where published scores should win outright, and it is not a technical one. If a number has to survive a board paper, a procurement process or a regulator, a self-reported internal result will not carry it. ARC Prize publishes its funding, its conflict-of-interest recusal rules and an independent academic panel precisely so its numbers hold up under that scrutiny. Your harness has none of that, and it should not pretend to. Our recommendation: shortlist on public benchmarks in an afternoon, then spend two weeks building a harness of 30 to 50 real tasks pulled from your own logs, with per-task token and cost logging and at least one scaffolding flag you can toggle. Rerun it on every model change and every loop change. If you only have budget for one of the two, build the harness — but publish the settings and the cost alongside the score, the same standard you would demand of a vendor.

Detailed Comparison

A side-by-side analysis of key factors to help you make the right choice.

Factor
Custom Eval HarnessRecommended
Published Benchmark ScoresWinner
What the number actually measures
The model plus your scaffolding as a single system: your prompts, your state handling, your tools, your retry policy.
The model under a provider-neutral setup that deliberately excludes vendor-specific API features.
Comparability across vendors
Only as fair as your own wrapper. OpenAI's harness used per-model settings, which is exactly what makes its 38.3 percent incomparable to a standardised run.
The entire point. ARC Prize runs every model through a standardised approach without provider-specific settings.
Predicts what you will actually ship
High. Sol moved from 7.8 to 38.3 percent on the same public tasks with no change to the weights, purely by retaining reasoning and compacting instead of truncating.
Low for agentic work. In the official harness the model's reasoning is discarded after every action step, which is not how anyone runs an agent in production.
Effort to get a usable number
Days to weeks: a task set, graders, budget instrumentation, and someone who keeps it honest as the codebase moves.
Minutes. Leaderboards are published, dated and free to read.
Reveals how much your loop matters
Directly. Two Responses API settings moved one model's score roughly fivefold and cut output tokens sixfold at the same time.
Invisible by design. A standardised harness holds scaffolding constant, so the largest single lever in your stack never appears in the number.
Auditability and conflict-of-interest control
Whatever you enforce yourself. Internal harnesses are rarely reviewed and easy to tune, consciously or not, toward the model the team already prefers.
Published policy. ARC Prize documents its funding, requires recusal from staff holding equity in a tested lab, and submits its methodology to an independent academic panel.
Reproducibility by a third party
Internal only. Nobody outside the team can rerun it, and a vendor claim measured this way stays self-reported — the 38.3 percent has had no independent private-set verification.
Reproducible by the publisher on a dataset and a publication cadence committed to in advance.
Cost visibility
Your real invoice on your real tasks, including retries and the tokens your loop wastes. Sol at 5 / 30 dollars per million and Opus 5 at 5 / 25 differ far less than the token counts two harnesses produce.
Published but generic. The ARC-AGI-3 leaderboard plots cost per task and only lists systems costing under 10,000 dollars to run — useful signal, not your bill.
Relevance to your workload
You choose the tasks, so the score is about your domain, your documents and your failure modes.
Abstract by construction. ARC-AGI-3 tests on-the-fly adaptation to novel interactive environments; the average human tester scores 48 percent.
Total Score4/ 94/ 91 ties
What the number actually measures
Custom Eval Harness
The model plus your scaffolding as a single system: your prompts, your state handling, your tools, your retry policy.
Published Benchmark Scores
The model under a provider-neutral setup that deliberately excludes vendor-specific API features.
Comparability across vendors
Custom Eval Harness
Only as fair as your own wrapper. OpenAI's harness used per-model settings, which is exactly what makes its 38.3 percent incomparable to a standardised run.
Published Benchmark Scores
The entire point. ARC Prize runs every model through a standardised approach without provider-specific settings.
Predicts what you will actually ship
Custom Eval Harness
High. Sol moved from 7.8 to 38.3 percent on the same public tasks with no change to the weights, purely by retaining reasoning and compacting instead of truncating.
Published Benchmark Scores
Low for agentic work. In the official harness the model's reasoning is discarded after every action step, which is not how anyone runs an agent in production.
Effort to get a usable number
Custom Eval Harness
Days to weeks: a task set, graders, budget instrumentation, and someone who keeps it honest as the codebase moves.
Published Benchmark Scores
Minutes. Leaderboards are published, dated and free to read.
Reveals how much your loop matters
Custom Eval Harness
Directly. Two Responses API settings moved one model's score roughly fivefold and cut output tokens sixfold at the same time.
Published Benchmark Scores
Invisible by design. A standardised harness holds scaffolding constant, so the largest single lever in your stack never appears in the number.
Auditability and conflict-of-interest control
Custom Eval Harness
Whatever you enforce yourself. Internal harnesses are rarely reviewed and easy to tune, consciously or not, toward the model the team already prefers.
Published Benchmark Scores
Published policy. ARC Prize documents its funding, requires recusal from staff holding equity in a tested lab, and submits its methodology to an independent academic panel.
Reproducibility by a third party
Custom Eval Harness
Internal only. Nobody outside the team can rerun it, and a vendor claim measured this way stays self-reported — the 38.3 percent has had no independent private-set verification.
Published Benchmark Scores
Reproducible by the publisher on a dataset and a publication cadence committed to in advance.
Cost visibility
Custom Eval Harness
Your real invoice on your real tasks, including retries and the tokens your loop wastes. Sol at 5 / 30 dollars per million and Opus 5 at 5 / 25 differ far less than the token counts two harnesses produce.
Published Benchmark Scores
Published but generic. The ARC-AGI-3 leaderboard plots cost per task and only lists systems costing under 10,000 dollars to run — useful signal, not your bill.
Relevance to your workload
Custom Eval Harness
You choose the tasks, so the score is about your domain, your documents and your failure modes.
Published Benchmark Scores
Abstract by construction. ARC-AGI-3 tests on-the-fly adaptation to novel interactive environments; the average human tester scores 48 percent.

