Toolkit

Falcon ASR is a 1.6 billion parameter (a measure of model size) speech-to-text model built specifically for Arabic, with strong support for the Emirati dialect. It also handles English, French, Spanish, and Portuguese using the same model weights, with no language flag needed — it detects what is spoken and transcribes accordingly.

Developed by the Technology Innovation Institute in Abu Dhabi, it achieved a word error rate (WER — the percentage of words transcribed incorrectly) of 20.92% across six Arabic test sets, beating the previous best published result by around 2.25 percentage points. On an internal Emirati evaluation, it outperformed all compared systems at both word and character level.

You can try it now via the Hugging Face Demo Space — paste or upload audio and see the transcript. Model weights are available to download if you want to run it locally.

Worth knowing: The Emirati evaluation used internal recordings and human-validated transcripts rather than fully public benchmarks, so treat those specific numbers as the company's own testing rather than independently verified results.

OpenAI has launched textGrain, an invisible watermark embedded in text generated by ChatGPT and Codex. It adds a statistical signal to the model's word choices — imperceptible to readers, but detectable by a dedicated tool. EU users will see it applied automatically in the coming weeks. API customers globally can opt in today.

The watermark does not change how the text reads or performs. According to OpenAI's own testing, it detected watermarks in roughly 80% of shorter passages and 95% of longer ones, at a 1% false positive rate. Detection drops sharply if words are swapped out.

To check whether text carries an OpenAI watermark, apply to access the detector at openai.com. Currently limited to approved researchers and organisations.

Worth knowing: A watermark is not proof of AI authorship. OpenAI says it cannot measure human contribution, and editing even 10% of words meaningfully reduces detection reliability.

OpenAI has published a practical guide for choosing between the three models in its GPT-6 family, available via the API (a connection that lets developers plug the model into their own apps or workflows), according to OpenAI.

There are three tiers:

• GPT-6 Astra — the most capable, at $10 per million input tokens (a token is roughly three-quarters of a word) and $50 per million output tokens. For genuinely hard reasoning work.
• GPT-6.1 Sol — similar capability at about one-fifth the cost. Good for complex coding or research tasks.
• GPT-6 Luna — the fast, cheap option at $0.10 input / $0.50 output per million tokens. Best for repetitive, well-defined tasks like classification or structured summaries.

The guide also covers prompt caching — reusing previous context to cut costs by up to 95% — and how to structure long-running tasks so they do not stall.

To get started: visit the OpenAI API platform, pick a model based on your task complexity and budget, and run a small test batch before committing to a full workflow.

Worth knowing: All three are API-only — this is for developers building applications, not for casual use through ChatGPT.

The Open TTS Leaderboard, available now on Hugging Face, ranks text-to-speech (TTS) models — software that converts written text into spoken audio — using objective, automated metrics rather than human voting panels.

It tracks three things: how accurately the audio reproduces the original words (measured by word error rate), how fast the model generates audio, and how closely a cloned voice matches the original speaker's voice. Over 8,000 TTS models are available on Hugging Face; this leaderboard focuses specifically on open-weights models, which are underrepresented on existing ranking sites.

If you are choosing a TTS model for a project, or just curious which open-weights options are worth testing, filter by language and check the Pareto plots — they show which models balance accuracy and speed without needing to read every row.

Access it here: Open TTS Leaderboard

Worth knowing: The leaderboard measures intelligibility and speed objectively, but does not measure naturalness or how pleasant a voice sounds — those still require human ears to judge, and scores here won't tell you which model sounds best.

Holo4 is a set of open-weights (downloadable and runnable models where the weights — the trained numerical values that define how the model behaves — are made publicly available, though training data and code are not) AI agent models that can operate across desktop interfaces, websites, Android, code environments, and business APIs. It comes in two sizes: 27B and 35B-A3B, where B refers to the number of parameters (the internal numerical settings that determine model behaviour). Both are available now via the H Models API.

What makes it notable: most agent models are built for one interface only. Holo4 handles all of them with the same model.

On OSWorld 2.0 — a benchmark for desktop control tasks, run by H Company — Holo4 27B scores 61.7%. That trails Anthropic's Opus 5.5 at 81.8%, but at significantly lower cost per task.

You can access it at Hugging Face or through the H Models API. Pricing applies; check the API quickstart for current rates.

