Copilot now reasons across multiple files for free (GPT-5 rollout)

Microsoft has quietly extended Copilot’s document-smarts: the web and Windows 11 Copilot surfaces can now reason across multiple uploaded files at once, bringing a ChatGPT-style multi-file analysis workflow to Microsoft’s assistant for free. The change — reported in a recent Windows Latest hands‑on — lets Copilot ingest several documents and “connect the dots” instead of answering about each file in isolation. That capability is being rolled into the consumer Copilot experience alongside the broader GPT‑5 rollout and Copilot’s evolving conversation modes, but some implementation details (notably the per‑request file-composition limit) are currently reported rather than fully documented in Microsoft’s public release notes. (microsoft.com)

Overview: what changed and why it matters​

Copilot already allowed users to attach many files to a conversation, but until now it generally processed uploads independently — answering about one file at a time rather than synthesizing information across them. The recent update makes Copilot capable of multi-file reasoning, enabling use cases such as:

• Comparing two job postings against a resume and recommending which role matches best.
• Building a travel plan from a budget spreadsheet, an itinerary PDF and a packing list.
• Generating study materials (quiz questions, flashcards, summaries) from several lecture notes or PDFs uploaded together.

Windows Latest’s coverage says Copilot now reads and reasons across up to three files at once in the web and Windows app interfaces; the story shows a demonstration of combining three study documents into a quiz and notes that the same Study‑style workflow is now available in Copilot. That account is supported by hands‑on screenshots and an explanation from a Microsoft representative quoted in the piece.
Microsoft’s broader messaging around Copilot also confirms the product is receiving deeper research, multipage “Pages” workflows, and improved file-aware features as part of a bigger Copilot evolution that includes Smart routing to newer models (GPT‑5) and conversation modes like Think Deeper and Deep Research. Microsoft announced GPT‑5 in Copilot in early August and has documented multiple related features — Deep Research for multi‑source analysis, Copilot Pages for long-form work, and file-aware tools in OneDrive — that make multi-document workflows a first‑class scenario. (blogs.microsoft.com, microsoft.com)

---

Background: Copilot, multi-file workflows and the model context​

Copilot’s evolution toward multi-document intelligence​

Copilot began as a chat and writing assistant integrated with Windows and Microsoft 365. Over the past year Microsoft has shifted it from a sidebar helper to a system-level AI layer: Copilot Vision for images and screen content, Copilot Pages and Deep Research for multi‑document projects, and a model router that dynamically chooses a faster or deeper model depending on the prompt’s needs. These additions were part of a major Copilot update that Microsoft outlined on its official blog and in product release notes. (blogs.microsoft.com, microsoft.com)
The most consequential platform move underpinning this capability is Microsoft’s integration of OpenAI’s GPT‑5 across its product line. Microsoft made GPT‑5 available inside Copilot and related products, and it uses server‑side model routing (Smart mode) to pick the appropriate variant — trading off speed and cost versus deeper reasoning when necessary. That routing is what lets Copilot attempt heavier multi‑file analysis without forcing users to pick a model. (microsoft.com, devblogs.microsoft.com)

How ChatGPT and competing services handled multi-file analysis​

OpenAI’s ChatGPT long had stronger multi‑file workflows via Projects/Advanced Data Analysis, which allow files to be uploaded to a single GPT or conversation and then synthesized. OpenAI’s public guidance documents explain upload limits, file size caps and per‑user quotas for file uploads — a useful technical reference when comparing how different assistants handle uploaded content. In ChatGPT’s case, Projects and the Advanced Data Analysis model were designed to synthesize many files together and run analysis over them. (help.openai.com)
Microsoft’s OneDrive Copilot had already offered file comparison and summarization over multiple files (OneDrive’s Copilot features document supports selecting and comparing up to five files), so multi‑file analysis in the broader Copilot ecosystem is an evolution rather than a brand‑new category for Microsoft. What changed this week is the consumer Copilot (web and Windows app) adopting ChatGPT‑style multi‑file synthesis on the free surface, according to reporting and hands‑on tests. (support.microsoft.com)

