ModelSpec is an analysis of alternatives for AI models: it decides which models fit a job from your requirements, sourced benchmarks and real cost, and shows its work.

GPT-5.4 vs GPT-6 Astra for Maths and proofs, as of 2026-09-23

Find a text model with the strongest evidence on mathematical problems and proofs.

Worked example as of 2026-09-23, from the public data image.

What the evidence says

The evidence separates them: GPT-6 Astra leads for Maths and proofs, with a 25% chance of being the best model in the full lineup for this job, as of 2026-09-23.

GPT-5.4 has a 21% chance of matching or beating GPT-6 Astra on this job's objective.

probability bands: a model joins the leader when P(its score >= the leader's) >= 0.25 under the score posteriors; a weighted capability with an 80% interval wider than 2.8 is not enough evidence

Decision status: partial. P(best) and score intervals retain the full ranked lineup for this job, including models beyond this pair and models with thin evidence. Models with unknown objective values remain unranked and are outside this probability. P(best) is not a head-to-head probability. Separation from this challenger need not mean a unique best model across that lineup.

Scores and score intervals use the objective's dimensionless normalised units. Cost per task is in USD for the selected offering and the applied spec's task size. When task size is omitted, the engine uses its documented default. Unavailable means the engine returned no value.

Model	Score	Score interval	P(best), ranked job lineup	Cost per task, USD	Selected offering: provider / region / tier	Capability estimates and intervals
GPT-5.4	0.776654	[0.523558, 1.02975]	7%	0.16	azure-ai-foundry / global-short-context / standard	maths: 0.925643173; interval [0.204892258, 1.646394088]; benchmarks 5; direct benchmarks 2
GPT-6 Astra	1.0	[0.747794, 1.252206]	25%	0.6	azure-ai-foundry / global-short-context / standard	maths: 1.561674136; interval [0.843459177, 2.279889095]; benchmarks 4; direct benchmarks 2

Strongest maths evidence, cost no object.

Requirements

model.class = text-generator: The answer must be generated as mathematical text or a proof.

Preferences

maths: Rank only by the maths capability estimate.

Applied template spec

spec_version

where

model.class = text-generator

optimize

weights

maths: 1.0

Evidence for GPT-5.4

Benchmark	Version	Measurement	Unit	Published interval	Evidence date	Measurement conditions	Provenance
aime_2026	AIME 2026, MathArena final-answer table	99.17	percent	Unavailable	2026-09-30	Effort: xhigh Subcategory: Unavailable Harness: Unavailable Measured by: independent_evaluator Sample size: Unavailable	Source
arena_sc_industry_mathematical	text_style_control / industry_mathematical, latest split, revision 1880dbebff5b	1499.48	Arena score (Elo scale)	Unavailable	2026-09-13	Effort: Unavailable Subcategory: Unavailable Harness: Unavailable Measured by: independent_evaluator Sample size: Unavailable	Source
arena_sc_math	text_style_control / math, latest split, revision 1880dbebff5b	1494.13	Arena score (Elo scale)	Unavailable	2026-09-13	Effort: Unavailable Subcategory: Unavailable Harness: Unavailable Measured by: independent_evaluator Sample size: Unavailable	Source
frontiermath_tiers_1_3_v2	FrontierMath-Tiers-1-3-v2-Private (Epoch AI run)	78.6	percent	Unavailable	2026-06-11	Effort: xhigh Subcategory: Unavailable Harness: Unavailable Measured by: independent_evaluator Sample size: Unavailable	Source
hle	Humanity's Last Exam, Scale Labs leaderboard	36.24	percent	Unavailable	2026-09-30	Effort: Unavailable Subcategory: Unavailable Harness: Unavailable Measured by: independent_evaluator Sample size: Unavailable	Source

Missing facts for GPT-5.4

model.hardware_fit_indeterminate: unknown
model.languages: unknown
model.parameters_total: unknown
origin.base_lineage: not_disclosed
origin.weights_hosting: not_disclosed
offering.api_model_id: unknown
offering.attestation.baa: not_disclosed
offering.attestation.soc2: not_disclosed
offering.data.retention: not_disclosed
offering.data.zero_retention: requires_contract
offering.router_slug: unknown
offering.speed.throughput: unknown
offering.speed.time_to_first_token: unknown

Evidence for GPT-6 Astra

Benchmark	Version	Measurement	Unit	Published interval	Evidence date	Measurement conditions	Provenance
arxivmath	ArXivMath, MathArena Overall table	92.68	percent	Unavailable	2026-09-30	Effort: max Subcategory: Unavailable Harness: Unavailable Measured by: benchmark_author Sample size: Unavailable	Source
brokenarxiv	BrokenArXiv, MathArena Overall table	90.67	percent	Unavailable	2026-09-30	Effort: max Subcategory: Unavailable Harness: Unavailable Measured by: benchmark_author Sample size: Unavailable	Source
frontiermath_tiers_1_3_v2	FrontierMath-Tiers-1-3-v2-Private (Epoch AI run)	93.68	percent	Unavailable	2026-08-30	Effort: max Subcategory: Unavailable Harness: Unavailable Measured by: independent_evaluator Sample size: Unavailable	Source
hle	Humanity's Last Exam, Scale Labs leaderboard	54.8	percent	Unavailable	2026-09-30	Effort: Unavailable Subcategory: Unavailable Harness: Unavailable Measured by: independent_evaluator Sample size: Unavailable	Source
hle_tools	Humanity's Last Exam (w/ tools)	57.2	percent	Unavailable	2026-09-03	Effort: Unavailable Subcategory: Unavailable Harness: Unavailable Measured by: provider_self_report Sample size: Unavailable	Source

Missing facts for GPT-6 Astra

model.hardware_fit_indeterminate: unknown
model.languages: unknown
model.parameters_total: unknown
origin.base_lineage: not_disclosed
origin.weights_hosting: not_disclosed
offering.api_model_id: unknown
offering.attestation.baa: not_disclosed
offering.attestation.soc2: not_disclosed
offering.data.retention: not_disclosed
offering.data.zero_retention: requires_contract
offering.price.batch_input: not_disclosed
offering.price.batch_output: not_disclosed
offering.router_slug: unknown
offering.speed.throughput: unknown
offering.speed.time_to_first_token: unknown

What this template teaches

Use one Must to select the right model class and one Prefer to state what best means.

Decision dec_25456c28ff3379f3b08db308 on public snapshot snap_05a963d2477ac2fb.

Read the full dated job answer.

The live board uses current data. This link applies the template to its full lineup.

Apply this template on Decide

Human lookup is free on Decide. CLI, API, and MCP access requires a key. Access and pricing.

Use-case examples · Model comparisons · How we decide