AI This Month Review
· September 2026
THE MONTH IN ONE LINE
Top-tier AI dropped a price tier within weeks, and agents started staying on - working in the background after you close the app.
By The Numbers
The stats worth knowing this month, distilled from the data behind the headlines.
€3B
Mistral Series D raise
Mistral raised €3 billion at a valuation of over €21 billion, led by Samsung Electronics, to push open-weight AI that companies can run without handing over their data.
8h → 1h
Weekly listings check, two-person firm
ATV Big Air Tour, a two-person company, cut a weekly job of checking republished event dates from about 8 hours to 1 with a scheduled ChatGPT briefing; the figures are the customer's own.
88%
Fewer tokens for video analysis
Gemini's agentic video understanding scans only the segments it needs, cutting token use by up to 88% and costs by up to 66% for video analysis.
47%
Irish workers rethinking careers
47% of Irish workers are considering new career paths as AI reshapes work, yet only 40% say they get enough AI training.
65%
Customer calls resolved by Ringg's agents
Ringg's voice and chat agents resolve up to 65% of customer calls across more than 7 million connected calls a month.
The Core Shifts
The structural changes shaping how AI tools work, cost, and get used - drawn from patterns across the month's coverage.
Flagship ability moves down a tier
July's review covered top prices holding or falling; this month, flagship-level ability turned up in the cheaper tier within three weeks of launch. Anthropic opened September with Claude Fable 5.1 and OpenAI followed with GPT-6 Astra, both at $10 per million input and $50 per million output tokens through the API (Astra is also included in ChatGPT Plus, Pro, Business and Enterprise). By September 22, Claude Opus 5.5 performed at the level of Fable 5.1 on most tasks at $4 and $20, and OpenAI's GPT-6.1 Sol on September 29 offered intelligence comparable to Astra at significantly lower token prices. SpaceXAI pitched Grok 4.7 at $2 and $6 as twice as fast at half the price of comparable models, and Gemini 3.8 Flash arrived at the same price as the model it replaces, already switched on for Google AI Pro and Ultra subscribers in the Gemini app, AI Mode in Search and Google Sheets.
The honest limit: "comparable" is the vendors' word. OpenAI calls GPT-6.1 Sol near-Astra, not equal to it; the Grok claim rests on one benchmark, CursorBench 4.0; and Google itself warns that 3.8 Flash runs more reasoning steps, so a lower rate does not guarantee a lower bill. Anthropic notes that Opus 5.5 often suspects it is being evaluated, which makes real-world behaviour harder to judge, and most cybersecurity tasks are re-routed to an older model. The pattern is real, but the size of the saving depends on your own tasks.
Price Gap: Claude Opus 5.5 performs at the level of Claude Fable 5.1 on most tasks, at $4 and $20 per million tokens against $10 and $50.
What this means for you
If you pay per token, run one real task on Opus 5.5 or GPT-6.1 Sol before defaulting to Fable 5.1 or Astra, and compare the result and the bill. If you pay for Google AI Pro, you are already on 3.8 Flash - hand it a long task you used to split into steps.
Agents that stay on
July's review covered agents finishing jobs you handed them; this month they kept running after you left, and learned from what they did before. Meta's Muse, launched in the US on September 8, takes a goal in a chat, then opens a browser, fills in forms, books and buys, and keeps working after you close the app, returning when it needs approval; it is free for most everyday use. OpenAI's dots, from September 29, are always-on agents powered by GPT-6 Astra that work in the background across more than 4,000 apps and learn from feedback; the first dot is included in Pro and Business Premium. Microsoft's new Copilot added Autopilot, a persistent agent, and Perplexity replaced Scheduled Tasks with Automations that run on a schedule or on events in apps like Slack or Gmail and remember previous runs. Gemini now creates Docs and schedules meetings in the background from Gmail or Chat.
The honest limit: most of this is gated or early. Muse is US only, Meta's encryption that would keep Meta itself out of your agent's machine is not in this release, and interactions train Meta's models unless you opt out. Dots for Pro are not available in the EEA, Switzerland or the UK, and OpenAI says pricing for additional dots is still to come. Copilot's Autopilot is in private preview and runs on usage-based billing. Every vendor still asks you to review consequential work.
