Its model, SubQ, is being presented as a breakthrough in long-context AI — the ability for a model to read and reason across much larger bodies of text, code, documents, contracts, records, and business information without breaking everything into tiny pieces first.
The claims are big. Some people in the AI world are excited. Others are sceptical. That is healthy.
For Australian small and medium businesses, the important question is not whether SubQ becomes the next dominant AI model. The more useful question is this:
What changes when AI can read much more of your business at once?
The Problem: Business Information Is Scattered Everywhere
Most businesses do not have one clean source of truth. Important information is usually spread across emails, PDFs, spreadsheets, supplier contracts, customer records, invoices, quoting templates, internal policies, website pages, CRM notes, accounting exports, staff knowledge, and old folders no one wants to clean up.
This is why many AI projects feel impressive in a demo but disappointing in real work. The model may be capable, but it is only seeing a small slice of the business. It answers from whatever information was pasted into the prompt, retrieved from a database, or summarised by another system.
That works for simple tasks. It becomes harder when the answer depends on many scattered documents or a long operational history.
A customer dispute might require emails, invoices, delivery notes, terms and conditions, and staff comments. A tender response might require past proposals, insurance certificates, capability statements, pricing tables, case studies, policies, and compliance documents. A business process review might require forms, spreadsheets, CRM exports, website pages, and conversations with staff.
The real issue is not simply that AI needs better prompts. It is that AI often needs more complete context.
Why Current AI Workflows Can Be Clunky
Modern AI systems often deal with large business information by using workarounds. The most common approach is to break documents into chunks, store those chunks in a search system, retrieve the most relevant pieces, and then ask the AI model to answer from those pieces.
This can be useful. It is how many document chatbots and internal knowledge assistants work. But it has limits.
If the retrieval system finds the wrong chunk, the answer may miss the point. If the relevant information is spread across ten documents, the system may only retrieve five. If the source material is messy, duplicated, outdated, or contradictory, the model may sound confident while quietly mixing things together.
This is one reason AI automation needs to be designed carefully. The model is only one part of the system. The surrounding workflow matters just as much.
Long-context AI is interesting because it may reduce some of these workarounds. Instead of asking the model to work from a few selected fragments, it may become more practical to give it the whole contract set, the full policy library, the complete spreadsheet, the entire customer history, or the full codebase.
That does not make the work automatic or risk-free. But it changes what is possible.
The Technical Bottleneck, in Plain English
Most major AI models are based on the Transformer architecture, introduced in the 2017 paper Attention Is All You Need. Transformers use a mechanism called attention. In simple terms, attention helps the model decide which parts of the input relate to other parts.
That is powerful, but expensive.
In standard dense attention, every token is compared with every other token. As the input gets longer, the amount of comparison work grows very quickly. Double the amount of information, and the work does not simply double. It can grow much faster.
That is one of the reasons very long AI inputs are expensive, slow, and difficult to manage. This limitation has shaped the way AI systems are built. It is one reason we see so much chunking, retrieval, summarisation, and orchestration around AI models.
What SubQ Claims to Change
Subquadratic says its SubQ models use a different approach called Subquadratic Sparse Attention. The basic idea is simple enough: do not compare everything with everything. Learn which comparisons matter, and focus attention there.
Instead of processing every possible relationship between every part of the input, SubQ aims to identify the relationships that are most useful and skip much of the wasted computation. The company claims this makes very large context windows much cheaper and faster to process.
In June 2026, Subquadratic released details for SubQ 1.1 Small. The company says the model can achieve near-perfect retrieval on needle-in-a-haystack tests up to 12 million tokens and, at 1 million tokens, uses far less attention compute than dense attention while running faster than FlashAttention-2 in an architecture-level comparison.
Those are serious claims. If they hold up in broader real-world use, long-context AI could become much more practical.
Why People Are Excited
Long-context AI is not just about making prompts bigger. For business use, it could change the shape of practical automation.
A model that can work across much larger bodies of information could help with:
- Reviewing a full set of supplier contracts
- Comparing internal policies against current practice
- Checking large spreadsheets for inconsistencies
- Summarising long customer histories before drafting a reply
- Reading a full website before suggesting improvements
- Reviewing a software codebase without splitting it into fragments
- Preparing grant, tender, or compliance responses from a large source library
- Finding patterns across operational records, reports, and notes
This matters because many business problems are not isolated. The useful answer is often spread across several files, systems, and decisions.
For example, a builder reviewing variation claims may need the contract, quote, email trail, photos, schedule, invoices, and notes from site meetings. A tourism operator may need booking terms, supplier agreements, customer messages, refund policies, and staff procedures. A professional services firm may need past proposals, client notes, pricing models, project reports, and compliance statements. An allied health provider may need policies, intake forms, appointment notes, funding rules, and communication templates.
These are not futuristic use cases. They are normal business problems. The difficulty is that today’s AI systems often struggle when the relevant material is too large, too scattered, or too poorly structured.
Why the Scepticism Is Fair
SubQ should not be treated as a guaranteed revolution. There are still important questions.
The model is not broadly available to the public. Many claims come from the company’s own announcements and selected evaluations. Broader testing across general reasoning, safety, multilingual performance, privacy-sensitive workflows, and messy real-world business data is still needed.
There is also a difference between a benchmark result and a dependable business tool. A model can perform well on long-context retrieval tests and still struggle with ambiguous instructions, outdated records, conflicting documents, or unclear business rules.
What to Watch For
The SubQ story is still being written. Here is what practical-minded business owners should pay attention to:
- Independent verification. Has anyone outside Subquadratic confirmed the numbers? Appen’s June 2026 verification of SubQ 1.1 was a meaningful step. More third-party testing is needed.
- Public access. A model you cannot test yourself is a model you cannot depend on. SubQ 1.1 Small is now partially available. Watch for broader access and clear pricing.
- Real-world performance. Needle-in-a-haystack tests measure one narrow capability. The harder test is whether the model stays coherent and useful across 10,000 pages of messy business documents.
- Integration paths. Even if SubQ delivers, it needs to plug into existing tools — CRM, email, accounting, document management — before it helps most businesses.
None of this is reason to dismiss SubQ. It is reason to watch carefully, test when possible, and avoid building critical workflows around unproven foundations.
The Practical Takeaway for Australian SMBs
You do not need SubQ to be real for this conversation to matter. The direction is clear: AI context windows are getting larger, attention is getting cheaper, and the bottleneck is shifting from “how much can the model read” to “do you have your information ready for it to read.”
Three practical steps you can take now, regardless of what Subquadratic ships:
- Get your policies written down. If your cancellation policy, refund terms, or safety procedures live in someone’s head, they do not exist for AI. Write them down.
- Structure your documents. Markdown files with clear headings beat scanned PDFs every time. Give AI systems clean, well-organised source material.
- Consolidate your sources. If customer history lives in three different systems, long-context AI will not fix that. It will just read three inconsistent sets of data and confidently repeat the contradictions.
The businesses that benefit most from long-context AI will not be the ones with the most data. They will be the ones that organised their data before the models were ready to read it all.
Want to know where AI could help your business first?
Long-context AI is promising, but the best automation opportunities usually start with one clear workflow.
Automic helps Australian businesses identify practical, low-risk automation opportunities across admin, customer follow-up, reporting, quoting, knowledge management, and internal operations.
Start with an Automation Readiness Assessment or speak with Automic about a focused workflow review.