To try it: create an account on the H Models API, follow the quickstart guide, and pass it a task description alongside access to your target interface.

Worth knowing: The 27B model requires substantial memory to run locally — cloud API access is the practical route for most users, which means your inputs are processed on H Company's servers.

Google has updated Gemini 3.8 Live with a new "Live Avatar" feature, according to The Verge. Gemini is Google's AI assistant. The Live Avatar adds an animated on-screen persona that lip-syncs and shows facial expressions in real time as the model speaks.

It supports 97 languages and can switch between them mid-conversation without the lip-sync breaking down. It can also surface information on screen while talking. Google says all output is watermarked using its SynthID system (an invisible digital tag embedded in AI-generated content).

The catch: it is currently only available to Gemini Enterprise customers — not the free or standard paid tiers.

If your organisation uses Gemini Enterprise, you can access it through the Gemini app now. Organisations can choose from preset avatars or create their own.

Worth knowing: This is an enterprise-only feature for now; if you are on a personal or standard business plan, you cannot access it today.

If you build or run applications on OpenAI's API — the programming interface that lets software connect to OpenAI's models — GPT-6 now has an improved prompt caching system worth understanding.

Prompt caching works like this: when your app sends the same instructions or context repeatedly, the model stores a copy and reuses it instead of reprocessing from scratch. OpenAI is offering up to 90% off on those cached tokens (the chunks of text the model processes), with eligible shared content cached within a 30-minute window.

New this week: a Prompt Caching Dashboard to track hit rates over time, a diagnostics tool that tells you exactly why a cache miss occurred (for example, if your tool definitions changed), and explicit cache breakpoints so you choose what to preserve.

You can now also adjust reasoning effort between requests without breaking the cache — useful if some follow-up tasks need less processing power than the original.

Access the tools and guide at openai.com.

Worth knowing: This is an API-level feature — it affects developers building on GPT-6, not end users of ChatGPT directly.

TypeSafe AI's Jev is a different kind of AI model: you send it text or structured data, and instead of generating a written response, it returns a number — a confidence score for a yes/no question, a probability across a set of choices, or a numeric rating. Simon Willison covers it in detail here.

It's built for classification tasks: spam detection, content labelling, relevance ranking, prioritisation. You send one document and as many questions as fit in the context window (the maximum amount of text the model can process at once), and all questions are scored in parallel.

Pricing is $0.042 per million input tokens — cheaper than GPT-5 Nano — and output is free. The Jev 1.13 jaggedness documentation notes it struggles with numbers, dates, and adversarial content. Get started at TypeSafe AI.

Worth knowing: Jev returns only a floating-point number — there is no explanation for why it scored something the way it did. If you use it to rank or filter people rather than content, any bias in the model is invisible and very difficult to detect after the fact.

Given today's coverage of AI watermarking, it is worth knowing that Google has made a SynthID detection tool available via Hugging Face. SynthID-Text (Google's watermarking system, explained in today's issue) embeds a hidden signal in AI-generated text. The Hugging Face demo lets you paste any passage of text and check whether it carries that signal — useful for editors, teachers, or anyone who regularly receives written submissions and wants a quick first check.

Find it at: huggingface.co/spaces/google/synthid-text

It is free to use, runs in a browser, and needs no account.

Worth knowing: SynthID detection only works for text generated by systems that use SynthID watermarking specifically. It will not flag AI-generated text from models that use a different approach, or no watermarking at all — so a clean result is not confirmation the text was written by a human.

Anthropic has launched Claude Docs and Claude Slides in beta — tools for creating shareable documents and presentations directly inside a Claude conversation, as reported by The Verge.

You describe what you want, Claude builds it, and you can edit manually or ask Claude to revise in plain language. Finished files live at a shareable link and can be opened for collaborative editing by other users, with comments working similarly to Google Docs. Alongside this, Anthropic has merged its separate Cowork and regular chat interfaces into a single "one Claude" experience, so these tools are available from any conversation without switching modes.

If you already use Claude Pro or Max (web, desktop, or mobile), this is rolling out now and requires nothing to switch on. Team and Free users get access later.

To start: open a Claude chat and ask it to draft a document or presentation on any topic.

Worth knowing: Claude Docs and Slides are in beta, and both your inputs and outputs are subject to Anthropic's standard data practices — check those before pasting anything confidential into a generated document.