Benchmark

    How well AI models write HEOR: measured, not asserted.

    General model leaderboards measure reasoning, code and trivia. None of them measure whether a model can write a defensible Consulting, HEOR or market-access deliverable. We built that measurement for our own model selection, published it, and keep it current: the preprint is fixed at publication, this page tracks what is actually available.

    Last updated 2026-09-17.

    Results

    Unanchored AI-Delphi writing scores on HEOR and market-access tasks. A model tested with the Untraceable Grounding Codex appears as its own row, so the Codex's effect is visible rather than blended into the model's score.

    Model aloneWith the Untraceable Grounding Codex
    1Claude Opus 4.8 + Codex
    79.1
    2GPT 5.5 + Codex
    78.5
    3Claude Opus 4.7
    77.6
    4GPT 5.5
    76.8
    5Kimi K2.6
    76.5
    6Claude Sonnet 4.6 + Codex
    76.3
    7Claude Sonnet 4.6
    75
    8Claude Opus 4.8
    74.4
    9GPT 5.4 Mini
    74.1
    10Claude Haiku 4.5
    73
    11Claude Haiku 4.5 + Codex
    72.6
    12GPT 5.4 Mini + Codex
    72.5
    13DeepSeek V4 Pro
    67.4
    6080
    Unanchored AI-Delphi composite score, measured on a 0–100 scale. The axis is truncated to 60–80 so the differences between frontier models stay readable. The bar length therefore compares models against each other, not against a perfect score. Every bar carries its own value.
    RankModelProviderGrounding CodexDate testedScore
    1Claude Opus 4.8AnthropicYes2026-06-1479.1
    2GPT 5.5RetiredOpenAIYes2026-06-1478.5
    3Claude Opus 4.7AnthropicNo2026-06-1477.6
    4GPT 5.5RetiredWithdrawn by the provider after this round; superseded by GPT-5.6.OpenAINo2026-06-1476.8
    5Kimi K2.6Moonshot AINo2026-06-1476.5
    6Claude Sonnet 4.6AnthropicYes2026-06-1476.3
    7Claude Sonnet 4.6AnthropicNo2026-06-1475
    8Claude Opus 4.8AnthropicNo2026-06-1474.4
    9GPT 5.4 MiniOpenAINo2026-06-1474.1
    10Claude Haiku 4.5AnthropicNo2026-06-1473
    11Claude Haiku 4.5AnthropicYes2026-06-1472.6
    12GPT 5.4 MiniOpenAIYes2026-06-1472.5
    13DeepSeek V4 ProDeepSeekNo2026-06-1467.4

    What the numbers show

    5 models were scored both ways. The Codex raised 3 and lowered 2, the largest single gain was +4.7 on Claude Opus 4.8. The gains land on the strongest models and the losses on the smallest: a retrieved knowledge layer is only worth what the model can do with it, and a lighter model might spend capacity reading the context instead of reasoning over it. Grounding is not a uniform uplift, it seems to work very well with frontier models, but might need improvement to work with all models.

    Why we built this

    Language models are built on language, and nobody measures the language.

    Every general leaderboard measures reasoning, code, mathematics and trivia. The one capability these models were built around: writing. We have not found a benchmark evaluating it.

    Regulated writing is a different task, and it is untested.

    A HEOR dossier or a CSR, a regulatory submission, a legal memo, a financial disclosure: each carries conventions, evidentiary standards and a reader who will challenge it. No public benchmark measures whether a model can write to that standard. This is true for pharma, healthcare, legal, finance, consulting or any other regulated field.

    So we built the measurement.

    A controlled task set, scored by an independent AI-Delphi panel, run on each model alone and again with the field-specific Grounding Codex we develop at Untraceable. That makes two separate questions answerable: how well does this model write in specialized fields, and does a domain knowledge layer actually help it.

    And we publish it

    Models change under the same name, so a benchmark published once is a benchmark that is already wrong. We re-run it and publish the results, including for models we do not offer, so users, colleagues and experts in our fields can judge for themselves which model to trust with complex work. Inside Untraceable you get those models, with the Codex on top if you want it, and the AI still never sees a real value. In the meantime, the platform only provides high quality and high security AI access accessible within our Azure interface, so not all models are available on the platform (see available models here).