---

What the new multi-file capability actually does (and what’s claimed)​

Key observable behaviors (based on reporting and official docs)​

• Multi‑file synthesis: Copilot can now take several uploaded documents and reason across them — identifying links, overlaps, contradictions, and opportunities to merge or transform information into a new output (summaries, quizzes, action lists). This is the primary functional improvement users will notice when uploading related files in the same message or session.
• Conversation modes: When combined with Think Deeper or Deep Research modes, Copilot applies more compute and longer reasoning to multi‑file tasks, returning richer, citation‑style outputs or multi‑part reports. Microsoft documents these modes and positions Deep Research as the subscription-grade, long‑running analysis mode. (support.microsoft.com)
• GPT‑5 grounding: The new multi‑file synthesis surfaces as part of the broader GPT‑5 rollout across Copilot — Smart mode routes complex multi‑file prompts to GPT‑5 reasoning variants as needed. Microsoft’s release notes confirm GPT‑5 is available in Copilot and that a router chooses the right model for the job. (microsoft.com, devblogs.microsoft.com)

The Windows Latest claim: “reads up to three files together”​

Windows Latest reports Microsoft told them that Copilot’s consumer surface can reason across up to three files at once (for example: two job listings + one resume). That specific numeric cap (three files) was confirmed to the reporter in a direct exchange, but is not — as of this writing — a broadly published limit in Microsoft’s public release notes or core support pages. Microsoft’s OneDrive Copilot and other Microsoft 365 surfaces use different caps (OneDrive docs reference up to five files for certain compare/summarize features), which suggests limits can vary by product surface and scenario. Treat the “three‑file” assertion as the company’s operational detail for the web/Windows chat flow reported to the journalist; it’s credible given product rollout patterns but not yet present in a formal, general‑availability support article. This is a flagged, reporter‑confirmed claim that should be validated in your own tenant or client before relying on it for production workflows.

---

Hands‑on impressions and practical use cases​

Study and Learn / study‑style workflows​

Windows Latest demonstrated a “Study and Learn” style usage: toggle a study mode, upload three documents, and ask Copilot to create flashcards or a quiz. The assistant returned quiz items, scored responses and offered explanations — an experience that closely mirrors ChatGPT’s Study Mode and ChatGPT’s Projects + Advanced Data Analysis workflows. Microsoft also promotes Copilot as a study aid elsewhere (study tips, practice tests and time management guidance), so the interface behavior aligns with Microsoft’s stated education goals. If your workflow involves study materials, Copilot’s multi‑file synthesis can accelerate question generation, create summaries, and turn disparate lecture notes into structured revision content.

Business and productivity scenarios​

• Contract and version comparison: Combine multiple contract drafts and an amendment to get a single, annotated summary of changes and risk flags. Microsoft’s OneDrive compare feature already covers up to five files, and the new Copilot multi‑file synth brings similar capabilities to the chat surface. (support.microsoft.com)
• Job matches and recruiting: Upload a candidate’s resume plus several job descriptions and ask Copilot to score fit, extract required skills, and propose interview questions aligned to gaps.
• Project synthesis: Upload design briefs, budgets and timelines; get a consolidated project plan with risk items and suggested next steps.

Limitations observed or likely to appear​

• File count and size: Even if Copilot accepts many files in an upload, the multi‑file reasoning window may be intentionally limited (Windows Latest’s three‑file figure) to bound compute costs and preserve latency. Larger numbers of files can still be attached, but Copilot may analyze them individually or route heavy synthesis tasks to Deep Research (a paid/subscribed mode).
• Format-dependent fidelity: Spreadsheets, images, and scanned PDFs may be handled by different pipelines (table/sandbox kernels vs. OCR), and outputs for structured data will be different from pure text synthesis. OpenAI and Microsoft documentation both note special limits and extraction behaviors for spreadsheets and images. (help.openai.com, datastudios.org)
• Hallucinations and verification: As with any LLM‑driven synthesis, Copilot can produce confident but incorrect connections. Outputs should be treated as assistant drafts that require human verification, especially for legal, financial, or safety‑critical content. Microsoft’s product guidance and independent testing both flag this as a persistent risk. (microsoft.com, theverge.com)

