AI products

Research preview of a visual thinking model that zooms into images, rewatches video at denser frame rates and uses tools that write code while it reasons; available through an API and a chat UI at $0.60 per million input and $3.00 per million output tokens.

Rules angle: Analysing video or images of people can engage privacy and biometric rules such as GDPR and the EU AI Act; check before using it on customer or CCTV footage.

16 Oct 2026 · Source

Upgrade to V4.1 Flash on the DeepSeek API, web and app; API model name (deepseek-flash) and pricing unchanged. DeepSeek also opened its in-house search API.

Rules angle: Chinese models are subject to CAC filing rules for public services in China.

17 Oct 2026 · Source

Open-weight 5.6B-parameter image model for text-to-image, editing and up to 10 reference images, with native 4K output; also in the BFL API.

Rules angle: FLUX Kommunity License allows free commercial use only for individuals and companies under $5M annual revenue; generated images must be machine-readably marked under EU Art. 50(2).

15 Oct 2026 · Source

Open 120B-parameter multimodal model for enterprise agents, plus Bolt 74B-A9B, which fits in 24GB of memory for local agents; Bolt E4B coming 23 Oct.

Rules angle: Open weights, datasets and training recipes let firms run it on their own infrastructure; EU GPAI rules may apply to those who modify and redistribute it.

15 Oct 2026 · Source

First open-weight model: 501B-parameter mixture-of-experts (23B active) with Apache 2.0 weights on Hugging Face, plus Reflection's API in public beta.

Rules angle: Permissive licence allows self-hosting; EU GPAI rules may apply to those who modify and redistribute it.

16 Oct 2026 · Source

Updated full-duplex voice models in ChatGPT Voice with more natural accents and new voices in French, German, Hindi, Japanese, Korean and Mexican Spanish.

Rules angle: Voice agents talking to customers should disclose they are AI.

15 Oct 2026 · Source

Open-weight multimodal model (816B total, 42B active parameters, 1M context) under the Muse Model License.

Rules angle: Open-weight GPAI has lighter EU duties unless it reaches systemic-risk scale.

15 Oct 2026 · Source

Speech-to-speech model in the Voice Agent API at $0.08 per minute.

Rules angle: Voice agents should disclose they are AI to callers.

14 Oct 2026 · Source

Now available in the paid Gemini API, Gemini app and AI Mode for Pro and Ultra subscribers; introductory price $2/$10 per million tokens until 30 Nov.

Rules angle: Frontier GPAI model; provider bears EU systemic-risk duties.

13 Oct 2026 · Source

Multimodal models for personal and work agent tasks with memory features.

Rules angle: Agent memory of personal data raises UK and EU data protection questions.

12 Oct 2026 · Source

Small, low-cost model for high-volume tasks such as summaries and classification.

Rules angle: Check automated-decision rules (e.g. Colorado from 2027) if used to classify people.

7 Oct 2026 · Source

Image generation and editing model in the Gemini 3 family.

Rules angle: Generated images must be machine-readably marked under EU Art. 50(2).

6 Oct 2026 · Source

1-trillion-parameter (52B active) multimodal model in public preview; weights promised by end of October.

Rules angle: EU-based provider of a large GPAI model; EU AI Act GPAI duties apply.

6 Oct 2026 · Source

Generally available real-time speech-to-speech model for voice agents.

Rules angle: Voice agents talking to customers should disclose they are AI.

5 Oct 2026 · Source

Live transcription and speech-generation models.

Rules angle: Synthetic voice output may need labelling under EU Art. 50.

1 Oct 2026 · Source

New frontier model released first to vetted cyber defenders through the Fairwind Program.

Rules angle: Google cites the US voluntary pre-release access process.

30 Sept 2026 · Source

Upgrade to GPT-6 Sol announced at DevDay, aimed at agentic coding and computer use at lower prices than Astra.

Rules angle: Computer-use agents raise the agent-safety issues now under FTC and EU scrutiny.

29 Sept 2026 · Source

Faster, lower-cost Claude 5.5 model for everyday work.

Rules angle: Same EU watermarking approach as other new Claude models.

28 Sept 2026 · Source

Cheaper GPT-6 models: Sol for complex work, Luna for high-volume tasks, in ChatGPT, Codex and the API.

Rules angle: Customer-facing chatbots need AI disclosure in the EU and several US states.

22 Sept 2026 · Source

First Claude 5.5 model; Anthropic says it matches Fable 5.1 on most work at 40% lower cost than Opus 5 and was tested by external evaluators before release.

Rules angle: Watermarked output for EU AI Act compliance; zero data retention available.

22 Sept 2026 · Source

Larger base model for coding and knowledge work at the same price as Grok 4.6 ($2/$6 per million tokens).

Rules angle: Note the UK ICO's ongoing Grok investigation when using it with UK personal data.

21 Sept 2026 · Source

552B-parameter mixture-of-experts model with native visual understanding, live on the DeepSeek API; replaces V4-Flash and V4-Pro routing.

Rules angle: Chinese AI technology exports may fall under MOFCOM's draft export-control catalogue.

10 Sept 2026 · Source

New flagship model, first to a limited set of organisations, then ChatGPT paid plans, the API, Azure and Bedrock; OpenAI rates it Critical for cybersecurity under its Preparedness Framework.

Rules angle: Above the EU systemic-risk threshold for GPAI models, so the provider has extra EU duties.

3 Sept 2026 · Source

Model for agentic and coding work, available in Muse Code and the Meta Model API.

Rules angle: Check model terms before customer-facing deployment.

2 Sept 2026 · Source

Lower-cost workhorse model for coding and agent workflows, plus a cyber-focused variant.

Rules angle: Agentic use may trigger EU Art. 50 disclosure where it interacts with people.

2 Sept 2026 · Source

Fable 5.1 generally available with cheaper cache reads; Mythos 5.1 is the same model with looser safeguards for vetted cyber and life-science users.

Rules angle: Outputs carry an invisible watermark under the EU transparency code.

1 Sept 2026 · Source

Open-weight multimodal 320B-parameter MoE model (18B active); weights on Hugging Face.

Rules angle: Chinese models are subject to CAC filing rules for public services in China.

26 Aug 2026 · Source

Speech-to-text model in the Gemini API that turns raw audio into formatted text.

Rules angle: Transcribing calls may involve personal data; check consent and retention rules.

26 Aug 2026 · Source