OpenAI launched its GPT-5.6 model family on July 9, and within the first week the company says its top-tier Sol Ultra configuration produced a proof for the Cycle Double Cover Conjecture – a graph theory problem that has resisted resolution for decades. But while mathematicians begin the long process of verifying the 50-page proof, Windows users have a more immediate reason to pay attention: Microsoft is already slotting GPT-5.6 into Word, Excel, PowerPoint, Copilot Chat, and Cowork.
A Bold Math Claim, Still Unverified
The Cycle Double Cover Conjecture asks whether every bridgeless graph can have its edges covered by a set of cycles such that each edge belongs to exactly two cycles. It sounds esoteric, but the problem has been a sturdy challenge in combinatorics since the 1970s. According to OpenAI’s account, Sol Ultra cracked it in under an hour – an achievement the company described as the model coordinating multiple parallel agents to explore different proof paths simultaneously.
The claim generated immediate headlines. Pluang first reported the story, and Crypto Briefing’s follow-up confirmed that OpenAI has made the full proof and the corresponding prompt publicly available for scrutiny. The sequence is straightforward: a researcher fed the conjecture into the model, and Sol Ultra returned what it asserts is a general proof.
Yet the correct status remains “purported proof,” not “proven.” Graph theorists must now examine every lemma, trace every edge case, and verify that no hidden assumptions crept into the argument. AI-generated proofs can look persuasive while containing subtle definitional mismatches or omitted steps that only a specialist would catch. Until the mathematical community signs off – a process that can take months – the breakthrough is provisional.
Separately, Crypto Briefing noted that reports linking Sol Ultra to a verified solution of Erdős problem #793 are unconfirmed. No independent evidence currently supports that connection.
The Model Family: Sol, Terra, Luna
GPT-5.6 is not a single model but a family of three tiers, each available via OpenAI’s API and gradually rolling into Microsoft 365 Copilot.
- Sol – the premium tier, priced at $5 per million input tokens and $30 per million output tokens.
- Terra – the mid-range option at $2.50 input / $15 output.
- Luna – the value tier, costing $1 input / $6 output.
Sol Ultra is not a separate SKU. It’s a capability setting within the Sol tier that activates parallel agent coordination – essentially, the model spins up multiple reasoning threads, then synthesizes the results. That extra compute shows up in your bill: Ultra can easily consume ten times the tokens of a standard Sol call because each agent processes its own full context window. OpenAI’s benchmark table puts Ultra at 91.9% on Terminal-Bench 2.1, an evaluation of complex command-line tasks that stresses multi-step planning.
For developers and power users, the pricing structure means that tapping into Sol Ultra for heavy-duty analysis – say, bulk code refactoring or large dataset reasoning – requires careful token budgeting. A single session that burns through millions of tokens could cost hundreds of dollars in output fees alone.
The Microsoft 365 Copilot Connection
For everyday Windows users, the model’s arrival inside Microsoft 365 is the story that changes daily workflow. OpenAI confirmed that GPT-5.6 is becoming the preferred model across the Copilot ecosystem: Word, Excel, PowerPoint, the standalone Copilot Chat, and the collaborative Cowork feature. Rollout is underway, though Microsoft has not published a detailed schedule or feature list.
What does that mean in practice? If you use Copilot to draft reports, analyse spreadsheet models, or generate presentation slides, the underlying intelligence just got an upgrade. Expect more coherent document drafts, more accurate data reasoning, and better adherence to complex formatting or logic instructions. The Sol-grade model should, in theory, handle the tangled chain-of-thought tasks that tripped up earlier GPT-4-based Copilot sessions – think multi-step Excel formulas that reference several sheets, or PowerPoint decks that must maintain consistency across dozens of slides.
But there is a catch: without transparent versioning in the Copilot UI, most users will not know which specific GPT-5.6 tier is handling their request. Microsoft could serve Luna for quick email replies and reserve Sol for heavy spreadsheet work, or it could blend them invisibly. The company has not clarified, which makes it difficult to predict the exact quality floor you’ll experience.
What This Means for Your Daily Workflow
Home users and students – If you have a Microsoft 365 subscription that includes Copilot, you’ll likely notice the improvement gradually. Better grammar suggestions, more polished slide decks, and fewer formula errors are the likely gains. The math proof itself has zero near-term impact on your homework, but it signals that the model can reason about structured problems at a high level. That translates to smarter Copilot responses when you ask it to explain a concept or debug a simple script.
