Editor's note
Two threads run through today's issue: a price war among AI models, and rule changes around how people reach search. On September 22, OpenAI launched GPT-6 Sol and Luna and Anthropic released Claude Opus 5.5, both with lower prices than their predecessors, while Xiaomi published a large open-weight model the same day — cost per unit of capability keeps falling. On the Web side, Google now spells out a four-day window for businesses to review suggested edits to their Business Profile, and the UK CMA proposed a search choice screen that would include AI assistants. The practical takeaway is less about picking a model and more about checking where customers find you and who controls your business information.
AIOpenAI
OpenAI launches GPT-6 Sol and Luna, cutting API prices roughly in half
On September 22, OpenAI released GPT-6 Sol and GPT-6 Luna across the API, ChatGPT and Codex. API pricing is $2 input / $10 output per million tokens for Sol and $0.10 / $0.50 for Luna — about half of the GPT-5.6 generation. Sol is aimed at complex work such as coding, while Luna targets high-volume, clearly defined tasks like summarizing and extraction.
Context
The two models bring much of the capability of GPT-6 Astra, the flagship released earlier this month, into faster and cheaper tiers. OpenAI says Sol makes about half as many mistakes as its predecessor. Free and Go users can also reach Luna through the desktop app, so the change touches everyday users, not just developers.
HaLVision's take
For small businesses, routine jobs like triaging inquiries or summarizing meeting notes rarely need a flagship model, and a Luna-class tier is often enough. Halving API costs can make previously marginal automations worth revisiting. The performance figures are the vendor's own, though, so we would run a small pilot on real data before switching anything in production.
WebGoogle Business Profile Help
Google Business Profile now states a four-day window to review suggested edits
Google updated its Business Profile help page to state that owners have four days after notification to accept or reject edits suggested by users (a change reported on September 23). If no action is taken, Google may publish the edit when it is supported by other public information, such as the business website. Some edits that meet policy may still be applied without prior notice.
Context
Anyone can suggest changes to a business listing on Google Maps — hours, phone numbers, categories and more — and unwanted auto-applied edits have long been a pain point for owners. What is new is a concrete deadline. The help page also names the business website as supporting evidence, which confirms that your own site is treated as a reference for your listing.
HaLVision's take
If an owner misses the notification, wrong hours or a wrong phone number could go live within four days. We would make sure notifications go to an inbox someone actually checks, and keep the store details on the website identical to the profile. Local SEO is not only about posts and review replies; this kind of defensive upkeep matters just as much.
AIAnthropic
Anthropic releases Claude Opus 5.5 with API prices cut by about 20%
Anthropic released Claude Opus 5.5 on September 22. API pricing is $4 input / $20 output per million tokens, down about 20% from Opus 5 ($5 / $25), and cache reads are 60% cheaper. It is available on the Claude Platform as well as AWS, Google Cloud and Azure, with Sonnet 5.5 and Haiku 5.5 due in the coming weeks.
Context
Anthropic says Opus 5.5 matches its larger Fable 5.1 model on most tasks and generates output more than 30% faster. External evaluators including METR tested the model before release. OpenAI launched GPT-6 Sol and Luna the same day, showing major vendors moving in the same direction: pushing top-tier capability into lower price bands.
HaLVision's take
Read alongside OpenAI's launch, near-flagship capability is now available at mid-tier prices from more than one vendor. Setting up workflows so you can compare several models per task, rather than locking into one provider, makes it easier to benefit from these cuts. As always, the performance claims are self-reported, so validate on your own business data first.
WebGOV.UK (Competition and Markets Authority)
UK CMA proposes search choice screens on Android and Chrome, with AI assistants as options
On September 23, the UK Competition and Markets Authority published strengthened proposals requiring Google to show Android and Chrome users a choice of search service at first use and then annually. AI assistants such as ChatGPT and Perplexity could be included if they meet technical and security standards. Providers on the screen would also have to fairly attribute publisher content; feedback is open until October 9.
Context
The CMA has been rolling out measures on Google Search through 2026, including publisher opt-outs from AI features and fair ranking requirements. This step adds choice at the point of entry, and notably treats AI assistants on the same footing as traditional search engines. A final decision is expected by the end of the year.
HaLVision's take
It is a UK rule, but it signals that the assumption ‘search means Google’ is starting to erode at the regulatory level too. For small businesses anywhere, being found via AI assistants — not just Google — makes it more important to publish accurate, first-hand information that is easy to cite. It is one more data point for taking LLMO seriously.
AIXiaomi MiMo
Xiaomi releases MiMo-V2.6 series as open weights, led by a trillion-parameter MoE model
On September 22, Xiaomi released MiMo-V2.6-Pro, MiMo-V2.6-Flash and MiMo-V2.6-Distill-Qwen-9B, publishing weights and a technical report on Hugging Face. Pro is a mixture-of-experts model with about 1.02 trillion total parameters and roughly 42 billion active per token, supporting a 1-million-token context. It scored 46 on the third-party Artificial Analysis Intelligence Index, ranking first among open-weight models at launch.
Context
Pro and Flash are omnimodal, handling text, images, video and audio within one model family. Beyond the weights, Xiaomi also released more than 7,000 reinforcement-learning task environments and a training framework, making the approach easier for researchers to reproduce. It is part of a broader trend of Chinese open models narrowing the gap with commercial ones.
HaLVision's take
Open-weight models widen the options for companies that want to keep data in-house or run AI on their own infrastructure. A trillion-parameter model is not something a small business will host itself, though, so realistic entry points are hosted providers or small distilled variants like the 9B model. Check the license terms and where your data is processed before adopting.
WebGoogle Search Central
Google adds CrUX ad metrics to its page experience documentation
On September 22, Google added Chrome User Experience Report (CrUX) ad metrics to the resources section of its Search Central page experience documentation. The four metrics are average ad count in the viewport, ad density, ad weight in network bytes, and ad weight in CPU milliseconds. Google's John Mueller said they are not direct ranking factors but tools for discussing and improving user experience.
Context
CrUX is field data collected from real Chrome users and underpins Core Web Vitals. Adding measures of how many ads a page shows and how heavy they are makes it possible to quantify pages that feel cluttered with ads. Google denies a direct ranking effect, but the addition still signals what it considers a good experience.
HaLVision's take
If you run an ad-supported media site or blog, this is a good prompt to revisit ad placement and volume. Most business sites do not run ads, but the same logic applies to heavy tags and third-party scripts. We would treat these metrics less as a ranking lever and more as a yardstick for keeping pages light enough that visitors stay.