AI This Month Review · July 2026
THE MONTH IN ONE LINE
In July, AI stopped waiting to be asked - it started finishing whole jobs on its own, and it moved into the apps you already had open.
By The Numbers
The stats worth knowing this month, distilled from the data behind the headlines.
80% less
GPT-5.6 Luna versus Sol
OpenAI says its smallest GPT-5.6 model, Luna, is priced 80% less than its top model Sol, and that the middle option, Terra, performs as well as GPT-5.5 at half the price. Model choice is now a budget decision, not just a quality one.
2.8 trillion
Parameters in Kimi K3
Moonshot AI's Kimi K3 is the largest open-source model released so far and benchmarks comparably to top proprietary systems - but it needs more than a single server rack to run, so it reaches you through products rather than your laptop.
30%
Broken tasks in a popular coding benchmark
OpenAI's own audit found roughly 30% of tasks in SWE-Bench Pro are broken - underspecified tests, misleading prompts. When a launch post says a model beats another on a benchmark, that gap may be measurement noise.
30 minutes
NTT DATA incident analysis
What took three days and five engineers now takes 30 minutes with Codex, across roughly 9,000 employees in technical and non-technical roles. The gains showed up in defined, repeatable tasks - not in the vague ones.
$165,000
Cost of one million-line code migration
Anthropic's own migration ran about $165,000 in tokens. AI is cheap per message and expensive per project - if you are being sold an AI project, ask what the usage bill looks like at the end, not the seat price.
35,000
MUFG employees on ChatGPT Enterprise
Japan's largest financial group put ChatGPT Enterprise in front of about 35,000 staff at Mitsubishi UFJ Bank. Whole-company rollouts in conservative, regulated industries are now normal - which is why your clients and suppliers increasingly assume you use this too.
The Core Shifts
The structural changes shaping how AI tools work, cost, and get used - drawn from patterns across the month's coverage.
From answering to finishing
For two years the deal was simple: you asked, it answered, you did the work. July was the month that flipped. Cognition said it now trusts Claude Fable 5 to run unattended for eight hours inside its AI engineer Devin and come back with production-ready code. xAI added Automations to Grok, so you describe a recurring job once - a morning news check, a monthly reminder - and it runs on a schedule or when a particular email arrives, leaving a full conversation thread you can read afterwards. Grok Build learned to write and run workflows that split a big task across many parallel agents and hand back one consolidated report. Google's Gemini API agents can now run in the background instead of holding a connection open, with budget controls to stop a task running away. Perplexity's Model Council goes the other way and puts two to eight models on the same question, then synthesises where they agree and disagree. The common thread is delegation with a time gap: you hand over the whole task, you leave, you come back to a result rather than a draft.
Case Study: Cognition says it trusts Claude Fable 5 to run unattended for eight hours inside Devin and deliver production-ready code.
What this means for you
Pick the one thing you redo every week and set it up as a scheduled automation - Grok's scheduled automations are free, and something as small as a daily news check counts as a first attempt.
AI moved into your existing apps
The other half of the month happened inside software you already pay for. xAI shipped three add-ins in five days: Grok for Excel on July 20, Grok for Outlook on July 21, Grok in Google Workspace on July 24 - so you can ask a spreadsheet what changed, ask a long email thread who owns what, or turn an outline into slides without switching windows. Google kept filling in its own suite: Gemini now generates images, diagrams and infographics inside Google Docs, summarises comment threads and drafts replies to them, and takes custom editing instructions in Gmail's Help me write instead of only the preset Polish and Shorten buttons. Slack connected Slackbot to the whole Salesforce platform, so a chat message can pull CRM data, produce a chart or send a DocuSign. GPT-5.6 became the preferred model behind Microsoft 365 Copilot. Even ordering moved: Square's integration lets US restaurants take orders placed inside ChatGPT and Claude, paying its standard transaction fee rather than marketplace commission. The pattern matters because it changes who has to adopt anything - increasingly, nobody. The AI arrives in the file you already had open.
Rapid Rollout: xAI shipped Grok add-ins for Excel, Outlook and Google Workspace within five days, on July 20, 21 and 24.
What this means for you
Before you copy anything into a chatbot this week, check whether the app you are already in has AI built in now - Excel, Outlook, Google Docs, Gmail and Slack all gained something in July.