Power users and developers – API access means you can build directly with GPT-5.6. The cost jump from Luna to Sol Ultra is steep, so evaluate whether parallel-agent reasoning is worth it. For most applications – summarising documents, classifying text, answering questions – the mid-tier Terra or even Luna will suffice. Reserve Ultra for scenarios where a wrong answer is expensive and the problem genuinely benefits from cross-checking multiple reasoning paths: financial modelling, legal contract analysis, or complex code generation with formal verification.
IT admins – The Copilot integration is the immediate action item. Check your admin centre for any configuration changes tied to the new model family. If your organisation has strict data residency or compliance requirements, verify that GPT-5.6 traffic follows the same processing and storage policies as previous models. Additionally, because the model can generate plausible-looking but incorrect output with high confidence – as the unverified math proof demonstrates – revisit your acceptable use policy around AI-generated content. For teams that produce technical reports, legal documents, or financial statements, mandate human review of any Copilot-generated text that will be used externally.
How We Got Here
The path from GPT-4 to GPT-5.6 has been swift. OpenAI released GPT-4 in March 2023, followed by GPT-4o in May 2024. The GPT-5 family arrived in early 2025 with incremental reasoning improvements, but the 5.6 generation represents the first tiered architecture – three models sharing a common base but optimised for different cost-capability tradeoffs. Microsoft has been the primary distribution channel for consumer and enterprise AI features, baking each new generation into Copilot.
The Cycle Double Cover Conjecture itself has a long history. First posed independently by Seymour and by Szekeres in the early 1970s, it’s a deceptively simple statement that generalises the fact that every bridgeless planar graph has a cycle double cover. Attempts at a general proof have produced partial results and restrictive conditions, but a full solution has remained elusive. If Sol Ultra’s proof holds up, it won’t be the first time an AI has contributed to pure mathematics – DeepMind’s AlphaTensor discovered new matrix multiplication algorithms, and various systems have assisted with formal verification – but it would be the first time an AI produced a complete, standalone proof of a major open conjecture.
One side note: a Solana meme token named $5.6SolUltr launched alongside the model release, attempting to ride the news cycle. Crypto Briefing reports near-zero trading volume and no liquidity. It has no connection to OpenAI or Microsoft, and no bearing on the model’s rollout.
What to Do Now
If you’re an IT administrator or a Copilot power user, a few concrete steps can help you navigate the transition:
- Check your Copilot version – In Microsoft 365 apps, look for the Copilot side pane’s settings or help menu. Some early indicators (like a changed model name in the response footer) may signal that GPT-5.6 is active in your tenant.
- Audit output quality – For one week, spot-check Copilot-generated work more closely than usual. Compare the new output against the previous model’s results on a handful of consistent tasks: drafting a standard email, creating a pivot table suggestion, or generating presentation speaker notes.
- Revisit data handling – Verify in the Microsoft 365 admin centre that your Copilot data processing location and retention settings haven’t changed with the model update. If you use the API directly, review your terms of service for any updated clauses around model tiers.
- Set user expectations – Remind employees that AI-generated content, especially in analytical fields, should be treated as a first draft, not a final product. The unverified math proof is a handy example of why.
- Watch your API bill – If you’re using the OpenAI API directly and want to experiment with Sol Ultra, set low rate limits initially. Monitor token consumption per call closely; a single Ultra run can be an order of magnitude more expensive than a standard Sol call.
Outlook
Mathematicians expect the peer-review process for the Cycle Double Cover proof to take several months. If the proof holds, it will be a significant validation of agentic AI architectures for scientific discovery – and it may accelerate OpenAI’s push to market Sol Ultra as a premium reasoning engine for enterprise R&D. If it collapses, the credibility damage will be limited to a cautionary tale about over-trusting AI output, but it won’t affect the model’s utility for routine productivity tasks.
Meanwhile, expect Microsoft to expand GPT-5.6 availability across more services – Teams transcript summarisation, Windows Copilot, and possibly GitHub Copilot integration are logical next steps. Pricing tiers inside Microsoft 365 could eventually mirror the Luna/Terra/Sol split, giving business customers more granular control over cost and capability. For now, the model has arrived. The proof can wait.