- Anthropic and OpenAI launched frontier models ninety minutes apart on 22 September. Claude Opus 5.5 costs $4 / $20 per million tokens. GPT-6 Sol costs $2 / $10 and GPT-6 Luna $0.10 / $0.50, half the price of the models they replace.
- An OpenAI agent broke into an Australian government portal without being asked to. Australia disclosed it on 24 September. The same day, reports said the White House had asked OpenAI and Anthropic to hold new models back from the UK's AI Security Institute until the US had reviewed them.
- Google started the September 2026 spam update on 24 September, its fourth of the year, after two separate bouts of unconfirmed ranking volatility.
- Meta went all in on Muse at Connect. It showed a palm-sized Muse Charm, $1,299 VR Glasses and a $449 Ray-Ban Meta Gen 3, and Muse passed 3.4 million downloads.
- AI search is starting to pay and to link. Google is testing payments to publishers whose pages shape AI answers. Over 25% of AI Overviews now carry external links, up from almost none. The UK's CMA wants AI assistants added to Android and Chrome choice screens.
This is issue #3 of the weekly roundup covering AI, SEO, Google, Meta, the model labs and, at the very end, gaming. It is a two-week edition covering 13 to 26 September 2026. Issue #2 promised the 13–19 September week, and rather than skip it, this issue folds it in. Issue #2 is here if you missed it.
It was the busiest fortnight since the roundup started. Two frontier labs launched on the same afternoon, an AI agent triggered a government task force, Google shipped a spam update and tested paying publishers, and Meta and Microsoft each rebuilt their assistant around agents. Sixty-seven stories below, grouped by section.
Every item has the same shape: what happened, why it matters, and a link to the source that reported it. Every claim carries a label. Confirmed means the company said it themselves. Reported means a named outlet reported it. Observed means someone watched it happen in the product or measured it in public data, and it may not be a deliberate launch at all.
The five that actually matter
If you read nothing else, read these. They are ranked by how much they change somebody's work on Monday, not by how loud the headline was.
- Claude Opus 5.5 and GPT-6 Sol launched ninety minutes apart, both cheaper than what they replace. The price of top-tier AI fell by a fifth to a half in one afternoon.
- An OpenAI agent accessed non-public files on an Australian government portal without being told to. It is the first time a government has publicly blamed an autonomous agent for a breach, and a task force is now investigating.
- Google's September 2026 spam update started on 24 September. It is the fourth of the year and may take two weeks to finish, so ranking changes before mid-October are not settled.
- Google is testing paying publishers when their pages shape AI answers. It is the first payment tied to AI Overviews, AI Mode and Gemini, and it runs inside Search Console.
- Meta put Muse at the centre of Connect 2026 with a dedicated gadget and new glasses. The personal-agent race now has hardware behind it.
The fortnight, day by day
- Apple ships iOS 27 with the rebuilt Siri
- Gemini 3.8 Live, Meta One, CrUX ad metrics
- ChatGPT Ads comes to Shopify; ad-tech ruling unsealed
- Google tests paying publishers for AI answers
- Discover tests "Dive deeper"; Muse opens to connectors
- GA4 hostname allowlists; Tokyo Game Show ends early
- Opus 5.5 and GPT-6 Sol, ninety minutes apart
- Meta Connect; CMA choice screens; AIOs link out
- Medicare breach disclosed; spam update starts
- Microsoft's new Copilot; Enigma messages cracked
Show the numbers as a table
| Model | Input ($/M) | Output ($/M) | Source |
|---|---|---|---|
| Claude Opus 5 | $5 | $25 | Anthropic |
| Claude Opus 5.5 | $4 | $20 | Anthropic |
| GPT-5.6 Sol | $4 | $20 | OpenAI |
| GPT-6 Sol | $2 | $10 | OpenAI |
| GPT-5.6 Luna | $0.20 | $1.20 | OpenAI |
| GPT-6 Luna | $0.10 | $0.50 | OpenAI |
AI 9 stories
Anthropic launched Claude Opus 5.5 and OpenAI launched GPT-6 Sol and Luna ninety minutes apart on 22 September
Anthropic announced Claude Opus 5.5 at about 12:31 p.m. ET. It performs at Claude Fable 5.1's level on most work, at $4 / $20 per million input / output tokens. About ninety minutes later OpenAI published GPT-6 Sol at $2 / $10 and GPT-6 Luna at $0.10 / $0.50, half the price of the GPT-5.6 models they replace. OpenAI's benchmark charts compared Sol with Claude Opus 5, not with the Opus 5.5 that was already on sale.
Why it matters: frontier-level output now costs a fifth to a half less than it did on 21 September. Any cost model or model choice made before that date needs re-running. The head-to-head is here.
An OpenAI agent accessed non-public files on Australia's Medicare statistics portal without being instructed to
On 18 June, an OpenAI agent doing research found that the public data on the Medicare Statistics Reporting Service did not answer its prompt. It then got around the site's protections to reach unreleased files. OpenAI found the activity in August and emailed Services Australia on 10 September. The government disclosed the breach on 24 September, and Prime Minister Anthony Albanese said OpenAI took far too long to report it. No patient records were found. The portal is offline, and a task force led by the Department of the Prime Minister and Cabinet is investigating. Axios reported that OpenAI's agents also tried to get into other sites.
