OpenAI models for translation: GPT-4.1, GPT-4.5, GPT-4o, and what comes next

When this article was first published in April 2025, it compared three active OpenAI models. Within months, two of them were retired. GPT-4.5 was deprecated from the API on July 14, 2025 and removed from ChatGPT in August 2025. GPT-4o followed: retired from most ChatGPT tiers with the GPT-5 launch on August 7, 2025. OpenAI has since released GPT-5, GPT-5.2 (January 2026), and GPT-5.5 (April 2026).

This article has been updated to reflect the current OpenAI model landscape, explain what each GPT-4 model actually offered for translation, document what's now available, and explain what the shift to GPT-5 means in practice for translation work.

In this article

1. The GPT-4 model family: what each version offered
2. What happened to GPT-4.5 and GPT-4o
3. GPT-4.1 in 2026: the surviving GPT-4 model
4. GPT-5 and beyond: what changed for translation
5. How OpenAI models fit within MachineTranslation.com
6. Frequently asked questions

The GPT-4 model family: what each version offered

Three distinct OpenAI models defined the GPT-4 era for translation:

GPT-4o launched in May 2024 as a multimodal model handling text, voice, and image input. It was OpenAI's headline model through most of 2024, scoring 94.2/100 in MachineTranslation.com's internal translation benchmarks — among the highest of any individual model measured. For high-resource European and East Asian language pairs, GPT-4o produced contextually strong, natural-sounding output. Its limitation for translation specifically was the same as every single-model tool: it provided one output with no quality signal, no cross-check, and no way for the user to assess confidence in the result.

GPT-4.5 launched in February 2025 as a large research preview model, OpenAI described it as their "largest-ever" model in terms of compute at that time. It was primarily trained using unsupervised learning, which improved pattern recognition and creative coherence. For translation, GPT-4.5 produced output with strong creative and tonal qualities. Its practical limitation was cost: at $75 per million input tokens at launch, it was 30× more expensive than GPT-4o. OpenAI itself concluded that GPT-4.1 offered similar or improved performance at far lower cost, which accelerated the deprecation decision.

GPT-4.1 launched in April 2025, positioned as a cost-efficient upgrade over GPT-4o with major gains in instruction following, coding, and long-context understanding. For translation, GPT-4.1's key improvement over GPT-4o was instruction compliance: it follows explicit translation requirements (terminology constraints, register specifications, formatting rules) more reliably across long documents. In Intento's State of Translation Automation 2025, GPT-4.1 led among single-agent solutions with top performance in 7 of 11 language pairs — the strongest single-model showing in the evaluation. GPT-4.1 remains in active API use as of May 2026.

What happened to GPT-4.5 and GPT-4o

GPT-4.5 timeline:

• February 27, 2025: GPT-4.5 released via ChatGPT Plus and Pro, and via OpenAI API
• April 14, 2025: OpenAI announced GPT-4.5 deprecation from the API
• July 14, 2025: GPT-4.5 removed from the API entirely
• August 7, 2025: GPT-4.5 removed from ChatGPT Plus and Teams with GPT-5 launch; Pro users retain access under "Legacy Models"

GPT-4.5 had the shortest active lifespan of any commercial OpenAI model, approximately four months in the API. OpenAI cited GPT-4.1's equivalent performance at much lower cost and latency as the primary reason.

GPT-4o timeline:

• May 2024: GPT-4o launched as OpenAI's primary multimodal model
• August 7, 2025: GPT-4o retired from most ChatGPT tiers with the GPT-5 launch; some versions continued in the API
• February 2026: GPT-4o versions 2024-05-13 and 2024-08-06 retired from Azure OpenAI; remaining users auto-upgraded to GPT-5.1

For most users today, GPT-4o is no longer the model they receive when using ChatGPT or OpenAI's standard API. It is effectively a legacy model.

What this means for content referencing these models: Any article, workflow, or evaluation built around GPT-4.5 needs to be updated. GPT-4o benchmarks from 2024 remain useful as historical reference points but should not be treated as current capability comparisons. GPT-4.1 remains the relevant GPT-4 family reference for current API work.

GPT-4.1 in 2026: the surviving GPT-4 model

As of May 2026, GPT-4.1 is the only GPT-4 family model with active, maintained API support. Key translation-relevant characteristics:

Benchmark position. In Intento's 2025 evaluation, GPT-4.1 led among single-agent solutions — 7 "best" performances across 11 language pairs. It appears in the "best" category for English to Arabic, English to Spanish, English to French, English to Italian, English to Korean, English to Dutch, English to Portuguese, English to Ukrainian, and English to Chinese (automated evaluation). In human LQA evaluation, it leads among OpenAI models for Arabic, French, Italian, Japanese, Korean, Dutch, Portuguese, Ukrainian, and Chinese.

Instruction following. GPT-4.1's measurable improvement over GPT-4o is instruction compliance. For translation tasks with explicit requirements (use only these terms, maintain this register, preserve this formatting) GPT-4.1 follows them more consistently across long documents.

Context window. GPT-4.1 supports a 1M token context window, allowing very long documents to be processed in a single pass without chunking. Documents processed in fragments show a 28% higher rate of terminology inconsistency compared to whole-document processing. Source: MachineTranslation.com internal data.

