What this experiment measured
This report records a translation-service investigation dated April 28, 2026. After a service moved from GPT-4o to GPT-5, response time reportedly more than doubled, putting an April 30 launch at risk. The investigation compared time to first token (TTFT), total response time, output-token counts, and tokens per second; explored alternatives for a future GPT-4o migration; and tested whether a lower reasoning setting made GPT-5 more suitable for translation.
These are historical measurements from one small application and workload, not Azure performance guarantees, a current model catalog, a retirement-date verification, or a formal translation-quality benchmark. The report's interpretation of reasoning overhead is a hypothesis; detailed reasoning-token telemetry was not supplied.
Deployment and application environment
The application ran on Azure App Service B1 for Linux in East US 2, using Node.js 20, Express.js, and OpenAI SDK v4. Its system-assigned managed identity held the Cognitive Services OpenAI User role on the Azure OpenAI resource. Azure Policy blocked API-key authentication. Actual application, resource, resource-group, and endpoint identifiers have been removed; use YOUR_APP_NAME, YOUR_AZURE_OPENAI_RESOURCE, and YOUR_RESOURCE_GROUP in your own environment.
| Model | Version | Deployment type | Allocated TPM | Setting used in six-model comparison |
|---|---|---|---|---|
| GPT-4o | 2024-11-20 | Global Standard | 30,000 | Defaults |
| GPT-4o-mini | 2024-07-18 | Global Standard | 30,000 | Defaults |
| GPT-4.1 | 2025-04-14 | Global Standard | 30,000 | Defaults |
| GPT-5 | 2025-08-07 | Global Standard | 30,000 | reasoning_effort: "low" |
| GPT-5.4-mini | 2026-03-17 | Global Standard | 30,000 | Defaults |
| GPT-5.4-nano | 2026-03-17 | Global Standard | 30,000 | Defaults |
Deployment names are intentionally omitted: Azure deployment aliases are environment-specific and should not be confused with model IDs. The report states that the tested GPT-5 version allowed only the default temperature of 1; reasoning_effort was the tuning parameter changed. It does not provide a controlled before-and-after tuning experiment with otherwise identical settings.
Application behavior
- One text submission started six model calls concurrently with Promise.allSettled.
- A result card for each model displayed its translation, TTFT, total duration, and tokens per second.
- The fastest result received a FASTEST label, and a history table recorded comparisons.
- No application source, deployment package, raw result export, or full request implementation accompanies the report. This page therefore describes the application rather than presenting invented runnable code.
Complete common system prompt
Every model used the following prompt. It is the complete inline prompt provided in the report, not an independently validated translation specification.
You are a professional Korean-English translator.
Rules:
- Translate the given text accurately while preserving the original meaning and tone.
- If the input is Korean, translate to English. If English, translate to Korean.
- Maintain proper grammar, natural phrasing, and cultural context.
- Do not add explanations or commentary. Return only the translated text.
- Preserve formatting (bullet points, line breaks, etc.) from the original.Eight translation scenarios
The original report supplied abbreviated descriptions rather than the complete test corpus. The English descriptions below retain their purpose and translation direction without claiming to reconstruct the missing prompts.
| ID | Scenario | Input description or excerpt | Direction |
|---|---|---|---|
| 1 | Simple statement | A statement beginning with Seoul being the capital of South Korea | Korean to English |
| 2 | Everyday conversation | An invitation to have coffee after a meeting | Korean to English |
| 3 | Technical documentation | A description of Azure Kubernetes Service | Korean to English |
| 4 | Business email | A greeting to a manager followed by a cloud-migration discussion | Korean to English |
| 5 | Technical blog | When deploying AI models at scale... | English to Korean |
| 6 | News article | Microsoft announced today that Azure AI Foundry... | English to Korean |
| 7 | Long-input stress case | Three paragraphs about worldwide advances in artificial intelligence | Korean to English |
| 8 | Idiomatic language | A description of someone who reads social cues and the atmosphere quickly | Korean to English |
Round 1: initial GPT-4o, GPT-5, and GPT-5-mini comparison
This initial round is distinct from the final six-model deployment list: it includes GPT-5-mini, not GPT-5.4-mini. Its GPT-5-mini version, deployment configuration, and complete reasoning settings are not specified. Only scenarios 1, 2, 3, 4, and 7 have rows in the report. No results for the other scenarios should be inferred.