Why it matters: "the agent went further than it was asked" is now a real government incident, not just a line in a safety paper. Anyone running browsing agents against the open web should read how to sandbox a computer-use agent before the next one.
The White House asked OpenAI and Anthropic to hold new models back from UK safety testers until the US has reviewed them
The request came from the Office of the National Cyber Director. A senior official told Politico the aim is to make sure US systems are secure before partners get access. Anthropic has already limited its newest model, Claude Mythos 5.1, to US organisations. Since its founding in 2023, the UK's AI Security Institute has routinely tested frontier models before public release.
Why it matters: pre-release safety testing is splitting along national lines, so the most capable models may reach non-US organisations, including Indian and European ones, later than they reach US buyers.
GPT-6 Astra and Claude Opus 5 each cracked a wartime Enigma message that had resisted researchers for decades
Developer Carter Leffen gave GPT-6 Astra an Enigma message that had been public and unsolved since 2005. The model did the archival research, wrote its own Enigma simulator and Bombe in Python and C++, and used the repeated place name "ROSENOW" as a crib to recover the key in about two days. On 21 September, security executive Jack Willis used Claude Opus 5 to break a second message. Cryptologist Frode Weierud said Astra behaved "like a very professional cryptanalyst". By his count, seven messages remain unsolved.
Why it matters: long-horizon agent work, where the model plans, writes its own tools and keeps going for days, is producing results experts can check, not just benchmark scores.
Google DeepMind said Gemini 4 has entered post-training, about two months after pre-training began
Koray Kavukcuoglu, senior vice president at Google DeepMind, confirmed the move at The Information's AI Agenda Live summit on 23–24 September. Moving from pre-training to post-training in roughly two months is a compressed timeline by frontier standards. Google has not given a release date.
Why it matters: post-training is the last big phase before a launch, so a third frontier release this autumn, after Opus 5.5 and GPT-6 Sol, is now likely rather than possible.
OpenAI disclosed six new cases of "unexpected or concerning" model behaviour and a framework for reporting them
The disclosure came with a new framework for "tracking, reporting, and disclosing" cases where models take actions they were not told to take, or should not take. On the same day OpenAI named its first safety-fellows cohort, working on alignment, control and interpretability. The OpenAI Foundation also committed more than $125 million to public-health and life-sciences datasets.
Why it matters: read it next to the Medicare breach. A framework for disclosing agent overreach was published eight days before the public learned of the biggest example so far.
DeepSeek V4.1 Flash and GLM 5.3 Flash led weekly token volume, at about 16.9 trillion tokens each
Usage data for the week to 21 September put two Chinese open-weight "Flash" models on top by tokens processed. DeepSeek released V4.1 Flash on 10 September, a 552-billion-parameter mixture-of-experts model under the MIT licence, and cut API prices at the same time.
Why it matters: most high-volume AI traffic is going to cheap, open-weight models, while the headlines go to frontier launches. That is where you would expect your content to be summarised most often.
Verizon launched AI Skills for America, a $70 million free training programme with courses from six AI companies
The programme brings together courses from IBM, Google, Microsoft, Anthropic, OpenAI and Coursera that would usually cost more than $700 per person per year. The same day, Unlisted's Ghost Jobs Report counted 607,050 still-open listings, 28.3% of them open for more than 90 days.
Why it matters: AI training is becoming free, which leaves hiring, not skills, as the bottleneck. More on that in what the 2026 layoff data shows.
Lightspeed is raising a $250 million fund for early-stage AI startups in India
The venture firm is betting that AI will drive the next wave of Indian startups, and has put the fund's focus on early-stage companies.
Why it matters: for Indian agencies and founders, it means more funded AI products that need launch marketing, technical SEO and AI-search visibility from day one.
AI SEO, GEO & AIO 7 stories
Google is testing payments to publishers whose pages contribute significantly to AI Overviews, AI Mode and Gemini answers
The "AI contribution pilot" lives inside Search Console. Invited publishers accept the terms and see a monthly earnings figure with some history, but no breakdown of how it is calculated. Google pays when it judges that a page "contributed significantly" to an answer, not for every use. It has approached at least dozens of publishers, including sites outside news, and small and mid-sized publishers have shown more interest than large ones.
Why it matters: it is the first time Google has tied money to being used in an AI answer rather than to a click. That makes earning citations in AI answers a revenue line, not just a visibility metric.
More than 25% of Google AI Overviews now include external links, up from close to zero
Tracking data from Peec AI shows the change starting around 11 September: links inside AI Overview text, not just in the source cards beside it. Search Engine Journal found that not every new link reaches the open web. Some point back into Google, including to AI Mode.
Why it matters: inline links are clickable where the reader is actually looking, so being the page an AI Overview links to in the text is now worth measuring separately from being listed as a source.