Key Statistics

Real data from verified industry sources to support your decision.

GPT-5.6 Sol scored 7.8 percent on the ARC-AGI-3 public set in ARC Prize's official harness and 38.3 percent through OpenAI's Responses API with retained reasoning and compaction — identical weights, different harness.

MLQ.ai

Claude Opus 5 scored 30.2 percent on the same public task set under the standardised harness: ahead of Sol's official 7.8 percent, behind Sol's self-reported 38.3 percent.

THE DECODER

Retained reasoning and compaction tripled the score while reducing output tokens sixfold, according to OpenAI.

MLQ.ai

The average human tester scores 48 percent on ARC-AGI-3, so both frontier models remain below human performance under either harness.

MLQ.ai

At Cursor, cloud agents authored roughly one in ten merged pull requests in December and more than half by July 2026; the company credits the environment, not a better model.

Cursor Engineering Blog

The ARC-AGI-3 leaderboard plots cost per task against score and only lists systems that cost under 10,000 dollars to run.

ARC Prize

OpenRouter lists GPT-5.6 Sol at 5 / 30 dollars per million tokens and Claude Opus 5 at 5 / 25 — the harness decides how many of those tokens you actually spend.

OpenRouter Models API

All statistics come from verified third-party sources. Source, year, and direct link are shown on each metric.

When to Choose Each Option

Clear guidance based on your specific situation and needs.

Choose Custom Eval Harness when...

  • You are shipping an agent rather than a chat box: state handling, compaction and retries are part of your product, and a score measured without them says nothing about your latency or your bill.
  • Two candidate models sit within a few points of each other on public leaderboards — at that distance the harness is a bigger lever than the model.
  • Your domain is narrow (claims handling, contract review, internal tooling) and no public benchmark covers the work you actually do.
  • You need a defensible cost-per-task figure for a budget or a client, measured on your own tasks rather than on someone else's.

Choose Published Benchmark Scores when...

  • You are at the shortlist stage and need to eliminate clearly unsuitable models before spending engineering time on any of them.
  • You need a citable third-party number for a board paper, a procurement document or a regulatory filing — self-reported internal scores do not survive that scrutiny.
  • You are tracking generational progress over time rather than choosing which model to run this quarter.
  • Nobody on the team owns evaluation yet: an unmaintained internal harness produces worse decisions than a published leaderboard.