Free Agent: Meta's Muse is free for most everyday use in the US and keeps working after you close the app.
What this means for you
Before connecting an always-on agent to your email or card, decide which actions must wait for your approval - spending, sending, signing - and set that first. On Muse, opt out of model training if you do not want your interactions used.
Agents join the team
August's review covered an assistant answering in a team channel; this month vendors gave agents job titles, approval rules and audit trails. xAI launched Grok Bot for Enterprise on September 3 with access, network and audit controls for running bots at scale, and backed it with its own case studies: a procurement bot that found over $100,000 in savings, and a support operation that absorbed a 175% rise in tickets without new headcount at $0.20-$0.30 per resolution. Asana described its Claude-powered agents as teammates with scoped roles, shared memory and work everyone can see, and Anthropic showed healthcare companies using Claude Tag in Slack to triage alerts and answer payer-rule questions. By month's end, SpaceXAI, Anthropic and OpenAI all had shared team agents out, mostly in beta (compared below).
The honest limit: the strongest numbers are the vendors' reports about themselves. Grok Bots still need explicit approval before spending money, accepting terms or emailing vendors, and their messages to vendors still need human editing for tone. Every setup means handing an agent access to Slack, Drive, Gmail or contracts, so permissions become the real work. Claude Tag is in beta and not covered by Anthropic's Business Associate Agreement, so it cannot touch protected health information.
Case Study: SpaceXAI handled a 175% increase in support tickets without new headcount, with resolutions costing $0.20-$0.30 per ticket.
What this means for you
Pick one recurring team chore, such as a weekly review of software subscriptions, and write down exactly which systems an agent would need to see before you trial one. If the list worries you, scope the job smaller.
Voice AI goes real-time
Voice stopped being a turn-taking feature this month and became a live conversation layer, with releases from OpenAI, Google, Meta, SpaceXAI and Alibaba's Qwen team. OpenAI's GPT-Live-1 lets developers build agents that listen and speak at the same time ("full-duplex", like a phone call rather than a walkie-talkie) for $0.05 a minute for the voice layer. Google's Gemini 3.8 Live reached consumers through Search Live and Gemini Live, and developers at $0.005 a minute for audio input. Qwen3.8-LiveTranslate interprets across 60 languages, keeping each speaker's voice, with average lag cut to 2.3 seconds. Transcription got sharper and cheaper: Grok Voice Transcribe 2.0 is twice as accurate as its predecessor at $0.10 an hour for batch work, Meta's Muse Voice Transcribe separates more than 20 speakers, and Gemini 3.8 Flash TTS gives granular control over a voice's pacing and emotion. Qwen3.8-Omni-Flash goes further, taking in audio and video together to edit or translate clips.
The honest limit: most of this ships as developer APIs, not apps you can open today. GPT-Live-1 still needs a separate model behind it to reason and use tools; Gemini's Live Avatar is Gemini Enterprise only; Google's voice replication is unavailable in Illinois, Texas, the EEA, UK, Switzerland and India; and LiveTranslate speaks back in only 29 of its 60 languages. Expect to meet these models inside call centres, meeting tools and apps rather than as products of their own.
Live Translation: Qwen3.8-LiveTranslate interprets across 60 languages and cuts average lag to 2.3 seconds.
What this means for you
Try Search Live or Gemini Live on a real question you would normally type, and interrupt it mid-answer to see how it copes. When a support line or meeting tool offers live AI interpretation or transcription, test it on a short call before relying on it.
Unapproved agents become a security problem
The month's sharpest number came from security, not a launch. When a Fortune 500 company switched on CrowdStrike's agent discovery in August, it found 18,000 AI agents on its machines against 300 it had approved, and about 17,700 were already running before anyone looked. They were not exotic: Claude Code, OpenAI Codex, Cursor and Kiro were among them. The point that applies at any size is that an agent runs with the permissions of the person who started it, so it reaches whatever that person can reach. This is the "shadow AI" problem (explained below) arriving as software that acts, not just chats. The vendors' answer this month was layers: Microsoft merged its security tools into one operations centre in Defender where analysts and AI agents work together, Anthropic and NVIDIA combined vaulted credentials, governed actions and audit trails for Claude agents, and Perplexity described sandboxing, prompt-injection screening and user confirmation before sensitive actions.
