# AI news, Wednesday, 21 October 2026

Compiled by Richard Brooks's AI news desk. Canonical: https://richard-brooks.com/ai-news/2026-10-21/

## Ant Group's inclusionAI open-sources 5 Ming-3.0 speech models under MIT

Language. Covers text-to-speech in 7 languages plus Cantonese and Sichuanese, audio understanding and dialect-aware speech recognition. Gives teams a self-hosted voice stack with cloning and emotion control, though recognition is Chinese and English only.

Source: inclusionAI (Hugging Face), https://huggingface.co/inclusionAI/Ming-3.0-tts-0.5B

## Anthropic says Claude ran a 64-hour protein design campaign against scorpion toxins

Agents. Claude generated 318,484 designs largely on its own, and 17 of 211 lab-tested proteins bound their target. It shows an agent running a long research job with 4 follow-up prompts, though no antivenom has been proven.

Source: Anthropic, https://www.anthropic.com/research/claude-scorpion-toxin-binder-design

## Cloudflare previews Clairic, a model that extracts typed values instead of text

Tools. Returns invoice totals, dates and estimates with confidence scores in 1 pass, with no generated text and input-only pricing. Suits document pipelines where an LLM's free-text output has to be checked. Open weights are promised.

Source: Cloudflare, https://blog.cloudflare.com/clairic-extraction-model/

## EU publishes first AI Act harmonised standard, EN 18286 for quality management

Policy. Providers of high-risk AI systems that follow EN 18286 gain a presumption of conformity with Article 17. It sits on top of ISO/IEC 42001 or ISO 9001, giving compliance teams a concrete route before high-risk rules apply.

Source: European Commission, https://digital-strategy.ec.europa.eu/en/library/first-harmonised-standard-supporting-ai-act-published-official-journal

## Gradium releases Phonon, open-weight text-to-speech that runs offline on devices

Language. The 31m English model is Apache 2.0 and runs on phones, laptops and in the browser, so apps can speak without sending text to a cloud. The 40m models add French, German, Spanish and Portuguese but need a commercial licence.

Source: Gradium, https://gradium.ai/phonon

## Inception launches Mercury 3, a diffusion LLM running at over 1,000 tokens a second

Models. Generates text in parallel rather than word by word, with a 1M-token context at $0.20 per 1M input tokens. Aimed at voice support, tutoring and browser agents, where response delay costs customers.

Source: Inception, https://www.inceptionlabs.ai/blog/introducing-mercury-3

## OpenAI says GPT-6.1 Sol migrates COBOL with 86.6% accuracy, up from 5.4%

Models. Scored on byte-for-byte output parity when porting COBOL to Java, Python or Rust. Relevant to banks and public bodies running legacy systems, though OpenAI still advises expert review before production use.

Source: OpenAI, https://openai.com/index/low-resource-programming-languages

## UK FCA says firms' incident reports must cover significant AI failures and near-misses

Policy. UK financial firms using AI agents should expect to log close calls, not just major outages, and to name who is accountable. The FCA's data chief Jessica Rusu set this out on 19 Oct; no new rule has been published yet.

Source: The Register, https://www.theregister.com/ai-and-ml/2026/10/20/uk-finance-watchdog-wants-firms-to-report-ai-failures-and-close-calls/5304078