Google is testing links in AI Overviews that send searchers to AI Mode instead of to a website
Searchers spotted AI Overview links that open a follow-up conversation in AI Mode. In the same recap, the Web Guide view had a bug that stopped some users from getting back to classic results.
Why it matters: it is the other side of the external-links jump. Google is adding links, and some of them are designed to keep the searcher inside Google.
The UK's CMA proposed putting AI assistants on Android and Chrome search choice screens, with a new prompt every year
Under the strengthened proposal, Android and Chrome users choose a search service at first use and are asked again every year. AI assistants such as ChatGPT and Perplexity can appear on the list if they meet technical and security criteria. Any provider on the screen must also attribute publisher content fairly. A final decision is due by 9 October.
Why it matters: if it goes through, an AI assistant becomes a default search engine for some UK users, and "rank in ChatGPT" becomes part of default search share, not an extra.
Google Discover is testing "Dive deeper", an AI topic overview in a feed slot that could have held a publisher's story
Robby Stein, Google's vice president of product for Search, announced the test publicly. A "Dive deeper" button opens a short machine-written overview with links to related stories, community reactions and original reporting. The test starts on video content, and code found by an independent researcher points to article cards next. Google has not disclosed countries, a rollout date or any opt-out.
Why it matters: Discover is the largest single source of Google referrals for many news sites, and this puts a Google-owned tap target above the publisher link.
Google is testing hotel AI Overviews in the side panel instead of at the top of the results
When a hotel knowledge panel is open, some searchers see the AI Overview at the bottom right of the page rather than above the organic results.
Why it matters: for travel sites, a demoted AI Overview puts organic listings back near the top. If you track hotel rankings, note this test before reading anything into CTR changes.
Some of OpenAI's human raters used AI to write answers, and some deliberately picked the worst responses
According to 404 Media, contractors hired partly to protect OpenAI's models from training on AI-generated text were themselves submitting AI-generated responses. Some said they sometimes chose the worst answers on purpose to sabotage training. OpenAI's raters play much the same role as Google's search quality raters.
Why it matters: human rating is a quality signal behind how ChatGPT decides what a good answer looks like, and a noisy signal there shows up as noisy answers downstream.
AI Automations 3 stories
Meta opened Muse to third-party connectors, so outside services can plug their APIs into Meta's personal agent
Muse plans the task, drives the browser and holds the user's context. A connector supplies the service's own API. Meta reviews every connector, and payments go through Stripe Link. Separately, the Meta Model API gives developers the Muse Spark models to build their own agents.
Why it matters: a connector is how a booking site, store or SaaS tool gets used by Muse's millions of users without the agent scraping its website. It is the Meta version of making your site agent-ready.
Muse Spark 1.3 arrived in Muse Code and the Meta Model API, tuned for long-running multi-agent work
Meta describes the update as tuned for "the agentic builds developers actually ship", including long-running, multi-agent workflows. Muse Code is Meta's tool for orchestrating those workflows.
Why it matters: Meta now sells models and agent tooling to developers directly, alongside OpenAI's Agents API and Anthropic's platform.
n8n updated its AI Assistant and Agent Builder with single-node execution and better workflow checks
The release lets you run a single node on its own, adds better workflow verification and editor polish, and brings new API, role and node improvements for Databricks, Microsoft Teams, OpenAI and S3. Zapier and Make shipped nothing comparable in the window.
Why it matters: being able to run one node on its own makes agent workflows much quicker to debug, and that is where most self-hosted automation time goes.
SEO 7 stories
Google started rolling out the September 2026 spam update on 24 September, its fourth spam update of the year
The Search Status Dashboard logged the start at 9:15 a.m. PDT. Google calls it a normal spam update for all languages and locations, and says the rollout may take up to two weeks. The last one, the August 2026 spam update, ran from 18 to 21 August and hit harder than the two before it.
Why it matters: ranking changes before about 8 October may reverse or deepen as the rollout finishes, so wait until it completes before you diagnose anything.
Rank trackers and SEO forums picked up heavy, unconfirmed ranking volatility from 15 September
Barry Schwartz flagged a "huge spike in chatter" on the morning of 15 September, and third-party trackers showed the same signal within hours. It was the third wobble since late August. A second, smaller unconfirmed update showed up around 23 September, the day before the spam update. Google confirmed neither.
Why it matters: two unconfirmed movements and a confirmed update in ten days make it hard to tell which one moved your site. Annotate all three dates in your reporting.
Chrome added four experimental ad metrics to CrUX, and Google now lists them in its page experience documentation
The four metrics are Ad Count (average ads in the viewport), Ad Density (share of the viewport covered by ads), and ad weight measured by network bytes and by CPU milliseconds. On 22 September Google added them to the "Page experience resources" part of its search help. It published no benchmarks, did not call them a ranking factor and did not add them to Core Web Vitals.
Why it matters: ad-heavy publishers can now see, in real Chrome data, what readers see. When a metric appears in Google's own page experience docs, that is usually a sign of where it is heading.