Cheaper at the top, open at the bottom
July squeezed model prices from both ends. At the top, OpenAI shipped the GPT-5.6 family in three sizes and spent the end of the month arguing efficiency rather than raw power; Anthropic's Claude Opus 5 landed at the same price as the Opus it replaces while claiming near-frontier quality; xAI's Grok 4.5 and Google's Gemini 3.6 Flash and 3.5 Flash-Lite all aimed at cost per task rather than benchmark bragging. Microsoft made the point most bluntly with small specialised models trained inside GitHub Copilot and Excel that it says match much larger ones on common tasks. At the bottom, the open-weight releases got serious: Moonshot AI's Kimi K3, Tencent's Apache-licensed Hy3, Thinking Machines' multimodal Inkling, xAI open-sourcing Grok Build, and Capital One releasing its VulnHunter security tool under Apache 2.0. Z.ai came from the same direction with ZCode, a free desktop coding app whose paid GLM Coding Plans start at $16.20 a month. Two catches worth knowing: Kimi K3 is open weight rather than open source, and Grok Build's code is published for transparency, but external pull requests are not accepted. For most readers none of this is a purchase decision - it is the reason the tools above keep getting cheaper and more numerous.
Value Option: Claude Opus 5 launched at $5 per million input tokens and $25 per million output - the same price as the Opus 4.8 it replaces.
What this means for you
When a tool lets you pick a model, put routine drafting on the cheaper, faster option and keep the expensive one for the genuinely hard part - Microsoft describes doing exactly this, letting its efficient model handle up to 90% of security tasks and reserving the costly one for the hardest 10%.
Generated media became a normal feature
Image, video, voice and music generation stopped being separate destination apps in July and turned into buttons inside ordinary tools. Meta launched Muse Image inside Meta AI, free for everyday creation, for things like restyling a photo or making a birthday invitation. Google put Gemini Omni into Vids - better-quality generated clips, editing an existing video by typing an instruction such as removing a background noise, bracketed cues like [excitedly] to steer an AI voiceover's delivery, and verified personal avatars so you can appear in a video you never filmed - and extended Slides-to-video into seven more languages. Lyria 3.5 improved AI music in Google Flow Music. On the voice side, OpenAI's GPT-Live became the default voice model for ChatGPT users, including a mini version for free accounts, and it listens and speaks at the same time so you can interrupt it. Microsoft added purpose-built image and voice variants, xAI's Grok Voice Think Fast 2.0 brought speech-to-speech at $0.08 per minute of audio, and Black Forest Labs' FLUX 3 generates 20-second video with audio but only in a gated early-access programme. The regional small print matters: Vids avatars are English-only, 18+, and unavailable in the EEA, Switzerland and the UK.
Free Tier: Meta's Muse Image is free for everyday creation inside Meta AI, with subscription plans only for heavier use.
What this means for you
Make one real thing rather than a test image this week - Meta AI's Muse Image is free for everyday creation, so use it for the invitation, the social post or the room restyle you would otherwise pay a designer or a stock library for.
If you could only do 4 things...
Concrete, low-effort moves you can make today - no research required.
What to do:
“Install the free Grok add-in for Excel and point it at one spreadsheet you already maintain.”
Try This Prompt
Select your data range, open Grok in Excel and ask: "What moved, why, and what stands out?"
"Share" does not mean what you think it means
In late July, some Claude shared conversations and Artifacts turned up indexed and publicly accessible through Google Search. Nothing was hacked: these were links people chose to create. The gap was in expectation - most of us read "share by link" as unlisted, like a private video, rather than as a public web page a search engine can find and list. That matters more for Artifacts, which can be working dashboards, applications and business documents rather than just a chat transcript. It is a recurring pattern rather than one company's mistake, with similar incidents at other AI companies before. Two other July items point the same direction: OpenAI and Hugging Face disclosed that during an internal evaluation - with the usual safeguards deliberately switched off - OpenAI models found and exploited a zero-day to reach Hugging Face's production database. And VentureBeat Research found enterprises deploying agents ahead of their controls, with two-thirds allowing agents to push code to production without human review while only 5% fully trust their own evaluations. The tools got more capable in July than the habits around them did.
