This Week's Releases: Apple's Siri AI and iPhone Duo, OpenAI's Agents API, Meta's Muse and DeepSeek V4.1 Flash
Updated: 6 days ago
Title: This Week's Releases: Apple's Siri AI and iPhone Duo, OpenAI's Agents API, Meta's Muse and DeepSeek V4.1 Flash
Date: 10 September 2026
Type: Blog
Author: SAASiQ (contact@saasiq.ai)
Word count: 1517 words
Reading time: 6 min
Published: 10-09-2026
Apple launched the iPhone 18 Pro and its first foldable iPhone on 9 September, and announced Siri AI, which goes into beta on 14 September. OpenAI released a set of tools for developers and banks on 10 September, Meta launched a personal AI agent in the US, and DeepSeek released an open-weight model priced at a small fraction of the leading US models. Google also brought its Gemini app to Windows.
iPhone 18 Pro and iPhone Duo
The iPhone 18 Pro and Pro Max run on Apple's A20 Pro chip, built on a 2nm manufacturing process. It has a 6-core processor, a 7-core graphics unit and a dual 16-core Neural Engine, the part of the chip that handles AI work, which Apple says doubles AI processing compared with last year's A19 Pro. The main camera is 48MP with a variable aperture, meaning it can change how much light it lets in, a first for iPhone, and the phones use Apple's own C2 modem. Storage runs from 256GB to 2TB. Launch coverage gives US prices of $1,199 for the Pro and $1,299 for the Pro Max. Pre-orders open on Saturday 12 September and the phones are in stores from Friday 18 September in more than 65 countries.
The iPhone Duo is Apple's first foldable phone, with a 5.4-inch screen on the outside and a 7.6-inch screen inside, and the same A20 Pro chip. It costs $1,999 in the US and £1,999 in the UK for 256GB. GB News and UK retailers list the larger models at £2,199 for 512GB, £2,599 for 1TB and £3,199 for 2TB. Pre-orders open on 16 October and it goes on sale on 23 October in more than 70 countries, including the UK. John Ternus, Apple's chief executive, called it 'the most transformational change to iPhone since the original'.
Siri AI
Siri AI is the rebuilt Siri that Apple announced at the event. Its beta starts on 14 September in English, with French, Japanese, Korean, Portuguese and Spanish following in October. Apple lists three main abilities: personal context, meaning it can draw on what the phone knows about its owner; on-screen awareness, so it can act on whatever is being looked at; and actions inside apps across the system. It runs partly on the device, using Apple's own Foundation Models, and partly on Apple's Private Cloud Compute servers, 'in collaboration with Google's Gemini models'.
It works on the iPhone 15 Pro and later, and on Macs and iPads with an M1 chip or newer. Apple's announcement says Siri AI is not initially available in the EU or China, and it does not list UK English specifically, so UK availability is something to check on apple.com/uk before planning around it. For IT teams that manage company iPhones, the device list is the useful part, since it decides which staff phones will offer the feature once it arrives.
OpenAI's releases on 10 September
OpenAI released three products aimed at businesses. The first is the Agents API, in public beta. It turns the harness behind OpenAI's Codex coding tool into a hosted service: a harness is the software around a model that lets it work through a long task, and this one handles orchestration, long-running sessions, context compaction (condensing earlier work so a long task fits in the model's memory), subagents that take on parts of a job, and sandboxes, the isolated computers where an agent runs code. There is no fee for the API itself. Customers pay for tokens, tools and sandbox time. SDKs are available for Python, TypeScript and JavaScript, Go, Java and Ruby. While it is in beta, data is held only in the US and Zero Data Retention is not supported, which matters for UK organisations and public bodies with data residency commitments.
The second is GPT-Live-1, a voice model that listens while it speaks, so a caller can interrupt it as they would a person. It costs $0.05 per minute of voice, billed by the second, with the reasoning model behind it billed separately. OpenAI reports 80.1 per cent on the Full Duplex Bench v1.5 test, a turn-taking delay of 0.798 seconds and 87 per cent success at calling tools. The independent Artificial Analysis Speech-to-Speech Index scores it at 69.8 per cent, seventh of 11 models.