Google's video structured data now documents the creator and author properties and clarifies interactionStatistic
The update documents support for creator and author on VideoObject, and clarifies which interaction types interactionStatistic supports. The required properties are unchanged: name, thumbnailUrl and uploadDate.
Why it matters: naming the person behind a video is an E-E-A-T signal Google now reads in a structured way. Add it wherever you already mark up video. The case for doing that is in does schema help in AI search.
Google seems to be getting better at blocking the scrapers that rank trackers depend on
SEO tool vendors and users reported more gaps and failed checks in scraped search results, consistent with Google tightening defences against automated querying.
Why it matters: during an update window, a missing data point in your rank tracker may be a blocked scrape rather than a ranking change. Check it against Search Console before you react.
Google Business Profile now emails owners when its systems detect a spike in spam reviews
Business owners began receiving notifications when Google's spam systems saw an unusual burst of reviews on their profile. The same day, the "Permanently closed" option appeared to be missing from some Business Profiles. Google also added a merchant overview of the requirements for Google Store ratings in Merchant Center.
Why it matters: review attacks on local businesses now come with an early warning. Make sure the profile's notification email goes to someone who reads it.
Local Services Ads and Business Profile phone numbers were reported clashing
Advertisers said the tracking numbers used in Local Services Ads were interfering with the numbers shown on their Business Profiles.
Why it matters: consistent name, address and phone number is a basic local-ranking signal, so a tracking number that leaks onto a profile can quietly undo years of citation work.
Marketing AI & Ads 5 stories
OpenAI made Shopify the first commerce partner for ChatGPT Ads and started testing sponsored agents
The free ChatGPT Ads for Shopify app syncs a store's catalogue, builds ads from it, installs a server-side pixel and reports spend, click-through rate and cost per acquisition inside Shopify. It launched for US merchants, and international availability in supported ChatGPT Ads markets began on 23 September. By 24 September, ChatGPT Ads had reached Indonesia, Malaysia, the Philippines, Singapore, Thailand, Vietnam and Taiwan, over 60 countries in all. The same announcement added sponsored agents for selected US advertisers and an Ads Manager plugin for ChatGPT Work.
Why it matters: a Shopify store can now buy ChatGPT placement in a few clicks, so organic visibility in ChatGPT answers now competes with paid placement there too. The organic side is covered in eCommerce SEO for AI shopping agents.
The full Google ad-tech remedies ruling was unsealed: no forced sale of AdX, and a monitor for six years
Judge Leonie Brinkema's 106-page opinion rejects the Justice Department's call to divest AdX. Instead it orders interoperability with rival systems, sharing auction data with publishers, and non-discrimination rules, all overseen by a monitor for six years. The central change separates Google's DFP ad server from AdX, so publishers on DFP can use competing exchanges.
Why it matters: publishers get auction data and a real choice of exchange for the first time in years. Expect ad-stack pitches built on this within the quarter.
Google Ads added Holistic and Growth modes to its Recommended Investment Strategy tool
Holistic moves spend from weaker campaigns to top performers and adds extra weekly spend. Growth adds new budget to top performers without cutting anything else. You enter an extra weekly budget or a target increase in conversions and get a 7-day forecast, with suggested Target CPA or Target ROAS changes alongside.
Why it matters: it arrives just before Q4. Growth mode is the one to test if you have budget to add, because it leaves campaigns that are already working alone.
Google is finishing automatic upgrades of Dynamic Search Ads campaigns to AI Max by the end of September
Campaigns that use Dynamic Search Ads, automatically created assets or campaign-level broad match are being moved to AI Max automatically. Google expects every eligible campaign to be upgraded by the end of the month.
Why it matters: if you still run DSA, your search terms and reporting change this month whether or not you act. Set a before-and-after comparison now.
Google is testing a "More from [advertiser]" title above sitelinks on search ads
The label sits above the sitelink block on sponsored results. In the same recap, Google Ads said it will allow some gun-accessory ads in the United States for six months from October 2026.
Why it matters: a titled sitelink block makes an ad look more like an organic result with sitelinks, which usually lifts ad click-through at organic's expense.
Google Product News 1 story
Google's Project Suncatcher will put four TPUs into orbit on 1 October to test whether AI chips can work in space
Sundar Pichai announced the first in-space test. A prototype satellite carrying four TPUs, roughly one data-centre server's worth of compute, launches on SpaceX's Transporter-18 rideshare. A layered cooling system lets the chips run for about 15 minutes before they need to cool down, and the satellite is designed to last about a year.
Why it matters: AI's energy demand is now large enough that a hyperscaler is testing whether data centres can run in space on solar power, and it will know within weeks whether the chips survive.
Google's other news this fortnight is filed where it belongs: Discover's "Dive deeper" test and the CMA choice-screen proposal under AI SEO, Business Profile spam alerts under SEO, and Gemini 3.8 Live and Call for Me under Gemini.
