Provider
OpenAI
Base model family
Model intelligence profile
GPT-5.4 benchmark results and model details
This family profile brings together 7 published benchmark records for OpenAI GPT-5.4. Every result keeps its original benchmark, configuration, scale, and source; unrelated scores are never averaged.
Published benchmark evidence available · Business Skills V3 not yet authorized
Release date
2026-03-05
Exact-family source only
Published benchmarks
7
Native scales kept separate
Artificial Analysis
Reference available
As of 2026-07-16
Benchmark-level evidence
Published GPT-5.4 benchmark results
These are individual benchmarks, not collection rollups. Scores remain on their original scales, and agent or harness results stay labelled as system configurations.
Spring Prompt benchmark
PredictTheWeek pilot
26.5%
Mean prediction score
Can LLMs anticipate next week’s Guardian agenda from last week’s coverage?
This benchmark is not on a weekly live cadence yet. The table and charts reflect a single multi-model comparison on one forecast window—a pilot snapshot, not an updating leaderboard. The evaluation setup (clustering, prompts, automated judge, and scoring checks) is still being refined; reported scores and details may change as we improve the pipeline.
1 published configuration
GPT-5.4
26.5%
- Accuracy (partial support or better)
- 33.0%
- Outcome-line coverage
- 9.1%
- Pilot composite
- 17.5%
Spring Prompt benchmark
ROASBench
18.39
Average ROASBench score
A 12-month performance-marketing simulation scored on business outcomes, planning, behavior, and persona fit.
1 published configuration
OpenAI: GPT-5.4
18.39
- Average contribution profit
- $-250,461
- Average ROAS
- 1.03×
- Completed runs
- 3
- Score variability
- ±1.54
Source: Spring Prompt
Recorded 25 Jul 2026
Published ROASBench cache · 12-month simulation
Source record →
source-native benchmark
Coding Agent Configurations
2 configurations
Coding Agent Index v1.2 point score
Artificial Analysis Coding Agent Index v1.2 point scores and task-specific operational measurements for exact agent and model configurations.
These are exact agent-plus-model configurations, not bare-model scores.
2 published configurations
Codex - GPT-5.4 (medium)
41.1
- Mean cost per task
- $2.27
- Mean wall time
- 7.1 min
Cursor CLI - GPT-5.4 (medium)
40.1
- Mean cost per task
- $1.52
- Mean wall time
- 8.3 min
source-native benchmark
Repository Issue Resolution
75.6%
OpenHands SWE-Bench resolved
Direct OpenHands SWE-Bench outcomes for resolving real repository issues with a pinned OpenHands agent and language-model configuration.
Scores are official OpenHands SWE-Bench resolved percentages for exact OpenHands and model configurations, not bare-model scores or SpringPrompt predictions. Twenty-nine exact task sidecars support marginal 95% bootstrap score intervals; five rows explicitly have no interval. No full-cohort rank confidence is claimed.
1 published configuration
OpenHands v1.18.1 + GPT-5.4
75.6%
- 95% task-bootstrap interval
- 71.8–79.4%
- Task evidence
- exact-headline-reconstruction
Source: OpenHands Index
Recorded 30 Jun 2026
OpenHands · SWE-Bench 2026.06.30-3015ac6
Source record →
source-native benchmark
STATE-Bench v0.7.main
55.7%
Overall pass@1
Source-native STATE-Bench completion, UX and cost results for stateful enterprise workflows.
A harnessed-model workflow result.
1 published configuration
GPT-5.4 · high
55.7%
- Mean UX
- 3.49
- Cost per task
- $0.0809
source-native benchmark
STATE-Bench v0.8.main
58.6%
Overall pass@1
Source-native STATE-Bench completion, UX and cost results for stateful enterprise workflows.
A harnessed-model workflow result.
1 published configuration
GPT-5.4 · high
58.6%
- Mean UX
- 3.67
- Cost per task
- $0.0874
source-native benchmark
Structured Output Reliability
87.02%
Direct benchmark score
Directly measured ability to return accurate values in the requested structured schema across the benchmark's evaluated text, image and audio modalities.