The honest limit: nearly all the figures come from CrowdStrike itself, and it would not say how many of the 17,700 were genuinely unapproved tools rather than legitimate but unrecorded ones; its attack demonstrations were staged and blocked. The fixes are enterprise products - Falcon Guardian goes to existing CrowdStrike customers with no published price, and Microsoft's centre is a preview for E5 or E7 licences. Perplexity's own post concedes no single safeguard is perfect.
Discovery Gap: One Fortune 500 company found 18,000 AI agents on its machines against 300 it had approved.
What this means for you
Ask your team which AI tools they have installed themselves and write the list down before deciding anything. For your own agents, disconnect any account an agent does not need.
Your business tools plug into the assistant
July's review covered AI arriving inside the apps you already pay for; this month the direction reversed, and those apps started plugging into the assistant. Gemini in Google Workspace now reaches Asana, Atlassian Rovo, HubSpot, Mailchimp, QuickBooks, Monday and Salesforce from Sheets, Gmail, Docs and Chat, using MCP (Model Context Protocol - see the explainer below). The Gemini app added Connected Apps including Adobe, Airtable, Linear and Peloton. Anthropic shipped a Salesforce plugin for Claude and opened a Claude Marketplace for plugins, connectors and agents, and ChatGPT's Library added Box, Dropbox and SharePoint next to Google Drive, so files can be searched without re-uploading. The assistant is becoming the place where you ask, and your CRM, books and files are what it reaches.
The honest limit: the plumbing is uneven. Ten of the Gemini app's fourteen new connections are US-only. The Salesforce plugin is a beta for organisations approved through Salesforce's sign-up, and the Claude Marketplace is open to companies with an existing Anthropic spend commitment. In Workspace an admin has to enable and manage third-party connectors, and existing file permissions still apply in ChatGPT's Library, so you only reach what your account already reaches.
Connections: Gemini in Workspace now connects directly to Asana, Atlassian Rovo, HubSpot, Mailchimp, QuickBooks, Monday and Salesforce.
What this means for you
Pick the one business app you check most (accounting, CRM or file storage) and see whether your assistant already connects to it; if it does, ask it one question you would normally answer by opening that app.
If you could only do 3 things...
Concrete, low-effort moves you can make today - no research required.
What to do:
“Install the Gemini desktop app on your Windows PC and use Alt + Space to fact-check a document without leaving the window you are in.”
Try This Prompt
Press Alt + Space with a draft open and ask Gemini: 'Check the facts and figures in this paragraph and list anything I should verify.'
The expensive default is quietly eating AI budgets
Uber burned its entire 2026 AI coding budget by April, and its president said there was no link yet between that usage and better products. The cause is mundane and applies at any size: on team and enterprise plans the most expensive model is often pre-selected and runs at high effort by default, so it ends up doing trivial work - one engineering lead found the top model on a large context window was roughly a third of monthly spend. Almost nobody in the reporting could put a number on what the spending bought. OpenAI's new ChatGPT admin analytics, which combine usage, cost and task data, are a step forward, but they are a starting point rather than a full return-on-investment calculation.
What This Means For You
Check which model your AI tool defaults to; if it is the most expensive one, set a cheaper default and move up only for hard tasks. Before renewing a paid plan, write down the one outcome it should improve and check whether it did.
ChatGPT for Word
A ChatGPT sidebar inside Microsoft Word, launched September 17, that drafts from your notes, summarises documents, revises text and adjusts formatting without leaving the document. It is available on every ChatGPT plan, including Free.
Best For
Anyone who writes or edits in Word - writers, students and business professionals - who already uses ChatGPT and is tired of copying text between windows.