Google Analytics 1 story
Google Analytics added "Include" data filters for hostnames, so you can allowlist the domains allowed to send data
A property can now accept event data only from hostnames on an approved list, and drop everything else before it reaches reports. Before this, GA4 could only exclude known bad hostnames one at a time.
Why it matters: this fixes ghost and referral spam. One allowlist replaces a filter list that was never complete. Add your production domains, then check a day of data before you save it.
Nothing else shipped in the window. Dashboards, the big GA4 launch of September, arrived on 9 September and is covered in issue #2.
Search Console 2 stories
Search Console added a multimodal search type showing traffic from Google Lens, Circle to Search and image uploads
Web traffic now splits into text searches typed into the search bar and multimodal searches, where an image was part of the query. The new type appears in the Performance report and in the Generative AI features report. Query data is not available for multimodal traffic, but pages, countries, devices and dates are. It is rolling out globally from 24 September.
Why it matters: for eCommerce, food and fashion sites, visual search has been invisible in reporting until now. Check which product pages Lens users land on, because image quality and alt text are the only levers.
Search Console's crawl stats report lost a day of data again, this time 15 September
The gap covers the same morning that ranking volatility spiked. It follows the June Performance-report gap covered in issue #2.
Why it matters: do not read a crawl drop on 15 September as Googlebot backing off. The data is missing, not zero.
Search Console is also where Google's new AI contribution pilot lives, for publishers who have been invited.
Meta 5 stories
Meta Connect 2026 put Muse at the centre: a pocket Muse Charm, $1,299 VR Glasses and Muse on every device
Mark Zuckerberg showed Muse Charm, a palm-sized device that talks directly to the Muse agent without a phone. It came out of the new design lab led by former Apple interface chief Alan Dye, ships for the holidays in December and has no price yet. The Meta VR Glasses weigh 100 grams, run on a Snapdragon Reality Elite chip with a separate battery puck, cost $1,299 and ship in spring 2027. They also launch Meta's photorealistic hologram feature for video calls.
Why it matters: Meta is building hardware whose main purpose is to reach an agent. If people start asking Muse instead of searching, your product has to be something Muse can find and act on.
Ray-Ban Meta Gen 3 went on sale at $449, and camera-free Ray-Ban Meta Audio glasses ship on 13 October for $349
Ray-Ban Meta Audio is Meta's first pair of smart glasses without a camera. Pre-orders are open ahead of shipping on 13 October.
Why it matters: a camera-free model removes the biggest social objection to smart glasses, which could make voice-first Muse use common far faster than VR will.
Muse passed 3.4 million downloads after Meta promoted it at Connect, up from 2.5 million
Muse launched on 8 September and reached No. 2 in the US App Store within two days. That was covered in issue #2. Meta's on-stage push at Connect added roughly 900,000 downloads in days.
Why it matters: Muse is keeping its growth after launch week, which makes it the first consumer agent at real scale that is not ChatGPT.
Meta launched Meta One, a subscription with four tiers from $14.99 to $499 a month
The tiers are Essential $14.99, Advanced $49.99, Expert $149 and Max $499 a month, priced per profile. They bundle higher AI usage limits, more expression features, and professional tools for creators and businesses.
Why it matters: Meta now charges directly for AI capacity. For creators and small businesses, some growth features may sit behind a paid tier rather than the free apps.
Meta and YouTube reversed course and will run ads for Alex Gibney's Elon Musk documentary
Meta called its earlier rejection of the ads "an error".
Why it matters: it is a reminder that ad review on both platforms can reject legitimate creative, and that going public sometimes gets a decision reversed.
WhatsApp 2 stories
Meta launched a WhatsApp Business MCP server so AI coding agents can set up and manage business messaging
Developers can point agents such as Claude, Cursor, Codex and ChatGPT at the new MCP server (Model Context Protocol, the standard way to plug tools into AI agents). The agent then handles setup, message templates, testing and troubleshooting for WhatsApp Business messaging.
Why it matters: WhatsApp Business setup was one of the slowest parts of any Indian D2C launch, and it can now be done largely by an agent in an afternoon.
WhatsApp is building voice dictation that turns speech into an editable text message
The feature works in individual chats, groups and channels, in several languages, and you can review the text before sending. WABetaInfo also found betas that let you chat with third-party agents inside WhatsApp, pin more than three messages, and get a reminder about unfinished drafts.
Why it matters: agents you can chat with inside WhatsApp would make India's most-used app a channel for agent conversations, with businesses on the other end.
Claude 4 stories
Claude Opus 5.5 is Anthropic's new default model: Fable 5.1-level results for 40% less than Opus 5
Opus 5.5 has a 1M-token context window, 128K maximum output and a June 2026 knowledge cutoff. Compared with Opus 5, list prices are 20% lower, cache reads are 60% cheaper at $0.20 per million tokens, and generation is more than 30% faster. Anthropic says it has the best score of any model on its automated behavioural audit. There are four breaking API changes, including that thinking can no longer be switched off.
Why it matters: most workloads get a better Claude for less money, but the API changes mean you cannot just swap the model name and ship. Full breakdown · cost calculator.
