Adding a second language to your app does not double your API bill. But it does shift it in ways most cost estimates miss. French and German run 15-20% above English token rates. Arabic sits 30-50% above. Japanese tokenizes at roughly English rates per unit of meaning, but output tokens for localized content can run higher if you are generating rather than processing native text. None of that shows up in an English-only benchmark.
This post ranks frontier and budget models across five languages (English, French, German, Japanese, Arabic) on classification, summarization, and chat, and shows where each model's multilingual quality actually breaks down.
The models (and what each costs)
| Model | Provider | Input per 1M | Output per 1M | Tier |
|---|---|---|---|---|
| Claude Fable 5 | Anthropic | $10 | $50 | Frontier |
| GPT-5.5 | OpenAI | $5 | $30 | Frontier |
| Gemini 3.1 Pro | $2 | $12 | Frontier | |
| Claude Haiku 4.5 | Anthropic | $1 | $5 | Budget |
| Gemini 3.5 Flash | $0.75 | $4.50 | Budget | |
| Qwen3 235B-A22B | Alibaba (hosted) | varies by provider | varies | Open-weights |
Pricing verified August 2026 at Anthropic, OpenAI, and Google AI Studio. Qwen3 is available through hosted providers (Together AI, Fireworks, Replicate) at variable rates.
The non-English token cost
Token pricing is per token, not per word. Non-Latin scripts tokenize differently than English, and the gap compounds at production scale.
French and German stay within 10-20% of English token counts for equivalent content, per the token-to-word breakdown we ran across all major providers. Arabic script runs 30-50% above English rates per equivalent unit of meaning. Japanese characters are 1-2 tokens each but carry more meaning per character than English words do, putting them at roughly English parity for native-language input, though localization output (English-to-Japanese generation) typically runs 20-30% longer in tokens.
At 10,000 document summarizations per day through Claude Fable 5 at $10/M input:
| Language | Relative token volume | Daily input cost (10k docs, ~200 words each) |
|---|---|---|
| English | baseline | ~$2.67 |
| French | +15% | ~$3.07 |
| German | +15% | ~$3.07 |
| Arabic | +40% (midpoint) | ~$3.73 |
| Japanese (native input) | ~+5% | ~$2.80 |
Arabic and Japanese localization pipelines routinely surprise teams mid-month. Setting per-language budget caps at the proxy layer is the cleanest fix; routing rules can also help you send Arabic and Japanese to the right model rather than the default.
How each model handles the five languages
Frontier: $2-10/M input
Gemini 3.1 Pro ($2/$12) is the strongest cross-language performer among tier-1 frontier APIs. Google's training corpus spans more non-English text than OpenAI's or Anthropic's, and the difference shows most clearly in Arabic: output coherence on generation tasks degrades less as sentence complexity increases. French and German run at near-English quality across classification, summarization, and chat. Japanese is strong for native-language processing.
For the audience choosing between these three models, Gemini 3.1 Pro at $2/M input delivers the best multilingual quality-per-dollar at every scale. The only scenario where Fable 5 makes sense is when the task is exclusively English-heavy and you only need occasional French or German for support copy. The same tone advantage extends to English: Fable 5 leads our business email ranking for the same reason — tone-sensitive tasks where register control matters.
Claude Fable 5 ($10/$50) performs at the top of every benchmark table, and that extends to French and German specifically. Arabic and Japanese show more variance on long-form generation tasks, narrowing the gap with Gemini 3.1 Pro. If you are shipping a product where French-language quality directly affects retention (customer-facing copy, legal drafts, anything a native speaker will read critically), Fable 5's French output is the safest choice. At $10/M input, that premium needs to be justified by actual output quality, not by the name.
GPT-5.5 ($5/$30) holds frontier quality on French, German, and English classification. The drop-off on Arabic and Japanese nuanced generation is larger than on Fable 5 or Gemini 3.1 Pro. Classification tasks (sentiment, intent labeling, topic routing) remain competitive across all five languages. If your multilingual use case is primarily structured extraction from user input, GPT-5.5 is viable at half the Fable 5 price. If it is open-ended generation or customer-facing chat in Arabic, Gemini 3.1 Pro is a stronger call at a lower price.
