Editor's note
Today put two threads side by side: AI getting cheaper and smarter, and search becoming ever more AI-mediated. Sakana AI shipped an orchestration model that routes across open models to chase top scores, while DeepSeek released a low-cost MIT-licensed model — both widening options that do not lean on expensive frontier systems. On the search side, Google's People Also Ask is now almost entirely AI Overviews, and Search Console again showed missing data, so visibility and data reliability both deserve a second look. For SMBs, the practical work is less about the flashiest model and more about deciding what to use where, and how to verify rankings.
AISakana AI
Sakana AI releases Fugu Max and Fugu Ultra v2, orchestrating open models
On September 11, Sakana AI released Fugu Max and Fugu Ultra v2 — not single large models but a learned orchestration system that routes tasks across a pool of open and specialist models, exposed behind one OpenAI-compatible API. Sakana says Fugu Ultra v2 beat some frontier models on a visual-reasoning benchmark without including any closed model in its pool.
Context
Rather than owning one giant model, the approach picks and combines existing open and specialist models to balance cost and quality — a distinct design philosophy. Claiming top scores without a closed frontier model in the pool challenges assumptions about how you source model capability.
HaLVision's take
It is suggestive for studios that would rather not depend on a single model. It fits our per-project model-selection stance, so we would measure accuracy and cost on real tasks before adopting it. Benchmark numbers do not always match a given use case, so we keep expectations grounded.
AIDeepSeek
DeepSeek releases the MIT-licensed lightweight model DeepSeek-V4.1-Flash
On September 10, DeepSeek released DeepSeek-V4.1-Flash, a lightweight model. It uses a Mixture-of-Experts design — a large total parameter count but only a fraction active per token — with a one-million-token context and image-plus-text input. The weights are published on Hugging Face under the MIT license, and API pricing is kept low.
Context
Activating only part of the parameters per token tends to cut inference cost while holding quality. Open-weight Chinese labs keep shipping inexpensive models, sustaining a contest on cost against major overseas labs.
HaLVision's take
Cheap open weights you can also self-host widen the options for cost-sensitive automation and internal tools. We advise clients to mix closed and open models by sensitivity and cost. Checking the license and operational setup, then piloting small, is the practical path.
AIAnt International
Ant International, Visa and Mastercard explore interoperable agent verification
On September 9, Ant International, Visa and Mastercard announced a collaboration on Know-Your-Agent (KYA) interoperability to streamline identifying and onboarding AI agents across networks. It aims to bridge each company's existing protocols toward shared trust signals and continuous transaction monitoring.
Context
As agentic commerce — AI agents shopping and paying on a person's behalf — comes into view, verifying who operates an agent becomes necessary. The effort tries to make competing protocols interoperable, building a common base while each network keeps its own verification steps.
HaLVision's take
It is early-stage, but it could lay groundwork for e-commerce sites to accept agent-driven purchases later. We think it is safer to wait for standardization to settle before implementing anything around payments and identity. For now we watch the trend and leave design room so we can adapt before it is required.
WebSearch Engine Roundtable
Reports say Google Search Console's page indexing report is missing June data
On September 11, multiple reports said Google Search Console's page indexing report was missing a large chunk of June 2026 data. A reporting bug is suspected, and some say the report has behaved erratically for several months.
Context
Search Console occasionally has display delays or missing data, often separate from actual indexing or ranking status. But when the report that underpins your monitoring is shaky, it becomes easier to misjudge whether a change worked or an anomaly is real.
HaLVision's take
When numbers go missing, the first step is to separate a reporting bug from a real problem. We advise clients not to take a single tool's display at face value, but to cross-check multi-day trends, the actual search results, and server logs. Similar display bugs have happened before, so the baseline is not to panic.
WebSearch Engine Roundtable
Google's People Also Ask answers are now almost all AI Overviews
In early September, analysis reported that about 97% of the answers shown in Google's People Also Ask were now AI Overviews, up from roughly 86% in August. The older format — a passage quoted from a ranking page — has become the minority.
Context
People Also Ask used to mostly surface a snippet lifted from a top-ranking page. Replacing that with a generated answer returns a summary drawn across multiple pages, but it may make clicks through to individual sites less likely.
HaLVision's take
The premise for earning clicks in search is shifting. We advise building clear, structured content that is easy to cite or reference in answers, alongside strengthening branded search and direct touchpoints. Beyond rankings, how to measure exposure inside AI answers is a growing question.
WebSearch Engine Roundtable
Google Merchant Center adds product video uploads
On September 11, a new section was spotted in Google Merchant Center that lets merchants upload product videos. Attaching video to a product listing appears aimed at helping shoppers understand items before buying and improving the display experience.
Context
Merchant Center is the core tool for feeding product data into Google's Search and Shopping surfaces, where richer images and attributes affect exposure and appeal. Video support adds a way to convey feel and scale that a still image alone struggles to show.
HaLVision's take
For SMBs with e-commerce, it is a practical addition that puts existing product videos to use. We suggest starting with short real-object or in-use clips for your top products, alongside optimizing images and attributes. Impact varies by product and sector, so expanding coverage while watching response is the realistic path.