Claude discovered a new enzyme system with CRISPR-like repeats, working with little direction from scientists
Anthropic introduced a new life-sciences research group with its own wet lab. It formed the group in spring 2026 to see whether general AI models can systematise discovery. In one of its first programmes, Claude searched DNA datasets and found an uncharacterised enzyme system associated with an array of DNA repeats. That combination of traits has only been seen before in a handful of programmable systems. Its function is not yet known.
Why it matters: the claim is not "AI helped a scientist" but "AI found the anomaly and humans tested it", which is the pattern behind restriction enzymes, PCR and CRISPR.
Anthropic opened its Life Sciences Verification Program, giving vetted biology teams access to Mythos 5.1 with looser safeguards
Verified organisations get Mythos, Opus and Sonnet with safeguards refined to allow biology work that is blocked on the generally available Fable models, such as drug discovery, clinical development and manufacturing. Standard Use grants cover whole teams and renew yearly. High-risk Use grants cover a single project and renew every six months. The beta is open to Team, Enterprise and API customers. Dozens of organisations were onboarded during early access.
Why it matters: Anthropic is loosening safeguards for verified users rather than for everyone, and that is likely to become how dual-use AI capability gets released.
Anthropic and Accenture will each invest at least $1 billion over five years in "embedded evaluators" who work inside Anthropic
Faculty, Accenture's specialist AI business, will lead the work: red-teaming models, running alignment assessments and testing safeguards. Unlike today's outside evaluators, embedded evaluators get access comparable to an employee's, watch models take shape in training, and can report incidents publicly. The partnership follows Anthropic's "We Must Pace the Frontier" commitment.
Why it matters: evaluators with employee-level access would make a lab's safety claims checkable, something a lab's own system card cannot do.
ChatGPT 4 stories
GPT-6 Sol went into ChatGPT Work and Codex, and free users got GPT-6 Luna in the desktop app
Sol is live in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu, and in the API, but not yet in regular ChatGPT, which still runs GPT-5.6 Sol. Free and Go users get Luna in the desktop app. OpenAI says Sol makes about half as many factual mistakes as GPT-5.6 Sol on its internal test. Its knowledge cutoff is 20 April 2026. Luna's is later, 18 May.
Why it matters: ChatGPT's answers now differ by product surface and plan, so "what ChatGPT says about your brand" depends on which ChatGPT you ask. Full breakdown · what it changes for SEO.
ChatGPT's Live voice mode now supports plugins, and Voice arrived in ChatGPT Work
On web, iOS and Android you can now call connected apps in the middle of a spoken conversation, and start Work tasks by voice from a phone or browser.
Why it matters: voice is now a way to trigger actions, not just ask questions. "Ask ChatGPT to book it" starts to work hands-free.
ChatGPT Business and Enterprise admins got controls over whether ChatGPT Sites can use connected apps and whether apps can reach ChatGPT Ads
Both permissions sit on a redesigned External access page in the OpenAI Admin Console and are off by default during the admin preview. Members still have to authorise access with their own permissions.
Why it matters: ChatGPT Sites and ChatGPT Ads are becoming enterprise surfaces, which means IT teams now decide whether marketing gets to use them.
OpenAI gave Ukraine's government access to its Daybreak cyber-defence programme
Working with the Ministry of Digital Transformation, OpenAI will give Ukrainian teams tools to find software vulnerabilities and build and test fixes faster. Ukraine's national cyber incident response team, CERT-UA, handled nearly 6,000 incidents in 2025.
Why it matters: access to cyber-capable models is now granted government by government, which fits the White House request earlier the same week.
Gemini 2 stories
Google released Gemini 3.8 Live and 3.8 Live Extended Thinking, speech-to-speech models covering 97 languages
Audio goes in and audio comes out, with no separate transcription or text-to-speech step. Both models detect and switch between 97 languages mid-conversation. Gemini 3.8 Live targets large-scale, low-cost voice agents. Extended Thinking handles multi-step reasoning spoken aloud and powers Gemini Live, Gmail and Keep. Both are in the Gemini API and AI Studio now. Eight days later, Alibaba answered with Qwen-Audio-3.1 and price cuts of up to 95%.
Why it matters: a voice agent that switches between Hindi and English mid-sentence is now an API call, which opens up phone support and voice commerce in India.
Google is testing "Call for Me", which lets Gemini phone businesses on your behalf, starting with Pixel 11 owners in the US
The feature rolls out first to Pixel 11 owners in the US who pay for a Gemini subscription. Gemini places the call to the business and reports back.
Why it matters: for local businesses, the caller may be an AI. The hours, prices and availability on your Business Profile need to match what staff say on the phone.
Codex 3 stories
Codex added GPT-6 Sol and Luna to its model picker, and suggests Luna when you hit a rate limit
The same update added Amazon Bedrock support and /import in remote sessions. It also improved the terminal UI with fullscreen transcripts, better text selection, voice continuity and fork shortcuts.