    Method

    • Every model writes the same controlled set of HEOR and market-access tasks, and the task set from the writing benchmark remains unchanged between models.
    • Scoring is by AI Delphi: several independent models score the same output over structured rounds, and their judgments converge. No single model grades the field.
    • These are the UNANCHORED results. The publication also reports human-anchored scoring; this page publishes the unanchored panel only, so every row is comparable to every other row.
    • Runs marked with the Untraceable Grounding Codex are given a curated domain knowledge layer: HTA guidance, methods conventions and reporting standards. The same model without it is a separate row, never averaged in.
    • Each row records the provider's exact version string and the date of the run. A model that changes under the same name is re-tested rather than edited.
    • Results are published whichever way they go, including for models we do not offer, and retired models stay in the table because the publication cites them.
    • These results drive which models we offer, and for which tool. A model that scores poorly on a task, or that the Grounding Codex does not help, is not put in front of a client for that task.

    The published method

    In testing

    Models released after the publication, queued for the next round. We publish the result whichever way it goes.

    • GPT-6 AstraOpenAI
    • Fable 5.1Anthropic
    • Claude Sonnet 5Anthropic
    • GPT-5.6 SolOpenAI, Released after the preprint, which measures GPT-5.5.
    • GPT-5.6 TerraOpenAI, Released after the preprint, which measures GPT-5.5.

    Questions about the benchmark

    Which AI model writes HEOR best?
    In the most recent round, Claude Opus 4.8 scored highest at 79.1 out of 100 when run with the Untraceable Grounding Codex, and Claude Opus 4.7 was the strongest model on its own at 77.6. Scores are unanchored AI-Delphi composites on a controlled set of HEOR and market-access writing tasks. The full ranking, the exact dates and the method are on this page, updated as new models ship: last updated 2026-09-17.
    Which AI model is best for pharma, medical writing and other regulated industries?
    Our measured task set today is HEOR and market access, and we are explicit about that rather than claiming more than we tested. We are expanding it to adjacent regulated writing, pharma and medical writing, regulatory affairs, biostatistics, and further into legal and finance. In the meantime we consider these results a reasonable guide for regulated professional writing generally: the demands are closely related: evidentiary standards, strict conventions, a reader who will challenge the document, and a cost of error that is legal rather than stylistic. So a model that writes a defensible HEOR deliverable is doing the work that regulated writing asks for. Treat it as a strong indication outside HEOR, and as measurement inside it.
    Why is there no general benchmark for AI writing quality?
    Public leaderboards measure reasoning, code, mathematics and factual recall. Writing, the capability these models were built around, is largely assumed rather than tested, and professional writing in regulated fields is not tested at all. We built this measurement because we needed it to choose the models inside Untraceable, then published it.
    Does a grounding or retrieval layer make an AI model better at regulated writing?
    Not uniformly. Of the 5 models scored both with and without the Untraceable Grounding Codex, it raised 3 and lowered 2, with the largest single gain being +4.7 on Claude Opus 4.8. The gains land on the strongest models and the losses on the smallest: a retrieved knowledge layer is only worth what the model can do with it, and a lighter model spends capacity reading the context instead of reasoning over it. We publish that result the way it measured.
    How is this benchmark scored, and who scores it?
    Every model writes the same controlled task set, and the output is scored by an AI-Delphi panel: several independent models score the same text over structured rounds and converge. No single model grades the field. These are the unanchored results only, the publication also reports human-anchored scoring, and the two are never mixed in one table.
    Can I use the best-scoring model inside Untraceable without exposing confidential data?
    Yes, that is the reason the benchmark exists. Untraceable cloaks every confidential value on your own computer before any model sees the text, so you can run the highest-scoring frontier model, with the Grounding Codex on top, while the model receives only cloaks and never a real value. These results are what decide which models we offer, and for which tool.