Skip If
Heavy users on tight plans: it draws on your plan's usage limits and, on some plans, the shared allowance used by Codex and other premium features. You also need to install the add-in from the Microsoft Marketplace.
Practical Action
Install the ChatGPT add-in for Word and use it on one real document this week, such as tightening a report section or summarising a long file.
Try This Prompt
In the ChatGPT sidebar in Word: 'Turn these notes into a one-page summary, then list the action items at the end.'
Head-to-Head
When multiple vendors ship the same capability, here's how your options compare.
A shared AI teammate for your team's channel or workspace
Three vendors shipped shared team agents in the last week of September; which one you can use mostly depends on the plan you already pay for.
A shared Grok Bot for a whole team's role or workflow that learns from shared context, connects to apps through plugins and keeps memories, while individual conversations stay private. Public beta on Teams and Enterprise plans.
Claude as a member of your Slack channels, now able to use each person's own connectors such as calendar, Drive or CRM for their requests; anything it posts is visible to the whole channel. Rolling out on Team plans first, Enterprise to follow; Claude Tag itself is in beta.
What This Means For You
Team already in Slack on a Claude Team plan - try Claude Tag with personal connectors; on ChatGPT Business - set up one Space for a live project; on Grok Teams or Enterprise - try one Team Bot for a single role. On none of these, wait: most of it is still beta or rolling out.
What Changed In Tools You Already Use
Updates and new features rolled out this month to the tools already in your stack.
ChatGPT
OpenAIImages 2.5 adds templates, sketch-to-image on mobile and editing by comment; plugins now work in Voice and with multiple accounts (personal and work Gmail together); a Data agent in ChatGPT Work answers questions on company data; automatic switching to Thinking was retired for Plus and Pro; GPT-6.1 Sol arrived with a $500-a-month Pro 500 plan.
Key Impacts
- ›On Plus or Pro, pick a Thinking option yourself for hard questions - ChatGPT no longer switches for you.
- ›Connect your work and personal calendars or Gmail in one chat if you juggle both.
- ›Templates in Images 2.5 are not yet in Work mode.
Claude
AnthropicCowork and chat merged into one Claude on Pro and Max, with Claude Docs and Claude Slides for co-created documents and presentations, and tasks that continue after you close your laptop.
Key Impacts
- ›No more choosing between chat and Cowork - start bigger tasks from any conversation.
- ›Set your check-in preferences so Claude asks before acting.
- ›Team and Free plans get it later.
Gmail, Docs, Meet and Google AI plans
GoogleAI Overviews in Gmail search went global for paid plans in English; Meet's note-taker adds one-page Quick notes; Docs can ground answers in your Gemini Notebook sources; Google AI plans add voice in Gmail, Docs and Keep, Sheets canvas mini-apps and the Google Pics image tool.
Key Impacts
- ›Ask Gmail's search bar a question in plain language instead of scrolling threads (English, paid plans; not personal accounts in the EEA, UK, Switzerland or Japan).
- ›Skim Quick notes after meetings and open full notes only when needed.
- ›On Google AI plans, Pics and Sheets canvas need Pro or Ultra.
Perplexity Computer
PerplexityHybrid Compute keeps sensitive files on your Mac behind a privacy gate, Portable Computer runs mostly locally on Windows PCs, and effort controls (Light to Ultra) pick the model for you.
Key Impacts
- ›Use a lower effort setting for simple tasks to spend fewer credits.
- ›Local modes need serious hardware: a Mac with 24GB unified memory, or an NVIDIA RTX GPU with 24GB VRAM on Windows.
Adobe Acrobat
AdobeAcrobat turns dense documents into interactive reports, summary slides or audio summaries, and applies Adobe Express templates to plain PDFs.
Key Impacts
- ›Turn a long PDF into a short audio summary to listen to on the go.
- ›Needs Acrobat Studio, Acrobat Express or AI Assistant Plus; review generated visuals before sharing.
Facebook, Instagram, WhatsApp and Meta AI
MetaMeta One subscriptions add more AI usage and extra features across Meta's apps, from $2.99 a month for a single product up to $499 a month for business plans; the core apps stay free.