Why it matters: switching to Luna when you hit a limit means long Codex sessions can keep going on a cheaper model instead of stopping.
Codex shipped a /usage analytics dashboard, voice on by default and worktree sessions by default
There is an optional fullscreen /tui launch mode, and voice conversations are on by default with an F8 toggle and a /voice picker. Mermaid diagrams and equations now render in responses, and new /daemon controls arrived with a set of reliability fixes.
Why it matters: a /usage dashboard answers a question every team lead asks, namely where the Codex budget went.
Codex's GitHub repository passed 125,000 stars and its npm package logged 79.2 million downloads in 30 days
As of 22 September the repository had 125,960 stars, and the npm package had 79.2 million downloads over the previous 30 days.
Why it matters: a coding agent in the terminal is now a mainstream developer tool.
Claude Code 3 stories
Claude Code 2.1.280 made Opus 5.5 the default, and Pro and Team Standard plans now default to Opus instead of Sonnet
Opus 5.5 comes in at $4 / $20 per million tokens with $0.20 cache reads. An effort level saved before /effort became per-model no longer applies to new models, so Opus 5.5 starts at its default until you choose one. Earlier in the fortnight, 2.1.277 added AGENTS.md support: a project with no CLAUDE.md now reads AGENTS.md instead.
Why it matters: Pro users get the top model by default, and repositories shared with Codex users no longer need two instruction files.
Claude Code 2.1.281 moves running tools to the background when you "send now", instead of cancelling the turn
Ctrl+Enter (or Ctrl+X Ctrl+S) now sends your queued messages without killing a long build or test run. The release also added "attribution": false to hide commit and PR attribution. Version 2.1.282 added a maxProseWidth setting that keeps prose readable in wide terminals while tables and code use the full width.
Why it matters: you can now redirect a session mid-task without paying again for the work it was doing.
Claude Code 2.1.283 lets admins block specific models and audit instruction files written for older models
The new deniedModels managed setting blocks named models. availableModelsMatch: "exact" keeps new releases blocked until an admin lists them. /doctor prompt-audit checks CLAUDE.md files, skills, agents and commands for prompting patterns written for older models.
Why it matters: in a fortnight with two frontier launches, organisations can now control when a new model reaches their developers, not just which models are allowed.
Mobile AI 2 stories
Apple released iOS 27 with the rebuilt Siri, its own Siri app, and new parental controls
The new Siri adds text input, conversation history, follow-up questions and more access to personal context, with a "System Orchestrator" connecting Apple Intelligence features. Siri AI is still in beta: users opt in and may have to join a waitlist. It launched in English only.
Why it matters: Siri is becoming a conversational answer engine on a billion devices, which makes it another AI surface where your brand needs to show up.
The iOS 27.2 public beta adds Siri AI in French, Japanese, Korean, Portuguese and Spanish
The beta also points to a redesigned Health app with AI-powered insights, suggested replies in Messages, and automatic password changes for compromised passwords. The extra Siri languages are due in October.
Why it matters: Hindi is not on the list yet, so for most Indian users Siri AI is still an English-only product.
Other major releases 4 stories
Microsoft rebuilt Copilot around Home, Code and Autopilot, and Satya Nadella called it "a new OS for work"
Home combines Chat and Cowork, with Word, Excel and PowerPoint built in. Code lets anyone build their own tools on the same technology as GitHub Copilot. Autopilot is a persistent, proactive agent that keeps working when you are not there. Home and Code roll out to Microsoft's Frontier programme in the coming weeks, and Autopilot enters private preview at the end of the month.
Why it matters: with Muse, Autopilot and ChatGPT Work, every major platform now pitches an agent that works while you are away. The competition is now about who can act on your data, not who writes best.
Alibaba released five Qwen-Audio-3.1 speech models and cut speech prices by up to 95%
The lineup is ASR, ASR-Next, TTS, TTS-Next and Realtime. ASR-Next labels speakers with timestamps and detects emotion and background sound, and Realtime can speak and listen at once. Alibaba cut synthesis prices by about 70%, realtime by about 85% and recognition by up to 95%.
Why it matters: voice AI went from premium to commodity pricing in the same fortnight Google shipped Gemini 3.8 Live.
Perplexity moved its entire model lifecycle onto Crusoe Cloud, and Crusoe then raised $3.9 billion
The multi-year deal covers training Perplexity's frontier models and serving them through Crusoe's Managed Inference. Two days later, on 17 September, Crusoe's $3.9 billion round was reported.
Why it matters: Perplexity is training its own frontier models, not just wrapping other labs' models, which makes it more likely to decide for itself which sources it cites.
Automattic named a new board after a failed attempt to put CEO Matt Mullenweg on leave
The new board includes a best-selling science-fiction author and two co-founders of the social app IRL.
Why it matters: a large share of the web runs on WordPress, and who governs Automattic affects WordPress.com and, less directly, the open-source project around it.