---

What Microsoft has officially documented (and where to look)​

• GPT‑5 integration and Smart mode: Microsoft announced GPT‑5 availability across Copilot surfaces on Aug 7, 2025 and published release notes describing Smart model routing. These documents show Microsoft’s intent to route complex prompts (like multi‑file synthesis) to reasoning‑grade model variants. (microsoft.com, devblogs.microsoft.com)
• Conversation modes (Quick / Think Deeper / Deep Research): Microsoft’s support pages explain when to use each mode and the expected latency/quality tradeoffs — Deep Research is intended for longer, more rigorous analysis and is commercially gated. These modes are the natural place to invoke multi‑file work. (support.microsoft.com)
• OneDrive Copilot file features: OneDrive’s Copilot supports comparing and summarizing multiple files (up to five in the OneDrive compare/summarize flows), demonstrating Microsoft’s multi‑file tooling within its cloud storage surface. That helps explain how multi‑file capabilities are being exposed across different Copilot endpoints. (support.microsoft.com)
• OpenAI file upload guidance: For context on file‑upload strategy and limits, OpenAI’s File Uploads FAQ lays out token and per‑file caps for ChatGPT Projects and Advanced Data Analysis, which is helpful to understand how other assistants (and sometimes shared model infrastructure) gate multi‑file tasks. (help.openai.com)

---

Technical analysis: how multi‑file reasoning likely works under the hood​

• Chunking and retrieval‑augmented synthesis: Long documents are typically split into chunks, embedded and indexed; the model retrieves relevant chunks from the file set to answer prompts without exceeding its context window. That hybrid approach (chunking + retrieval) is standard across modern assistants. Evidence from OpenAI documentation and public engineering writeups indicates this pattern is in use. (datastudios.org, community.openai.com)
• Model router and cost control: Copilot’s Smart mode uses a server‑side router to pick a lightweight model for short queries and escalate to a heavier GPT‑5 reasoning model for complex, multi‑document tasks. That router is essential for balancing latency, cost and quality when doing multi‑file syntheses for free users. Microsoft’s release notes and product blogs explicitly describe this architecture. (microsoft.com, devblogs.microsoft.com)
• Specialized pipelines for spreadsheets and images: When CSVs/Excel files are uploaded, systems often spawn a sandboxed kernel (a “data mode”) to parse and operate on tables directly, producing charts and structured analysis. OCR or vision pipelines convert images or scanned PDFs into text before synthesis. These specialized paths influence fidelity and turnaround time for multi‑file workflows. OpenAI and third‑party analyses document similar patterns. (help.openai.com, datastudios.org)

---

Privacy, security and governance — the checklist every user and admin should run through​

• Data sensitivity assessment
• Treat uploads as cloud‑processed content by default. Even if some features run on‑device (Copilot+ PCs can offload compute), the consumer chat and web surfaces route content through Microsoft servers. Don’t upload secrets or regulated data without confirming governance. (blogs.microsoft.com, support.microsoft.com)
• Tenant and compliance controls
• Enterprise users should use Microsoft 365 Copilot and Azure AI Foundry flows for governed processing. These enterprise flows include tenant‑scoped controls, audit logs and documented data residency options. Microsoft’s enterprise messaging around Copilot highlights these capabilities. (microsoft.com, mc.merill.net)
• Output verification
• Plan for a human‑in‑the‑loop sign‑off for any critical outputs (contracts, regulatory filings, legal briefs). Multi‑file synthesis increases the chance of plausible but incorrect extrapolations. (microsoft.com)
• Rate limits and quotas
• If your workflows are heavy (batch document analysis), confirm the exact per‑user quotas for the consumer Copilot surface in your region. Public reporting suggests Copilot’s free surface offers more liberal access to deeper reasoning than some generic free tiers, but Microsoft’s audited quota tables are not uniformly published, so test in your environment or consult your account representative. (microsoft.com, help.openai.com)

