Tech Meridian← ALL MODELS
PROMY MERIDIAN

OPENAI · MODEL RELEASE TRACKER

GPT-5

GPT-5 is a version of the GPT family that was evaluated on the XYBench benchmark and was included in OpenAI's analysis of model monitorability between GPT-5 and GPT-6.

CURRENT SNAPSHOT2/5 DIMENSIONS WITH DATA

The dimensions that change the decision.

PRICING
Estimated cost per complex caseAbout $8,400 per case when running GPT-5 on its own with a short, basic prompt (estimated cost in the Microsoft simulation).
DxO effect on cost (complex cases)When paired with Microsoft’s DxO, GPT-5’s estimated cost per complex case fell by about $3,000.
CONTEXT WINDOW

Not established from the available sources.

MODALITIES

Not established from the available sources.

BENCHMARKS
Included in benchmark study of theoretical problemsGPT-5 was among the OpenAI models (o1, o3, GPT-5, GPT-5.1, GPT-5.4, and GPT-5.5) evaluated on problems drawn from research papers posted between December 2025 and May 2026, covering fields including probability, statistics theory, economic theory, high-energy theory, gravitation and cosmology, and quantum physics.
Success rate in executable planning for multi-component CPS using GPT-5Reported 87.9% success rate on seven heterogeneous multi-component cyber-physical system scenarios when using GPT-5 with the authors' hierarchical feedback-guided planning framework.
Accuracy on complex NEJM cases (simulation)GPT-5 scored 78.9% correct on complex cases in the simulation when used on its own with a short, basic prompt.
DxO effect on accuracy (complex cases)When paired with DxO, GPT-5’s accuracy on complex cases rose by about four percentage points.
Benchmark: HandMeThat v2 (FISER vs CEMIR)GPT-5 was evaluated in a comparative study of FISER and CEMIR on the HandMeThat v2 benchmark; the authors report no statistically significant difference in performance between FISER and CEMIR for GPT-5 across any difficulty level.
Speedup cap in review‑every‑output workflowIn GDPval's review‑every‑output workflow, review time alone caps GPT-5's speedup at 3.6×.
Reward-hacking detection (at 5% false-alarm)On METR's MALT evaluation (at a 5% false-alarm rate), GPT-5 caught nearly 98% of reward-hacking cases.
Included in METR's MALT evaluationGPT-5 (and GPT-5 nano) were examined by METR in analyses using the MALT dataset.
Performance on XYBench (misconception handling)Evaluated on the XYBench benchmark; GPT-5 often suggested a solution but tended to remain focused on the user's original, mistaken approach and was less likely than human reference answers to steer users toward an alternative solution that better addressed their actual goal.
AVAILABILITY

Not established from the available sources.

VERIFIABLE FACTS

Every value stays attached to a source and date.

BENCHMARKS · Included in benchmark study of theoretical problemsDEVELOPER CLAIM

GPT-5 was among the OpenAI models (o1, o3, GPT-5, GPT-5.1, GPT-5.4, and GPT-5.5) evaluated on problems drawn from research papers posted between December 2025 and May 2026, covering fields including probability, statistics theory, economic theory, high-energy theory, gravitation and cosmology, and quantum physics.

LIMITATIONS · Test used solutions published after models' training cutoff; internet access blockedDEVELOPER CLAIM

The study used solutions that were published after the training cutoff for GPT-5.5 and preceding models, meaning GPT-5 and the other tested models could not rely on those solutions as training data; researchers also blocked internet access during testing so models could not look up solutions online.

BENCHMARKS · Performance on XYBench (misconception handling)DEVELOPER CLAIM

Evaluated on the XYBench benchmark; GPT-5 often suggested a solution but tended to remain focused on the user's original, mistaken approach and was less likely than human reference answers to steer users toward an alternative solution that better addressed their actual goal.

SAFETY · Included in OpenAI monitorability analysisDEVELOPER CLAIM

GPT-5 was included in OpenAI's analysis of model monitorability (the study covered models from GPT-5 through GPT-6 Astra); researchers tracked early warning signs such as silent reasoning without a scratchpad (No-CoT Capability) and targeted control over chains of thought (CoT Controllability).