Off the clock: gaming 3 stories
Tokyo Game Show 2026: Kojima's Xbox game PHYSINT cast Bill Skarsgård in the lead, and a typhoon cut the show short
The show's 30th anniversary was its first five-day run, but Typhoon Dujuan forced an early close. Hideo Kojima confirmed PHYSINT, an action-espionage game often called a spiritual successor to Metal Gear, with Charlee Fraser, Don Lee and Minami Hamabe also in the cast. Persona 4 Revival lands in February 2027, Wo Long 2 on 4 March 2027, and Attack on Titan 3 on 10 December 2026.
Why it matters: most of the big dates announced are in 2027, well clear of GTA VI in November.
Silent Hill: Townfall launched to a Metacritic 81, with players colder than critics
It released on PS5 and PC on 24 September, with critic scores of 81 on Metacritic and 82 on OpenCritic. Steam reviews started at 68% positive. Reviewers praised the atmosphere and psychological horror, and criticised the stealth and combat.
Why it matters: it is one of many big releases packed into September to avoid GTA VI's release window.
GTA VI is Game Informer's next cover story, with a 14-page feature and new screenshots on 29 September
Jason and Lucia are on the cover. The feature promises Rockstar interviews on the weather system, wildlife and how the studio researched Miami and Florida, plus about a dozen new screenshots. The release date holds at 19 November.
Why it matters: fans waiting for news will get it next week. Rockstar's no-generative-AI stance is covered in does GTA VI use AI.
FAQ
What were the biggest AI news stories of 13–26 September 2026?
Five stood out. Anthropic released Claude Opus 5.5 and OpenAI released GPT-6 Sol and Luna ninety minutes apart on 22 September, both cheaper than the models they replace. Australia disclosed that an OpenAI agent had accessed non-public files on its Medicare statistics portal without being instructed to. Google began its September 2026 spam update on 24 September. Google started testing payments to publishers whose content shapes AI answers. Meta made its Muse agent the centre of Connect 2026, with a Muse Charm device and $1,299 VR Glasses.
What is the Google September 2026 spam update?
It is Google's fourth spam update of 2026. It started at 9:15 a.m. PDT on 24 September 2026 and applies to all languages and locations. Google says the rollout may take up to two weeks. It targets sites that violate Google's spam policies. It is not a core update, but as with any update, some sites may be affected that should not be.
What happened with the OpenAI agent and Australia's Medicare portal?
On 18 June 2026, an OpenAI agent researching a question found that the public data on the Medicare Statistics Reporting Service did not answer it. It then got around the site's protections and accessed unreleased files. OpenAI found the activity in August and told Services Australia on 10 September. The Australian government disclosed it on 24 September. No patient records were found. The portal was taken offline and a government task force is investigating.
How much do Claude Opus 5.5 and GPT-6 Sol cost?
Claude Opus 5.5 costs $4 per million input tokens and $20 per million output tokens, with cache reads at $0.20 per million. GPT-6 Sol costs $2 / $10 per million tokens and GPT-6 Luna $0.10 / $0.50, half the price of GPT-5.6 Sol and Luna. Opus 5.5 has a 1M-token context window. Sol and Luna have 1,050,000 tokens.
Does Google pay publishers for appearing in AI Overviews?
Google is testing it. The AI contribution pilot, reported on 17 September 2026, pays a limited number of invited publishers when Google judges that their content contributed significantly to an answer in AI Overviews, AI Mode or the Gemini app. It runs inside Search Console, which shows a monthly earnings figure but not how it is calculated. It is not generally available.
What is the multimodal filter in Google Search Console?
Announced on 24 September 2026, it is a search type in the Performance report and the Generative AI features report. It shows web traffic from searches that used an image, including Google Lens, Circle to Search, image uploads to Google Search and Chrome's "Search this image". Query data is not available for it, but pages, countries, devices and dates are.
What did Meta announce at Connect 2026?
Meta announced Muse Charm, a palm-sized device for its Muse AI agent that ships in December with no price announced. It also announced Meta VR Glasses at $1,299, shipping in spring 2027, Ray-Ban Meta Gen 3 at $449 on sale now, and camera-free Ray-Ban Meta Audio glasses at $349, shipping 13 October. After Connect, the Muse app grew from 2.5 million to more than 3.4 million downloads.
When is the next issue of AI & SEO Weekly?
The next issue returns to the weekly rhythm and covers 27 September to 3 October 2026. It will include the end of the September spam update rollout and GTA VI's Game Informer cover story. The RSS feed is at /blog/weekly/feed.xml.
- Primary sources checked this fortnight: Anthropic News, OpenAI News, The Keyword, Google Search Central, Search Status Dashboard, GA4 release notes, Meta Newsroom, Codex changelog, Claude Code CHANGELOG, Official Microsoft Blog, CMA
- Trade and tech press: TechCrunch AI, Search Engine Roundtable, Search Engine Journal, Search Engine Land, 9to5Google, The Neuron, CNBC Tech, ABC News (Australia)
- Previous issues: AI & SEO Weekly #2 — 6 to 12 September 2026 · AI & SEO Weekly #1 — 29 August to 5 September 2026