---

Practical advice: how to test and adopt multi-file Copilot in your workflow​

• Start small:
• Test with three representative documents of the type you’ll use (notes, PDFs, spreadsheets). This helps you understand whether the assistant treats them as a combined context or answers each independently. Windows Latest’s test used three files for a study quiz; replicate that experiment for your content type to see behavior.
• Use the right mode:
• For quick synthesis use Quick; for deeper combinatory reasoning toggle Think Deeper or Deep Research (Deep Research may be limited to paid tiers). Expect Deep Research to take longer but to return a more structured, citation‑aware report. (support.microsoft.com)
• Feed the assistant explicit instructions:
• In your prompt, tell Copilot to “treat these three files as a single corpus and extract X, Y, Z,” so the router can decide whether to escalate to a reasoning model.
• Verify outputs systematically:
• Ask the assistant to highlight evidence lines and show which document a claim came from. Merge that into your review checklist before acting on the generated output.
• Automate cautiously:
• If you plan to automate multi‑file synthesis at scale (e.g., nightly reports), pilot the flow and measure hallucination rates; consider a human review or a deterministic check before publishing.

---

Strengths, risks and strategic takeaways​

Strengths​

• Productivity multiplier: Multi‑file synthesis removes tedious manual collation and highlights relationships across documents users would otherwise inspect one by one.
• Democratizing advanced reasoning: Exposing deeper reasoning on the free consumer Copilot surface (backed by GPT‑5 Smart mode) lowers the barrier to advanced AI assistance for individuals and small teams. (microsoft.com)
• Education and learning: Study‑style workflows — generating quizzes, flashcards and summaries from multiple lecture documents — scale study and revision tasks efficiently.

Risks​

• Unclear operational limits: The precise per‑request file composition cap appears to vary by surface and is not uniformly documented; Windows Latest reports “three files” for the consumer chat, while other Microsoft surfaces accept up to five. That inconsistency can confuse adoption decisions. Flagged claim: Windows Latest’s three‑file cap is reporter‑confirmed but not yet fully documented in public Microsoft support pages.
• Hallucination and over‑confidence: Multi‑document synthesis amplifies the chance of plausible but incorrect extrapolations if the model tries to unify inconsistent or incomplete sources. All outputs require validation.
• Privacy and compliance: Users may inadvertently upload regulated content to a cloud service without appropriate protections. Admins must categorize and govern uploads carefully, especially for legal, healthcare, or financial data. (microsoft.com, support.microsoft.com)

---

Final assessment and recommendations​

Microsoft’s addition of ChatGPT‑style multi‑file analysis to Copilot’s consumer surfaces is an important milestone for everyday AI productivity. The feature lets users move from single‑file Q&A to genuine synthesis workflows — a shift that mimics the most useful parts of ChatGPT’s Projects/Advanced Data Analysis and brings them into Windows and the Copilot web app for free. Microsoft’s GPT‑5 rollout and Smart model routing provide the technical scaffolding for this behavior, and Microsoft’s documented tools (Deep Research, Copilot Pages, OneDrive Copilot) confirm a strategic focus on multi‑document workflows. (microsoft.com)
However, adoption requires caution. The exact multi‑file composition limits vary by surface and currently rely on a mix of published docs and hands‑on reporting (Windows Latest’s three‑file claim). Users and administrators should:

• Validate the behavior in their own tenant or client before designing production workflows around a numeric file cap.
• Treat generated outputs as drafts that need human verification, particularly for sensitive content.
• Use enterprise Copilot or Azure AI Foundry for governed, auditable multi‑file processing when working with regulated data.

If you plan to try this feature today: test with representative content, toggle Think Deeper/Deep Research for more rigorous outputs, and insist on traceability in the assistant’s replies (show the source file and line for any critical assertion). This combination of careful testing and governance will let you benefit from Copilot’s new multi‑file reasoning while avoiding the common pitfalls that accompany powerful, cloud‑based generative AI.

---

The incremental gains from multi‑file synthesis are immediate and practical: less manual aggregation, faster insight generation, and new classroom or hiring workflows. The longer-term question is how Microsoft balances free access to reasoning‑grade models with capacity controls and enterprise governance. For now, Copilot’s multi‑file capability is a useful productivity upgrade — a welcome one — provided teams adopt it with verification, privacy controls, and a clear understanding of the product’s surface‑specific limits.

---

Source: windowslatest.com Microsoft Copilot (web), Windows 11 app gets ChatGPT's multi-file analysis feature for free