The third is ChatGPT for Financial Services, built on GPT-6 Astra for ChatGPT Enterprise. Institutions have to apply for it, and Morgan Stanley and Evercore were the design partners. It comes with data from Daloopa, PitchBook, LSEG News and Crunchbase built in, connectors for Bloomberg, FactSet, S&P and MSCI, and about 50 connectors in total using MCP, the standard way of linking AI assistants to other software. It produces pitchbooks in a firm's own templates, Excel models and dashboards, and compliance teams can export the workspace logs. Nick Turley, head of ChatGPT, said: 'This is the canonical product we are hoping the industry adopts.' Pricing was not disclosed.
Meta's Muse
Muse is a personal AI agent from Meta, launched on 8 September in the US for adults. It runs on the web at muse.ai, on iOS and Android and inside WhatsApp, and Meta says its AI glasses will follow. There is a free tier, a Power tier at $20 a month and a Maximum tier at $100 a month. Muse can send email, book travel, fill in forms, shop using a one-time card number and lower bills, and it connects to email, calendars, payments, health, smart home and other services.
Each user's agent runs in its own 'Muse Secure VM', a separate virtual computer, watched by a second agent Meta calls Sentinel, and risky actions need the user's approval through a confirmation card. Meta says Muse cannot access passwords or payment methods and 'doesn't share people's conversations or data with Meta's ads systems'. It is not available in the UK.
ChatGPT Images 2.5
OpenAI's image tool was updated on 8 September for all ChatGPT, ChatGPT Work and Codex users on every tier. A new @Sketch feature lets users draw a rough outline for the image they want, and OpenAI says images arrive up to 50 per cent faster than with Images 2.0. There are templates for posters and merchandise, comments can be added directly on an image, and transparent backgrounds are cleaner. Images carry C2PA metadata, a standard record inside the file showing it was made by AI.
Developers get two new models through the API: GPT-Image-2.5 Flare, the default, and Sunburst, for precise editing. Both cost $8 per million image-input tokens ($2 cached), $30 per million image-output tokens and $5 per million text-input tokens.
DeepSeek V4.1 Flash
DeepSeek released V4.1 Flash on the night of 9 September, according to VentureBeat. It is a mixture-of-experts model, built from many specialised parts of which only a few are used for each request: 552 billion parameters in total, with 8 billion active while it reads a prompt and 16 billion while it writes the answer. It can read images, handles 1 million tokens of context, and is released under the MIT licence with its weights on Hugging Face, so anyone can download it and run it on their own hardware.
Off-peak it costs $0.15 per million input tokens, $0.003 for cached input and $0.60 per million output tokens. OpenAI's GPT-6 Astra, released the week before, costs $10 and $50. DeepSeek charges double during its peak hours, weekdays from 01:00 to 04:00 and from 06:00 to 10:00 UTC, and the second window is 07:00 to 11:00 in the UK while clocks are on British Summer Time.
DeepSeek's own benchmarks put it at 74.2 on DeepSWE v1.1, a software engineering test, against 74.0 for Anthropic's Claude Opus 5 and 73.0 for GPT-5.6 Sol. On Terminal-Bench 4.0, which tests work at the command line, it scores 31.2 against 51.8 for Opus 5. DeepSeek is a Chinese company, and public bodies should check the hosted service against their own procurement and security guidance before using it. The open weights can instead be run on infrastructure the organisation controls.
Gemini for Windows
Google released a Gemini desktop app for Windows 10 and 11 on 10 September, for users aged 18 and over, available globally with some regional differences. Pressing Alt and Space opens Gemini over whatever window is in use. It can pass multi-step tasks to Google's Gemini Spark agent and pull information from Gmail and Drive, and some features, including Spark and Omni, need a Google AI subscription. It downloads from gemini.google/desktop.
Before switching
Apart from the Artificial Analysis score for GPT-Live-1, the benchmark figures above come from the vendors. The more useful test for a business is its own: a sample of real tasks run through the current tool and the new one, comparing the output and the bill.
On managed Windows estates, the Gemini app is one more AI tool staff can install from the web, and whether it is allowed is a decision for the organisation's application control list. The Agents API's US-only data residency applies for as long as it stays in beta.
Siri AI's beta starts on 14 September, and iPhone Duo pre-orders open on 16 October.
SAASiQ - Intelligent Solutions for SaaS ©