| Scenario | Model | TTFT (ms) | Total (ms) | Output tokens | Tokens/s |
|---|---|---|---|---|---|
| 1 | GPT-4o | 715 | 763 | 17 | 22.3 |
| 1 | GPT-5 | 2,901 | 2,943 | 16 | 5.4 |
| 1 | GPT-5-mini | 6,906 | 6,990 | 20 | 2.9 |
| 2 | GPT-4o | 908 | 1,035 | 30 | 29.0 |
| 2 | GPT-5 | 6,584 | 6,635 | 32 | 4.8 |
| 2 | GPT-5-mini | 7,318 | 7,523 | 30 | 4.0 |
| 3 | GPT-4o | 717 | 872 | 38 | 43.6 |
| 3 | GPT-5 | 3,801 | 3,904 | 37 | 9.5 |
| 3 | GPT-5-mini | 4,776 | 5,048 | 37 | 7.3 |
| 4 | GPT-4o | 670 | 987 | 58 | 58.8 |
| 4 | GPT-5 | 16,293 | 16,516 | 58 | 3.5 |
| 4 | GPT-5-mini | 12,712 | 14,008 | 56 | 4.0 |
| 7 | GPT-4o | 374 | 1,154 | 115 | 99.7 |
| 7 | GPT-5 | 12,467 | 13,823 | 129 | 9.3 |
| 7 | GPT-5-mini | 19,420 | 19,888 | 121 | 6.1 |
These tokens-per-second values match output tokens divided by total elapsed time, to rounding. They therefore include the initial wait and should not be interpreted as pure post-first-token decoding speed. For example, scenario 4 GPT-5 spent 16,293 of 16,516 milliseconds before the first token; the long wait dominates its end-to-end throughput.
Round 2: six-model averages across eight scenarios
Mean translation response duration
April 2026 benchmark / eight scenarios / lower is faster
| Rank by mean duration | Model | Mean total (ms) | Mean output tokens | Mean tokens/s |
|---|---|---|---|---|
| 1 | GPT-4o-mini | 924 | 55 | 53.4 |
| 2 | GPT-5.4-mini | 1,195 | 68 | 48.0 |
| 3 | GPT-4o | 1,249 | 62 | 42.8 |
| 4 | GPT-4.1 | 1,756 | 62 | 30.8 |
| 5 | GPT-5.4-nano | 1,831 | 53 | 30.1 |
| 6 | GPT-5 | 4,324 | 346 | 73.5 |
Round 2 does not provide per-scenario TTFT, individual run values, variance, percentiles, repeat counts, or confidence intervals. Mean tokens/s is retained as reported; a mean of per-scenario rates need not equal mean tokens divided by mean duration.
What the results support
- GPT-5 with reasoning_effort set to low averaged 4,324 ms, approximately 3.46 times GPT-4o's 1,249 ms in this comparison. The report's shorthand of roughly four times slower is a rounded characterization.
- GPT-5 reported 346 output tokens on average, about 5.6 times GPT-4o's 62 and roughly five to six times most comparison models. Despite its highest reported mean tokens/s, it had the slowest total completion.
- The report attributed the combination of delay and higher token counts to reasoning work unnecessary for a straightforward translation task. That is a plausible explanation, not a demonstrated causal breakdown: no visible-output versus reasoning-token accounting or instrumentation code was supplied.
- Higher measured token consumption may affect cost and provisioned-capacity demand, but this Global Standard experiment did not measure PTU utilization or prove an equivalent PTU multiplier.
- The observations point to workload/model fit, rather than establishing a structural performance problem with every Azure-hosted GPT model.