WHAT CHANGED

Stored passport versions, without reconstructed history.

Passport updated18 facts
SAFETY · Included in OpenAI monitorability analysis+ GPT-5 was included in OpenAI's analysis of model monitorability (the study covered models from GPT-5 through GPT-6 Astra); researchers tracked early warning signs such as silent reasoning without a scratchpad (No-CoT Capability) and targeted control over chains of thought (CoT Controllability).
BENCHMARKS · Performance on XYBench (misconception handling)+ Evaluated on the XYBench benchmark; GPT-5 often suggested a solution but tended to remain focused on the user's original, mistaken approach and was less likely than human reference answers to steer users toward an alternative solution that better addressed their actual goal.
Passport updated16 facts
BENCHMARKS · Included in METR's MALT evaluation+ GPT-5 (and GPT-5 nano) were examined by METR in analyses using the MALT dataset.
BENCHMARKS · Reward-hacking detection (at 5% false-alarm)+ On METR's MALT evaluation (at a 5% false-alarm rate), GPT-5 caught nearly 98% of reward-hacking cases.
Passport updated14 facts
BENCHMARKS · Speedup cap in review‑every‑output workflow+ In GDPval's review‑every‑output workflow, review time alone caps GPT-5's speedup at 3.6×.
LIMITATIONS · Training‑data cutoff precedes tested SEC filings+ The paper reports the experiment used briefs built from SEC filings that were made after the model's training data were collected (the filings postdate the model's training‑data collection).
Passport updated12 facts
BENCHMARKS · Benchmark: HandMeThat v2 (FISER vs CEMIR)+ GPT-5 was evaluated in a comparative study of FISER and CEMIR on the HandMeThat v2 benchmark; the authors report no statistically significant difference in performance between FISER and CEMIR for GPT-5 across any difficulty level.
Passport updated11 facts
PRICING · Estimated cost per complex case+ About $8,400 per case when running GPT-5 on its own with a short, basic prompt (estimated cost in the Microsoft simulation).
BENCHMARKS · DxO effect on accuracy (complex cases)+ When paired with DxO, GPT-5’s accuracy on complex cases rose by about four percentage points.
PRICING · DxO effect on cost (complex cases)+ When paired with Microsoft’s DxO, GPT-5’s estimated cost per complex case fell by about $3,000.
LIMITATIONS · Limited improvement on emergency department cases+ On emergency department cases, there was little improvement in either accuracy or cost between GPT-5 on its own and GPT-5 paired with DxO.
CAPABILITIES · Among top-performing models in MAI-DxO tests+ Microsoft reported that MAI-DxO achieved its best results when using OpenAI’s GPT-5 (and GPT-5.5) among the models it tested.
BENCHMARKS · Accuracy on complex NEJM cases (simulation)+ GPT-5 scored 78.9% correct on complex cases in the simulation when used on its own with a short, basic prompt.
Passport updated5 facts
BENCHMARKS · Success rate in executable planning for multi-component CPS using GPT-5+ Reported 87.9% success rate on seven heterogeneous multi-component cyber-physical system scenarios when using GPT-5 with the authors' hierarchical feedback-guided planning framework.
Passport updated4 facts
BENCHMARKS · Included in benchmark study of theoretical problems+ GPT-5 was among the OpenAI models (o1, o3, GPT-5, GPT-5.1, GPT-5.4, and GPT-5.5) evaluated on problems drawn from research papers posted between December 2025 and May 2026, covering fields including probability, statistics theory, economic theory, high-energy theory, gravitation and cosmology, and quantum physics.
LIMITATIONS · Test used solutions published after models' training cutoff; internet access blocked+ The study used solutions that were published after the training cutoff for GPT-5.5 and preceding models, meaning GPT-5 and the other tested models could not rely on those solutions as training data; researchers also blocked internet access during testing so models could not look up solutions online.
Passport created2 facts