Most subtitle translation tools that exist today were designed in 2018 to 2022. They were built around a translation engine — Google Translate, or a similar neural machine translation (NMT) system — that was best-in-class at the time. That engine is now measurably behind. Modern AI translation, powered by large language models like Gemini, GPT-4o and Claude, beats classical machine translation by 8–15% on COMET quality benchmarks for complex content. The gap is largest on exactly the kind of text subtitles are made of: dialogue, slang, idioms, fast-paced conversational speech.
This is the technical context for why Sublo exists, and why we made the choices we did when we built it. The honest version of the argument, with the numbers behind it. Skip ahead if you want — but the data is more interesting than the conclusion.
Quick disclosure: I build Sublo. The numbers and benchmarks below come from independent research; the interpretation is mine. I've tried to be specific about where Sublo's design wins and where it doesn't.
What actually changed between 2022 and 2026
Classical neural machine translation — the technology behind Google Translate, DeepL and the translation engines inside most subtitle browser extensions — was a major leap forward when it replaced statistical machine translation around 2016. It was trained sentence-by-sentence on parallel corpora, optimized for word-level accuracy, and worked extremely well on grammatical, well-structured text.
Large language models translate differently. They are not specialized translation systems; they are general models that translate as a side effect of having read enormous amounts of multilingual text. Two things follow from that:
They understand context. A classical NMT system translates each sentence in isolation. An LLM can take in surrounding lines, infer character voice, follow a running joke. For subtitles — where one line often only makes sense in the context of the previous five — this matters a lot.
They handle the messy edges. Slang, idioms, register shifts, colloquial speech, code-switching between languages mid-sentence. The texture of real dialogue. Classical NMT systems were trained on news, web crawls and parliamentary proceedings. They were never very good at conversation. LLMs read Reddit, fan fiction, transcripts and dialogue corpora. They are.
The benchmark numbers
The standard quality metric for machine translation is COMET — a learned metric that correlates better with human judgment than older measures like BLEU. Higher COMET means closer to human-quality translation.
A 2025 evaluation by TokenMix across multiple language pairs and content types found that frontier LLMs (Gemini, GPT-4, Claude) outperform Google Translate by 8–15% on COMET scores for complex content — legal documents, marketing copy, technical writing. The gap narrows to 2–5% on simple, well-structured content where classical NMT is already strong.
Subtitle text is firmly on the complex end of that spectrum. It is full of incomplete sentences, references to off-screen context, register shifts and culture-specific expressions. The 8–15% range is where it lives.
Google's own data confirms the direction. In December 2025, Google announced a major upgrade to Google Translate — replacing its underlying NMT engine with a Gemini-based system. Google claims state-of-the-art performance on the WMT25 Machine Translation benchmark, the standard yearly evaluation for the field. The interesting subtext: Google itself is moving away from the NMT pipeline that most existing subtitle tools are still wired into.
If you are using a subtitle extension that integrated with the classical Google Translate API back in 2020 and has not changed its translation backend since, you are now using a system that Google itself is replacing.
COMET score improvement (2025 benchmark)
Sources: TokenMix 2025 cross-domain LLM-vs-NMT evaluation (COMET); Google Cloud Translation evaluation (M-ETA score, Gemini vs NLLB). Bar lengths visualize relative magnitude; values show the actual gap.
Where the gap is biggest
The aggregate COMET numbers hide the most interesting story. The gap between LLM translation and classical NMT is not uniform — it shows up most on specific categories of text.
Named entities and cultural references. In one 2025 study, Gemini scored almost twice as well as the previous NLLB translation model on M-ETA (a metric for named-entity translation accuracy). For a K-drama or anime where character names, food, places and pop-culture references appear constantly, this is the difference between "I followed every reference" and "I had to pause to Google three things per episode."
Idioms and metaphors. The 2025 paper A Testset for Context-Aware LLM Translation in Korean-to-English Discourse-Level Translation tested 600 Korean text instances with six challenging linguistic phenomena, including idioms, slang and inter-sentential context. Their finding was nuanced: state-of-the-art LLMs like GPT-4o still struggle with the hardest idioms, but they substantially outperform classical NMT systems on the same set. And chain-of-thought-prompted LLMs perform even better — a quality lever that classical NMT simply cannot pull.
Conversational rhythm. Real dialogue contains overlapping turns, half-finished sentences, emphasis cues. NMT flattens all of it into clean neutral prose. LLMs preserve register: a teenager sounds like a teenager, a formal news anchor sounds like a formal news anchor. For language learners specifically, this register information is core learning content. Losing it is a real cost.
Why this matters more for subtitles than for text
Translation quality matters everywhere, but it matters in a specific way for subtitles. Three reasons:
The reader cannot stop to think. A subtitle is on screen for two to four seconds. If the translation is awkward, the viewer cannot pause, reread, or compare against the original. Quality has to land on the first pass. Stiff translation that is technically accurate but unnatural fails harder in subtitles than in any other medium.
