Polylang AI Translation v4.21: Model Prices, Claude Sonnet 5 & GPT-5.4 Nano

Robbert Vermeulen

Robbert Vermeulen

· 5 min read

Polylang AI Translation v4.21 makes picking an AI model a lot less of a guessing game. Every model in the dropdown now shows what it costs, Claude Sonnet 5 and Opus 4.8 are available, and GPT-5.4 Nano is the new OpenAI default. This post also covers the 4.21.1 and 4.21.2 patch releases, which fix a set of issues around Bricks, ACF, Elementor and translation runs that got stuck.

See what every model costs before you pick it

Choosing a translation model used to mean looking up pricing on three different provider websites. Now the model dropdown in the plugin settings shows the input and output price per million tokens next to every model, for OpenAI, Claude and Gemini alike.

That makes the trade-off visible at a glance. GPT-5.4 Nano costs $0.20 per million input tokens, Claude Opus 4.8 costs $5.00. For a large WooCommerce catalogue that difference adds up quickly, and now you can decide with the numbers in front of you.

We also refreshed the model lists for every provider and flag models that are about to retire, such as Gemini 2.5 Flash, so you’re not caught out when a provider pulls a model.

Claude Sonnet 5 and Opus 4.8, through Anthropic’s native API

Both new Claude models are available in the dropdown, with settings tuned for translation work: deterministic output and no extended thinking, which keeps translations fast and consistent. Claude Sonnet 4.6 stays the default, so switch to Sonnet 5 or Opus 4.8 in the settings if you want to try them.

Under the hood, Claude translations now go through Anthropic’s native Messages API instead of a compatibility layer. You get Anthropic’s own error messages when something goes wrong, and we fixed a conflict between the Claude client and WordPress internals along the way.

GPT-5.4 Nano is the new OpenAI default

GPT-5.4 Nano is the cheapest credible translation model on any provider right now, at $0.20 / $1.25 per million tokens, and it handles translation very well. If you’ve never picked a model yourself, you’re now on GPT-5.4 Nano. If you did pick one, nothing changes.

Gemini users get two improvements too: Gemini 3 models are handled more efficiently during translation, and the new Gemini API key format works without any extra setup.

Fixes in 4.21.1 and 4.21.2

Two patch releases followed 4.21.0, most of them the direct result of support tickets. If any of these sound familiar, this is the update you’ve been waiting for.

Bricks: page content silently left untranslated. Since v4.19.0 a Bricks page could end up with a translated title and slug but an untranslated body, while the activity log reported everything as completed. A second variant of the same problem hit translations that ran from a background worker. Both are fixed. If you have affected pages, re-translate them after updating.

ACF: source post data overwritten or wiped. Translating a post with ACF Repeater fields could overwrite the repeater values on the source post with the translation. And relationship or post object fields could disappear from the source and all translations when a translation was saved with those fields empty. Both fixed.

Elementor: HTML widget content is now translated. Text inside Elementor’s HTML widget was skipped. It’s now picked up like any other widget.

Out of credits? You’ll know. When your OpenAI or Claude account ran out of credits, a translation run would sit on “processing” forever with no error anywhere. The run now fails visibly with a clear message, so you can top up and try again instead of wondering why nothing happens.

OpenRouter reasoning models. Reasoning models on OpenRouter that returned no content caused an “Invalid AI response structure” error. Handled now.

Bug fixes

  • Fixed AI refusals being written into blank translation fields.
  • Fixed the max output tokens setting to always allow at least 16,000 tokens.
  • Raised the connection timeout for AI requests to 30 seconds for slow networks.
  • Fixed the Single Translator reporting completion before the worker actually started.
  • Restored the Bricks array translation opt-in that was accidentally dropped.
  • Fixed a field restarting endlessly when the source content changed mid-translation.

What’s changed

  • Input and output price per million tokens shown for every model in the dropdown.
  • New Claude models: Sonnet 5 and Opus 4.8, with translation-tuned settings.
  • Claude now uses Anthropic’s native Messages API.
  • GPT-5.4 Nano is the new OpenAI default model.
  • Refreshed model lists for all providers, with retiring models flagged.
  • Improved Gemini 3 handling and support for new Gemini API keys.
  • 4.21.1 and 4.21.2: fixes for Bricks page content, ACF repeater and relationship fields, the Elementor HTML widget, OpenRouter reasoning models and runs stuck on “processing” when out of credits.

Update now

This update is available now for all customers. Go to Plugins → Updates in your WordPress admin, or download the latest version from your account dashboard.

If you’re not using Polylang AI Translation yet, check it out here. Questions, feedback, or need a hand? Open a support ticket and we’ll help you out.

Want the full history of every release? See the complete changelog.

Robbert Vermeulen

Robbert Vermeulen

Founder EPICWP & 13y+ WordPress dev

View all posts →