Model selection and migration recommendations
| Candidate | Recommendation recorded on April 28 | Qualification |
|---|---|---|
| GPT-4o-mini | First candidate: fastest mean duration, 924 ms; described as low-cost and adequate for translation | No price table or scored quality evaluation accompanies that judgment. |
| GPT-5.4-mini | Second candidate: 1,195 ms, close to GPT-4o, and a newer migration option | Validate model availability, quality, and lifecycle for the target deployment. |
| GPT-4o | Third candidate: familiar existing baseline, 1,249 ms | A historical baseline, not a reason to ignore retirement planning. |
| GPT-4.1 | Usable but slower in this workload, 1,756 ms | The report describes it as stable without reporting reliability statistics. |
| GPT-5 | Not recommended for this latency-sensitive translation workload | This is not a universal claim that GPT-5 cannot translate or is unsuitable for every translation use case. |
| GPT-5.4-nano | No demonstrated latency advantage over mini in this test, 1,831 ms | The experiment does not establish that nano has no benefit for other workloads or cost objectives. |
The proposed migration path was GPT-4o (1,249 ms) to either GPT-4o-mini (924 ms) or GPT-5.4-mini (1,195 ms). Those figures are the historical measured means, not expected production service levels.
Validation before changing production
- Repeat the same workload on the intended PTU deployment if production uses provisioned throughput. Global Standard and PTU can have different latency patterns.
- Evaluate translation quality separately, including a metric such as BLEU and task-appropriate review. The source does not provide BLEU scores or blinded human judgments.
- Try a shorter system prompt, such as
Translate. Output only., as a separate experiment. A speed improvement was suggested, not measured. - Preserve model/version, deployment type, reasoning setting, token-accounting definitions, region, concurrency, and measurement boundaries when reproducing the results. The simultaneous six-model calls and shared B1 application are part of the measured environment.
- Collect repeated runs and tail-latency/error data before setting an operational target; the published averages alone cannot establish reliability or a service-level objective.
Lifecycle notes recorded in the report
The following dates and statuses reproduce the report's April 28, 2026 lifecycle notes for context. They have not been revalidated here and must not be treated as the current service schedule. Azure API deployments and ChatGPT product availability are separate.
| Model | Version | Status stated in the report | Historical note |
|---|---|---|---|
| GPT-4o | 2024-05-13 | Deprecating | Older version already in deprecation |
| GPT-4o | 2024-08-06 | Deprecating | Older version already in deprecation |
| GPT-4o | 2024-11-20 | GA | Retirement listed as October 1, 2026; Standard auto-upgrade listed as March 31, 2026 |
| GPT-4o-mini | 2024-07-18 | GA | No announced date recorded |
| GPT-4.1 | 2025-04-14 | GA | No date recorded |
| GPT-5 | 2025-08-07 | GA | No date recorded |
| GPT-5.4-mini | 2026-03-17 | GA | No date recorded; described as new |
| GPT-5.4-nano | 2026-03-17 | GA | No date recorded; described as new |
The report also states that GPT-4o, GPT-4.1, and GPT-4.1-mini were removed from ChatGPT on February 13, 2026, while noting that API lifecycle is separate. Consult the public Azure OpenAI model-retirement documentation for current deployment-specific dates and upgrade behavior.
Materials and reproducibility limits
The folder contains one report and no companion application source, full test corpus, screenshots, binary assets, or raw measurements. The complete shared system prompt is included inline above. No downloadable application or result dataset is offered because none was present. All reported numerical result rows are retained; abbreviated Korean prompts have been described in English rather than replaced with invented full prompts. Resource names and the live application address are intentionally excluded.
Resources
No standalone companion files are distributed for this topic. Use the in-page examples and review the editorial notes below for source availability.
Editorial notes
This page consolidates the following source documents into an English technical guide:
- GPT Translation Performance Comparison Report — April 28, 2026
- Actual Azure resource, resource-group, application, deployment-alias, and endpoint identifiers were removed or replaced with YOUR_ placeholders; no private source or live application links are included.
- The source supplies only abbreviated prompt descriptions, no complete test corpus or application code, and no raw result exports; missing artifacts have not been reconstructed.
- Korean prompt excerpts are represented by English descriptions. All reported numerical results, environment details, the complete common system prompt, and historical lifecycle notes are retained.
- No screenshots or binary originals are distributed. Lifecycle, pricing, quality, and reasoning-overhead assertions are clearly separated from measured findings.
No original credentials or private repository links are included. Do not put populated configuration files or copied production outputs back into this public site.