The reader is decoding two streams at once. A language learner is listening to the original audio while reading the translated subtitle, looking for the mapping between the two. If the translation is too literal, the mapping makes no sense ("why did they translate that as that?"). If the translation is too loose, the mapping disappears. LLMs hit the middle ground more often.
Dialogue is the worst-case for classical NMT. The training data for classical NMT is biased toward written, edited, grammatical text. The training data for modern LLMs includes massive volumes of transcribed dialogue, scripts and informal text. The dialogue gap is one of the biggest in the entire field. Subtitles sit exactly in that gap.
The cost story that made all of this possible
There is a parallel story that is just as important as the quality story. The cost of running LLM inference has dropped roughly 30–50% per year since 2023, and accelerated sharply in late 2025: LLM API prices dropped by approximately 80% between early 2025 and early 2026.
Concrete numbers: GPT-4 launched at $30 per million input tokens in early 2023. The equivalent capability via GPT-4o now costs $2.50 — a 12x reduction in three years. Gemini 2.5 Flash is cheaper still. The Chinese DeepSeek R1 launched at one-tenth of Western pricing, accelerating the industry-wide price war.
This is what makes a tool like Sublo viable at €2.89 per month. Two years ago, the API call costs alone would have made that price impossible. Today, the cost of translating every line of a 45-minute K-drama episode is a fraction of a cent. The price-quality curve has shifted so far that you can now ship LLM-based subtitle translation as a mass-market consumer product.
Frontier model launch prices, 2023–2026
Sources: CloudZero 2026 LLM API pricing comparison; OpenAI public pricing history.
This is also why tools built on the older NMT pipeline have a hard upgrade path. Their cost structure was designed around free or near-free translation. Switching to LLM translation at scale would change their unit economics fundamentally. Most of the older subtitle extensions have not made that switch — they are stuck on the engine that was right for 2020.
Where Google Translate is still fine
Being honest about the data also means being honest about where the gap is small or zero. Three cases where classical NMT still does the job:
Well-structured, formal content. News articles, technical documentation, legal text in major languages. Google Translate and DeepL are still excellent here. The COMET gap drops to 2–5%, and at that range, the difference is hard to feel as a user. If you are translating Wikipedia for fun, you do not need an LLM.
High-resource European language pairs. English-Spanish, English-German, English-French. Classical NMT had a decade to optimize these pairs and the training data is enormous. LLMs win, but the margin is small enough that picking based on speed or cost can be more important than picking based on quality.
One-shot lookup translations. If you are using browser translation to read a single foreign-language page, the speed and free price of Google Translate matter more than the last 8% of quality. The gap shows up in long-form usage, not in glance translations.
The places the gap matters are exactly the places people watch foreign-language video for entertainment and language learning: conversational dialogue, idioms, cultural references, character voice. For those, the data is unambiguous.
What is still hard for everyone (the honest part)
LLM translation has crossed a meaningful quality threshold but it is not perfect. The Korean-to-English research above is explicit that the hardest idioms still trip up GPT-4o. Chinese four-character idioms remain a known weakness. Wordplay, puns and culturally untranslatable jokes are still hit-or-miss for every system in existence. Translating a Korean comedy where the joke depends on a Korean pun is going to read as a slightly-off rendering no matter what tool you use.
What changed is the baseline. The default translation of casual dialogue used to be wooden. It is now natural. The default translation of a song lyric used to be a literal nonsense rendering. It is now plausible. The remaining failure cases are at the edges of human translation difficulty, not in the middle.
What this means for the tools you actually use
If you are choosing a subtitle translation tool in 2026, the translation engine underneath is the single biggest determinant of quality you will experience. Tools that integrated with Google Translate's classical API in 2020 and never changed are now visibly behind. Tools that integrated with the latest Gemini, GPT or Claude APIs are visibly ahead. The user-facing UI differences across these tools are surface-level compared to that gap.
This is the gap Sublo was built to close. Sublo translates streaming subtitles using Gemini AI, in real time, across Netflix, Disney+, HBO Max, Amazon Prime Video, Apple TV+, Crunchyroll, YouTube and the rest. The translation quality lift over Google-Translate-based tools is the gap above — 8–15% better on the kind of dialogue you actually watch. The pricing — €2.89 per month — is what the dropped LLM API costs make possible.
For a side-by-side breakdown of how Sublo compares to the most common alternatives, see our comparison of subtitle translators in 2026 or the specific pages for Sublo vs Language Reactor and Sublo vs Trancy.
The short version
AI translation crossed a quality threshold in 2024-2025 that classical machine translation cannot match by simply being upgraded. The gap is 8–15% on complex content like subtitles, biggest on dialogue, idioms, named entities and cultural references. API costs dropped 12x in three years, which is what made LLM-based subtitle tools possible as mass-market products. Most existing subtitle translation extensions were designed around the old pipeline and have not switched. Sublo was built around the new one.
If you have not tried AI-powered subtitle translation on the streaming services you actually use, the gap is bigger than you expect. It is also free to test.
Try Gemini-powered subtitle translation on Netflix, Disney+, HBO Max, YouTube and 8 more streaming services — free, no account required.
Install Sublo