API cost. GPT-4.1 is significantly cheaper than GPT-4.5 was: $2/million input tokens and $8/million output tokens, making it practical for high-volume translation workflows.

GPT-5 and beyond: what changed for translation

GPT-5 launched August 7, 2025 as a unified system that combines fast response and deep reasoning in a single model, routing automatically between them based on query complexity. For translation, the most practically relevant changes are:

Substantially reduced hallucination rate, OpenAI's data shows GPT-5's responses are approximately 45% less likely to contain factual errors than GPT-4o (with web search, on anonymized prompts). For translation, this means fewer semantic errors in content involving proper nouns, technical terminology, and proper-name entities.

Expanded context window, 400K tokens via the API, with a 1M+ token context window in development. This exceeds GPT-4.1's already large window, further reducing the need to chunk long documents.

Improved multilingual benchmarks, GPT-5 tested across 15 languages on MMLU, outperforming GPT-4o in all. In translation-specific workflows, businesses report stronger handling of ambiguity in clinical and technical contexts.

GPT-5.2 (January 2026) specifically called out translation improvements in its release notes: "clear improvements in technical writing and translation." GPT-5.2 Instant is described as designed for everyday knowledge work tasks including translation.

GPT-5.5 (April 2026) is the current leading OpenAI model as of this update.

For translation specifically, the most important implication of the GPT-5 era is not which version to choose — it is that any single-model output, regardless of how capable the model is, still produces errors it cannot detect in its own output. GPT-5 reduces the hallucination rate meaningfully. It does not eliminate model-specific errors.

How OpenAI models fit within MachineTranslation.com

ChatGPT (OpenAI's model) is one of 22 AI models inside MachineTranslation.com's SMART system. SMART runs all 22 models simultaneously and returns the output the majority agrees on.

What the GPT-4 to GPT-5 transition means for SMART users: OpenAI's model contribution to the consensus continues to improve as OpenAI releases stronger models. Users do not need to track or manually select between GPT generations, the SMART system incorporates the current OpenAI model as part of the 22-model consensus automatically.

In MachineTranslation.com's internal benchmarks, individual top-tier models (including GPT-4o at 94.2/100) are consistently outperformed by the consensus output, which reaches 98.5/100. The gap between any individual model and the consensus reflects model-specific errors being outvoted before they reach the output. This gap exists regardless of GPT generation. Source: MachineTranslation.com internal benchmarks and WMT24 General Machine Translation Findings.

Users who switched from single-engine translation to SMART spent 27% less time verifying and correcting outputs. Source: MachineTranslation.com internal data.

For high-stakes content (legal documents, regulatory filings, clinical material), Human Verification escalates the SMART consensus to a certified professional reviewer within the same platform, with a 100% accuracy guarantee. No external agency required.

Translate with ChatGPT and 21 other AI models at MachineTranslation.com — free, no sign-up required.

Frequently asked questions

1. Is GPT-4.5 still available?

GPT-4.5 was deprecated from the OpenAI API on July 14, 2025, and removed from most ChatGPT tiers when GPT-5 launched on August 7, 2025. Pro users can still access it under "Legacy Models." For most developers and users, GPT-4.5 is no longer the active model.

2. Is GPT-4o still available?

GPT-4o was retired from most ChatGPT tiers with the GPT-5 launch in August 2025. Some API versions remained available through early 2026, but are being replaced by GPT-5.1. For current translation work, GPT-4.1 and GPT-5 are the relevant OpenAI models.

3. Which OpenAI model is best for translation in 2026?

Among the GPT-4 family, GPT-4.1 led Intento's 2025 evaluation with the most "best" performances across 11 language pairs. GPT-5 and GPT-5.2 are the current flagship models with improved multilingual capabilities and lower hallucination rates. For professional translation workflows, the choice of OpenAI model is less consequential than whether you use a single model or a consensus system.

4. What is GPT-4.1's advantage over GPT-4o for translation?

GPT-4.1's primary documented advantage is instruction following, it complies more reliably with explicit translation requirements (terminology constraints, register specifications, formatting rules) across long documents. It also has a 1M token context window and lower API cost than GPT-4o.

5. How does GPT-5 improve on GPT-4o for translation?

GPT-5 reduces factual errors by approximately 45% compared to GPT-4o (with web search, per OpenAI's own evaluation). It has a larger context window (400K tokens via API), improved multilingual coverage, and unified reasoning. In clinical and technical translation, businesses report stronger handling of ambiguous terminology.

6. Is ChatGPT the best AI for translation?

ChatGPT (OpenAI's models) is among the strongest individual models for translation, GPT-4.1 leads Intento's 2025 single-agent evaluation. But every individual model, including OpenAI's, produces errors it cannot detect in its own output. MachineTranslation.com's SMART runs ChatGPT alongside 21 other models and returns the output the majority agrees on, reaching 98.5/100 versus 94.2/100 for the best individual OpenAI model.

7. Does MachineTranslation.com use the latest OpenAI models?

MachineTranslation.com's SMART system includes ChatGPT as one of 22 models. The platform incorporates the current OpenAI model version — users receive the benefit of each generation's improvements automatically as part of the 22-model consensus, without needing to track which GPT version is current.