Budget: $0.75-1/M input
Gemini 3.5 Flash ($0.75/$4.50) is the value pick for multilingual pipelines with volume. Our French translation benchmarks showed Flash tying GPT-5.5 on total wins at a sixth of the cost per call. That pattern holds for French and German generally. Japanese and Arabic generation falls a half-tier behind frontier on open-ended tasks but stays competitive on classification and structured output extraction. If your pipeline runs 100,000+ calls per month in French or German, Flash is the default until benchmarks on your specific prompts say otherwise.
Claude Haiku 4.5 ($1/$5) handles French and German reliably on shorter generation tasks and classification. Japanese and Arabic performance degrades on longer outputs. For short-form tasks: form validation, intent classification, one-sentence summaries. Haiku 4.5 is a solid budget option across four of the five languages for these. For Arabic chat or multi-paragraph generation, budget for Flash instead.
Open-weights: Qwen3 235B-A22B
Qwen3 leads public multilingual benchmarks including MMLU-ProX (29 languages) and multilingual MGSM as of mid-2026. Japanese and Chinese coverage reflects Alibaba's training data emphasis on East Asian languages. Arabic performance is competitive with Gemini 3.1 Pro. Among best free LLMs, Qwen3 is the strongest option if your infrastructure can handle a hosted open-weights model.
The catch is operational: you are running through a third-party hosting provider rather than a tier-1 API, with the SLA guarantees that entails. For teams already running self-hosted inference, Qwen3 is the answer for Japanese and Arabic. For teams on managed APIs, Gemini 3.1 Pro covers the same ground with better uptime guarantees.
Subscription vs API
| Provider | Model | API input/output per M | Subscription | Best for |
|---|---|---|---|---|
| Anthropic | Fable 5 / Haiku 4.5 | $10/$50 or $1/$5 | Pro $20/mo, Max $100-200/mo | Low-volume quality-critical French/German |
| OpenAI | GPT-5.5 | $5/$30 | Plus $20/mo, Pro $200/mo | English-heavy with occasional FR/DE classification |
| Gemini 3.1 Pro / 3.5 Flash | $2/$12 or $0.75/$4.50 | AI Premium (Gemini Advanced) $19.99/mo | Arabic + volume pipelines |
For a production multilingual pipeline, API beats subscription at almost every scale. Subscriptions cap daily usage, which is the worst outcome for a batch processing job. The API also lets you route by language: French and German to Gemini 3.5 Flash, Arabic to Gemini 3.1 Pro, English to whatever you already run. A single provider subscription cannot do that.
Break-even math: Anthropic's Pro plan ($20/mo) beats the API rate for Haiku 4.5 at roughly 20,000 tokens/day. Above that threshold, the API costs less. For Fable 5, the Pro plan covers about 2,000 tokens/day of output before API wins. Most production workloads cross that in under an hour.
Verify current tiers at Anthropic, OpenAI, and Google. Subscription pricing changes independently of API rates.
Which to use
French and German at scale: Gemini 3.5 Flash. Near-frontier quality, $0.75/M input, and a 15% token premium you can plan around. Only upgrade to Gemini 3.1 Pro if your specific prompts show a visible quality gap.
Arabic: Gemini 3.1 Pro. Strongest cross-language training among tier-1 APIs. Accept the 30-50% token premium as a fixed line item and budget for it from the start.
Japanese (native input): Qwen3 if you can run it; Gemini 3.1 Pro as the tier-1 fallback. Both outperform GPT-5.5 and Fable 5 on Japanese specifically in public benchmarks.
Mixed five-language pipeline: Route at the proxy layer rather than picking one model for everything. The LLMTest proxy handles per-language routing with a single endpoint and unified billing: Latin-script requests go to Gemini 3.5 Flash, Arabic and Japanese route to Gemini 3.1 Pro, English stays on your existing model.
Before committing to any model for a multilingual workload, replay your actual prompts across the candidates through LLMTest benchmarks. The quality gap between Flash and Pro narrows considerably on extractive tasks; the cost difference does not. Running your own prompts takes the guesswork out of whether the premium is worth it for your specific content type.
If you want to see the cost and quality breakdown on your own multilingual prompts, run them through LLMTest. It takes about 10 minutes to get a ranked result against three candidate models.