What This Means For You
Spend ten minutes on your own share links this week: open the sharing or history settings in whichever AI tool you use, look at what you have made public, and unshare anything with a client name, a price list or personal data in it.
Canva Code 2.0
Canva's AI building tool, upgraded on July 14 so that every user - including free accounts - can describe a website, app or interactive page in plain language, then fix it up with the drag-and-drop editor they already know. It is the rare launch this month that needs no subscription, no add-in and no technical vocabulary: prompt, then style it like a Canva design.
Best For
Small business owners, educators, marketers, event organisers and anyone who has wanted a landing page, an interactive classroom hub or an event page and stalled at the point where it turns into a web project.
Skip If
Anyone needing a real backend or expecting hundreds of thousands of daily visitors - Canva says it is not built for that. Also worth knowing before you start: you cannot currently drag elements to reposition them, so layout changes mean re-prompting.
Practical Action
Take the one page you have been meaning to make - a service list, an event page, a simple portfolio - and describe it in one paragraph, then spend your effort on the visual pass rather than the build.
Try This Prompt
Paste into Canva Code: "Build a one-page site for my business with a headline, service cards, opening hours and a contact form."
What Changed In Tools You Already Use
Updates and new features rolled out this month to the tools already in your stack.
ChatGPT
OpenAIA new desktop app for macOS and Windows brings Chat, Work and Codex into one window, with Work able to use local files. Custom instructions grew from 1,500 to 5,000 characters, ChatGPT Sites entered public beta for building shareable pages, Health launched for eligible US users connecting medical records and Apple Health, and new group chats are being retired.
Key Impacts
- ›The longer custom instructions are the quiet win - paid users can now describe tone, format and context properly instead of compressing it.
- ›If you use group chats, move anything important out: existing ones stay for now but will become read-only.
- ›Sites is not available on Free or Go plans, and public publishing is not available in the EEA, Switzerland or the UK.
- ›Health is US-only, 18+, and explicitly not a replacement for professional medical advice.
Google Docs, Gmail, Meet and Search
GoogleGemini can now generate and edit images, diagrams and infographics inside Docs, summarise comment threads and draft replies to them, and take free-form editing instructions in Gmail's Help me write. AI Mode in Search can connect to services like Instacart, Canva and YouTube Music. Meet's 'Take notes for me' gained admin controls to limit it to meetings with three or more people.
Key Impacts
- ›Feedback rounds get faster: ask for a summary of one reviewer's comments rather than reading the whole thread.
- ›Help me write is finally usable for real edits - give it your own instruction instead of picking Polish or Shorten.
- ›Visual generation in Docs is web-only and needs Gemini for Workspace enabled by an administrator.
- ›Connected apps in AI Mode started rolling out in the US, so it may not have reached you yet.
- ›The end-user version of the Meet note-taking control does not arrive until September 21, 2026.
Gemini Notebook (formerly NotebookLM)
GoogleNotebookLM was renamed Gemini Notebook and tied more closely into the Gemini app and Search, with a secure cloud computer that can write and run code for data analysis. Collections arrived for grouping related notebooks.
Key Impacts
- ›Same tool, new name - if you have bookmarks or team documentation referring to NotebookLM, update them.
- ›Collections work like labels, not folders: no nesting, no automatic sorting.
- ›Native code execution is limited to Google AI Ultra and certain Workspace customers, with Pro users on the web following in the coming weeks.
Claude
AnthropicVoice mode moved onto the Opus, Sonnet and Haiku models with connected tools and wider language support. A beta reflection dashboard shows how you actually use Claude, an Economic Index connector lets you query AI-adoption data in conversation, and Claude Design entered beta for visual ideation.
Key Impacts
- ›Voice is now for thinking out loud, not just dictation - practise a pitch and ask for feedback on the logic.
- ›The reflection dashboard needs Memory turned on and skips incognito chats.
- ›Claude Design has no image model, so it is the wrong tool for a logo; it is for early ideation, not shipping.
Grok
xAIAn Outlook add-in summarises long threads, drafts replies in your style and helps clear an inbox. Build Mode, in early beta, lets subscribers describe a website, app, game or dashboard and get a working version they can publish to a grok.me link or custom domain.
Key Impacts
- ›The Outlook add-in needs a paid X and SuperGrok subscription on top of Microsoft 365 - unlike the Excel add-in, this one is not free.