Key Impacts
- ›Nothing changes if you stay on the free apps.
- ›Compare tiers before buying - the most AI usage sits in the higher plans, and offers vary by region.
Theme Spotlight
A closer look at the concepts and ideas worth understanding this month.
Why your assistant can suddenly reach your CRM
MCP is an open standard that lets an AI assistant discover and use outside tools while it works, instead of each connection being hard-wired in advance. For non-developers it shows up as one-click connectors that link apps like Gmail or HubSpot to a chatbot. Fixed APIs still do the steady, high-volume data moving underneath; MCP sits on top for requests that change from task to task.
Where APIs still win
Fixed APIs remain better for stable, high-volume data transfers; MCP suits requests that change from task to task.
The trade-off
MCP can use a lot of tokens on complex tasks, so a direct API call is more efficient for repetitive data work.
Try It
In an assistant you already use, connect one app you check daily and ask a question only that app can answer.
Why staff go around IT
Shadow AI is staff using AI tools their organisation has not approved. Perplexity's guide argues it is usually a supply problem: people go around IT when approved tools are rationed, capped or worse than what they can get themselves, and banning tools tends to push use onto personal devices rather than stop it.
Scale
A KPMG survey of 48,000 people found about half admitted using AI against company policy.
What helped
In one financial-services example, personal AI account use fell from 76% to 36% once official tools were offered.
Paying less for context you already sent
Prompt caching lets AI agents reuse more context from previous requests, so the repeated input costs less and responses come back faster. It matters most for long-running AI tasks.
OpenAI GPT-6
Cached input-token reads get discounts of up to 90%, with a dashboard to track how often the cache is hit.
The catch
Changing tools, settings or input can still stop the cache being reused.
Price & Access Tracker
API prices 50% lower than GPT-5.6 promotional pricing; Luna costs $0.10 input and $0.50 output per million tokens.
Input and output tokens 20% cheaper and cached tokens 60% cheaper for usage billed by the token.
$2 input and $10 output per million tokens, typically up to 30% less per task than Sonnet 5.
Off-peak cached input at $0.003 per million tokens; open weights (free to download and run on your own hardware) under the MIT licence.
Free download as open weights; Flash on the API at $0.14 input and $0.28 output per million tokens.
Five audio models (three upgraded, two new) with price cuts of up to 95%.
$0.10 per hour of audio as a limited-time offer until the end of the year.
Anyone with a Google account can create 1080p AI video scenes for free.
Free AI browsing assistant in beta, in France and North America; conversations are not saved on Mozilla's servers by default.
Available to all users globally in the Gemini app on web and mobile.
AI study hub now available globally and free for students.
What This Means For You
Building on an API? Re-price your setup this month - several models cut rates by a fifth to a half. Everyone else: make a free HD video in Google Vids or a track with Lyria 3.5 before paying for a separate tool.
AI Milestones & Watchlist
Things To Try This Weekend
Hands-on projects you can knock out in a weekend with the tools you already have.
Turn a few product or event photos into a short HD promo video for social media.
Standard Tool
Google Vids (Gemini Omni)
Requirements and Cost
Free with a Google or Google Workspace account; Workspace Business and Enterprise plans include larger generation pools.
- 1Open Google Vids with your Google account and start a new video.
- 2Add a few photos of a new product, arrival or upcoming event.
- 3Describe in plain words the scene and mood you want and generate a short clip.
- 4Adjust scene length and transitions, then export the video for your social channels.
Set up a flow that answers routine emails and notifies your team in Chat without you touching it.
Standard Tool
Google Workspace Studio
Requirements and Cost
Included in Workspace Business Starter, Standard and Plus, Enterprise Standard and Plus, and Education editions; admins can disable steps or require approval.
- 1Open Workspace Studio and start a new flow for one routine email you get often, such as a new-client or new-hire enquiry.
- 2Add a step that replies to that email with your standard instructions or next steps.
- 3Add a step that copies your template file from Drive into a folder for that person.
- 4Add a step that sends a Chat message to the colleague who needs to know, then test the flow on one email.