Our Recommendation

Use published benchmarks to shortlist, use your own harness to decide, and never let the two swap roles. The July 2026 ARC-AGI-3 dispute is the cleanest evidence anyone has produced for this. OpenAI's position — "benchmarks never measure just the model but also the technical setup around it" — is correct. ARC Prize's position — that official scores must use a standardised approach without provider-specific settings so that cross-vendor comparisons stay honest — is also correct. François Chollet drew the line that actually resolves it: harnesses built to solve the benchmark, or that encode knowledge of its format, are off limits; general-purpose API features available to every customer are fair game, provided the settings and the cost are reported. That is a rule you can apply to your own evaluations tomorrow. For a buying decision, the practical consequence is a sequence, not a choice. Leaderboards eliminate candidates quickly and cost nothing: if a model is not in the conversation on public reasoning or coding benchmarks, no amount of scaffolding will rescue it. But once two or three candidates sit within a few points of each other, the published number stops discriminating and your harness starts. A roughly fivefold score swing from two API flags is larger than almost any gap between frontier models on a leaderboard, which means the scaffolding decision dominates the model decision at exactly the moment teams stop thinking about it. Cursor's engineering data points the same way from a completely different direction: cloud agents went from roughly one in ten merged pull requests in December to more than half by July, and the company credits the environment — giving agents their own machines and letting them repair their own setup — rather than a better model. Two independent 2026 data points, one conclusion: the system around the model is the variable that moved. There is one case where published scores should win outright, and it is not a technical one. If a number has to survive a board paper, a procurement process or a regulator, a self-reported internal result will not carry it. ARC Prize publishes its funding, its conflict-of-interest recusal rules and an independent academic panel precisely so its numbers hold up under that scrutiny. Your harness has none of that, and it should not pretend to. Our recommendation: shortlist on public benchmarks in an afternoon, then spend two weeks building a harness of 30 to 50 real tasks pulled from your own logs, with per-task token and cost logging and at least one scaffolding flag you can toggle. Rerun it on every model change and every loop change. If you only have budget for one of the two, build the harness — but publish the settings and the cost alongside the score, the same standard you would demand of a vendor.

Frequently Asked Questions

Common questions about this comparison answered.

No. They measure something specific and narrow, and they are honest about it: the model under a standardised setup without provider-specific settings, so that numbers from different vendors can sit in the same table. François Chollet drew the useful line in the July 2026 exchange — harnesses custom-made to solve a benchmark, or that encode knowledge of its format, are off limits, while general-purpose API features available to every customer are acceptable as long as the settings and the cost are clearly reported. The error is not reading leaderboards; it is treating a leaderboard number as a forecast of your production behaviour.
More than the gap between frontier models. On the ARC-AGI-3 public set the same GPT-5.6 Sol weights produced 7.8 percent in the official harness and 38.3 percent through OpenAI's Responses API with retained reasoning and compaction — roughly a fivefold swing, and with six times fewer output tokens rather than more. For comparison, Claude Opus 5's official score on that set is 30.2 percent. If your scaffolding discards reasoning between steps or hard-truncates context when the window fills, you are running your model with the parking brake on.
In sequence, both, and the order matters. Use public benchmarks to cut a long list to two or three candidates in an afternoon — that is what they are good at, and it costs nothing. Then decide between the survivors with your own harness, because at that distance the published numbers no longer discriminate. If you can only fund one, fund the harness: it is the only one of the two that measures the thing you are actually paying for.
Thirty to fifty real tasks pulled from your own logs rather than invented ones, a grader you trust (human review is acceptable at this size), per-task logging of input tokens, output tokens and wall-clock cost, and at least one scaffolding flag you can toggle — state retention between steps is the one that moves numbers most. Rerun it on every model change and every change to the loop, and record the settings alongside the score. That last step is what separates an evaluation from an anecdote, and it is the same standard the ARC Prize dispute concluded vendors should be held to.

Need help deciding?

Book a free 30-minute consultation and we'll help you determine the best approach for your specific project.

Free consultation
No obligation
Response within 24h