Point order reproduces the source's direct Overall score. Rank ranges come from overlap of marginal record-cluster bootstrap intervals for Overall; they are not simultaneous confidence intervals for rank.
1 published configuration
GPT-5.4
87.02%
- 95% interval
- 86.62–87.40
- Evaluated modality coverage
- 100%
Source: Structured Output Benchmark (SOB)
Recorded 17 Jul 2026
Sob-v1@da785a8521c8954283b2989d01e54d80c4e023c6:upstream-provider-configs:temperature-0-where-supported:max-output-2048:reasoning-disabled-or-minimum-where-required:official-modality-weights
Source record →
Independent operational reference
Artificial Analysis details
Representative configuration: GPT-5.4 (xhigh). Lifecycle: deprecated; released 2026-03-05.
Input price
$2.50 / 1M tokens
Output price
$15.00 / 1M tokens
Median output speed
150.3 tok/s
Operational reference only. Price and throughput are not model-quality scores and never affect Spring Prompt comparisons.
Early-user signal
Initial community opinions
A paraphrased editorial synthesis of 4 Reddit discussions created in the 14 days after launch (5 Mar 2026 to 19 Mar 2026). Anecdotal context only — it never affects scores or rankings.
In the search-indexed launch-window sample, GPT-5.4 drew positive reports for complex coding, self-correction, writing, and customization, though the perceived change depended on the surface and task. Some users retained older Codex or GPT variants for smaller work because they found them faster, cheaper, or more predictable, and the unified model naming caused early confusion.
Independent early-test signal
Early technical field tests
An editorial paraphrase of 2 launch-window field tests from 2 independent authors on X, including 2 reports with a described method or inspectable artifact. The window runs from 5 Mar 2026 up to 19 Mar 2026, and exact model identity was editor reviewed.
Early GPT-5.4 evidence showed a marked improvement on a professional-agent benchmark, while a concrete distillation workflow exposed a different tradeoff: fast data generation but disappointing teacher-data quality and rapid quota consumption. The combined picture is stronger agent capability without a universal workflow conclusion; output quality, quotas, and task framing still matter.
These reports are selectively surfaced and are not a representative sample. They never affect benchmark scores, rankings, winners, or comparison outcomes.
Business Skills V3 · proposed
How Spring Prompt plans to test GPT-5.4
The setup below is proposed and may change until preflight and execution approval are complete. Existing benchmark evidence above does not authorize or stand in for a Business Skills V3 result.
- Proposed reasoning
- High reasoning
- Provider revision
- Exact provider revision will be resolved and frozen only after preflight and execution approval
- Tools and service
- No tools · Provider-default service tier
- Planning configuration
- gpt-5.4-high
- Generation controls
- Provider-managed reasoning; no temperature override proposed
Future first-party coverage
15 proposed Business Skills V3 task areas
These links describe evaluation contracts, not published GPT-5.4 results.
Professional Knowledge Work & StrategyBusiness Workflow AutomationCustomer Support & Service ResolutionSales & MarketingLegal Document WorkFinance, Accounting & SpreadsheetsData Analysis & SQLResearch & Evidence SynthesisBusiness Writing & EmailDocument Comprehension & ExtractionSummarization & Grounded KnowledgeSoftware Engineering & CodingFrontend, UI & Web AppsTool Calling & Structured OutputPresentations, Charts & Visual Communication
Model comparisons
Compare GPT-5.4 side by side
Editorially reviewed comparisons appear first. Every page matches only benchmark records with the same reviewed protocol key and does not manufacture an overall winner.
Publication safeguards
What must pass before a V3 result appears
- 1First-party task-local comparisons and eligible external evidence must both be present.
- 2Model and provider configuration identity must match the reviewed release exactly.
- 3Coverage, reliability, judge diagnostics, and sealed stability checks must pass.
- 4Uncertainty and missing evidence remain visible when results are published.
Stable family URL
Evidence can grow without changing the page
New reviewed benchmark snapshots, operational facts, community themes, and early field-test syntheses can be added here while the canonical model-family identity remains fixed.
More from OpenAI