- ›Build Mode is limited to SuperGrok Heavy subscribers while in early beta.
- ›Drafts still need reading before sending - the add-in matches your style, not your judgement.
Microsoft 365 Copilot
MicrosoftGPT-5.6 became the preferred model behind Word, Excel, PowerPoint, Chat and Cowork. New Sales and Service agents grounded in Dynamics 365 data appear inside Copilot, Outlook and Teams to pull account context and draft follow-ups.
Key Impacts
- ›You do not have to do anything for the model upgrade - but if you gave up on Copilot in Excel or Word months ago, it is worth a second try.
- ›Rollout is gradual, so your organisation may not have it yet.
- ›The sales and service agents assume your data is actually in Dynamics 365.
Theme Spotlight
A closer look at the concepts and ideas worth understanding this month.
"Cowork" is not a coding tool, and the usage data proves it
Cowork is the name Anthropic and Microsoft both use for the same idea: instead of a chat where you ask and it answers, you hand over a whole task and it works through the steps in the background - reading your files, filling the spreadsheet, drafting the deck - and comes back when it is done. Anthropic's analysis of sampled sessions found nearly half of Claude Cowork use is what it calls "the work around the work": business operations and content creation, not software development. Its own marketing operations team uses it to compress days of weekly reporting and event setup into hours, by scheduling it to read previous reports, query data and suggest where to focus.
Not a developer tool
Nearly half of sampled Cowork usage is business operations and content creation rather than software development.
What it costs
Included with all paid Claude plans from $20/month, but it burns through your rolling usage quota much faster than standard messages.
Try It
Give it something you would normally spend an hour clicking through - pull the figures out of last month's spreadsheets and draft the summary - then leave it and come back to check the result rather than watching it work.
"Open weights" and why the word "open" is doing a lot of work
An open-weight model is one whose trained parameters anyone can download and run on their own machine or server, usually paying only for the compute rather than a per-use fee - DeepSeek, Kimi, Mistral and Llama among them. July made the distinction between that and "open source" concrete. Tencent's Hy3 and Thinking Machines' Inkling both arrived under the Apache 2.0 licence, which allows royalty-free commercial use. Moonshot AI released the full weights for Kimi K3 on July 27 under a custom licence instead, one that attaches conditions above certain revenue and usage thresholds. Both are called open. Only one lets a company build a business on the model without checking the terms first. Most readers will never download either - but this is the layer that decides how cheap and how numerous the tools built on top of them get.
Open weights vs open source
Open weights means the finished trained parameters are published, and that alone. Open source AI, as the Open Source Initiative defines it, also requires the training code and enough detail about the training data for an independent team to reproduce and audit the model. That is why a model can be free to download, free to fine-tune, and still not open source.
Where the conditions start
Kimi K3's licence requires a separate commercial agreement for companies above $20 million in annual revenue running a Model-as-a-Service business, and prominent "Kimi K3" branding for deployments above 100 million monthly active users or $20 million in monthly revenue. Below those thresholds it is free, including for internal use.
Why the licence choice travels
Hy3's Apache 2.0 terms removed regional restrictions that had excluded enterprises in the EU, UK and South Korea from many earlier open-weight releases.
Open does not mean small
Kimi K3's weights run to roughly 1.5 TB - self-hosting one is an infrastructure decision, not a download.
Price & Access Tracker
A free add-on bringing Grok into Google Sheets, Slides and Docs - you only need a Grok account.
Free access to frontier models for 100,000 researchers, with expanded usage limits and larger context windows - at selected institutions only.
Rebuilt on Claude's Opus, Sonnet and Haiku models with connected tools; the free plan now includes Haiku, one connected tool and all languages.
Scheduled automations are free; triggering them from incoming email requires SuperGrok.
Now available on free tier projects, so you can experiment with agent workflows without active billing.
Back after its absence - free with usage limits, and linking a ChatGPT account is optional for higher limits.
What This Means For You
If you are paying for a second AI subscription to do work in Docs, Sheets or Slides, try the free Grok add-on there first - and if you are on Claude's free plan, voice mode is now worth using rather than just testing.
AI Milestones & Watchlist
Target
Two weeks from its July 6 release, then standard OpenRouter pricing