Unlimited free ChatGPT vs governed enterprise AI on Azure

On 6 August 2026, OpenAI removed the limits on text-based conversations for all ChatGPT users. Free and Go tiers now run on GPT-5.6 Luna, a new model that replaced GPT-5.5 as the default, and every tier gets a "Think" button that lets the model reason longer on hard questions. The unlimited chats and the Think button are rolling out this week. ChatGPT recently passed 1 billion weekly users, and the free product just became meaningfully better: OpenAI's internal evaluations show Luna making roughly 62 percent fewer factual errors than the previous default. If you run IT, security or AI strategy for a Swedish or EU enterprise on Microsoft's stack, this is not consumer news you can scroll past. The strongest argument you had against shadow AI, that the free tool was rate-limited and noticeably worse than what the company could provide, just lost most of its force.

What OpenAI actually shipped

The announcement contains four concrete changes, and it is worth being precise about them before drawing conclusions.

Unlimited text chats for Free and Go users. Rate limits on text conversations are gone, subject to abuse guardrails. Limits on file uploads, image generation, voice and other tools remain in place.
GPT-5.6 Luna as the free default. Luna replaces GPT-5.5 for Free and Go tiers. OpenAI's testing puts factual errors at about 62 percent fewer than GPT-5.5 Instant.
A "Think" button for everyone. Free users can now trigger deeper reasoning on a per-message basis, a capability that used to be gated behind paid tiers.
GPT-5.6 Sol for Plus and Pro. Paid subscribers get an upgraded model tuned for everyday conversation, with about 68 percent fewer factual errors than GPT-5.5 Instant, plus a slider that adjusts how much thinking the model applies per query.

Read that list from the perspective of an employee in your finance, legal or engineering team. The free tool they may already be using without approval now has no daily ceiling, hallucinates far less, and can reason harder on demand. The gap between "the ChatGPT I use at home" and "a serious work tool" narrowed substantially in one release.

Why this changes the shadow AI calculus

Shadow AI, employees using consumer AI tools with work content outside any sanctioned channel, has always been driven by a simple equation: the unsanctioned tool is faster to access and good enough, and the sanctioned alternative is either absent, slow or worse. Enterprises have leaned on three counterweights.

First, friction: free tiers had message caps, so heavy work use pushed people toward paid accounts, which show up on expense reports and can be governed. That counterweight is now gone for text. An employee can run an entire workday of drafting, summarising and analysis through the free tier without ever hitting a wall.

Second, quality: the free default model was visibly weaker, so anyone doing serious work had a reason to ask for a corporate tool. A 62 percent reduction in factual errors, if it holds up in daily use, blunts that argument too. The free tier is no longer the demo version.

Third, policy: many organisations simply banned consumer ChatGPT. Bans have a poor track record when the banned tool is free, excellent and one browser tab away. With a billion weekly users, ChatGPT is now as familiar to your workforce as search engines, and policies that ignore that reality tend to produce hidden usage rather than no usage.

The governance gap is the point. Under OpenAI's consumer terms, content from Free and Plus accounts can be used to improve OpenAI's models unless the individual opts out. There is no admin console, no audit log, no data processing agreement with your company, and no way to apply retention or eDiscovery policies. None of that changed on 6 August. What changed is that the tool carrying those properties became unlimited and considerably better, which means more of your data will flow through it unless you offer something comparable.

Your realistic response options on the Microsoft stack

For an Azure-first organisation, the answer to "employees want ChatGPT-class AI" is not one product but a portfolio. Each option trades cost against capability and governance depth.

Microsoft 365 Copilot Chat: the free governed baseline

Most Swedish enterprises already license Microsoft 365, which includes Copilot Chat at no extra cost. It gives employees a capable web-grounded chat experience with enterprise data protection: prompts and responses are not used to train foundation models, and usage sits inside your tenant's compliance boundary. This is your minimum viable answer to the free ChatGPT tier. If employees have nothing sanctioned to open, they will open ChatGPT; if Copilot Chat is one click away in Edge, Teams and microsoft365.com, a meaningful share of casual usage moves inside the boundary at zero incremental licence cost.

Microsoft 365 Copilot: paid seats where context pays

The paid Copilot seat adds grounding in your own tenant data: mail, calendar, files, meetings. This is the capability consumer ChatGPT cannot replicate, because it cannot see inside your tenant. The economic question is not "is Copilot better than free ChatGPT at general chat" but "for which roles does grounding in company data create enough value to justify the per-seat price". Sales, project management and executive support tend to clear that bar; occasional users often do not. Buy seats by role, not wall to wall.

Azure OpenAI and Azure AI Foundry: build where the value is specific

For workflows where generic chat is the wrong shape, document processing, customer service, internal knowledge assistants, Azure OpenAI gives you the model layer under your own governance: your prompts and outputs are not used to train the underlying models, you choose the Azure region, and the service inherits your existing Azure identity, networking and monitoring controls. Swedish organisations can deploy in Sweden Central and keep inference in country for the models available there. This is also where a token-based cost model beats per-seat licensing: an internal assistant serving 2,000 employees a few queries a day usually costs far less on pay-per-token than 2,000 Copilot seats.

ChatGPT Business or Enterprise: when the workforce insists

Some teams will want ChatGPT specifically, and OpenAI's business tiers exclude customer content from training by default and add admin controls. That is a defensible choice for pockets of power users. For an Azure-first EU organisation it is usually the supplement, not the backbone: it sits outside your Microsoft compliance boundary, adds a separate vendor relationship to manage under GDPR, and duplicates capability you may already be paying Microsoft for.

A decision framework: block, channel or provide

Treat every employee AI use case as landing in one of three lanes, and make the lane assignment explicit in policy.

Block: use cases involving personal data, customer data, source code under IP constraints, or anything regulated. Enforce with more than policy text: DLP rules in Microsoft Purview, browser controls in Edge for Business, and network-level visibility of AI service usage. Blocking only these categories, rather than all AI, keeps the policy credible.
Channel: generic drafting, summarising public information, brainstorming, language polish. Point this traffic at Copilot Chat, which costs you nothing extra and keeps the data under enterprise protection. Success here is measured by consumer-tool traffic going down, not by bans going up.
Provide: high-value, role-specific workflows. Fund paid Copilot seats or Azure OpenAI applications where a business case exists. Every workflow you provide well is a workflow that will not leak to a consumer tool.

The test for whether your response is working is behavioural. If employees still paste work content into free ChatGPT after your rollout, the sanctioned option is losing on speed, quality or access, and that is your problem to fix, not theirs.

The Swedish and EU dimension

GDPR exposure scales with usage. When an employee pastes customer information into a consumer chatbot, your organisation is making an undocumented transfer of personal data to a processor you have no agreement with. That was true last month too, but unlimited free access raises the volume of the behaviour you must assume is happening. Data protection officers should treat this release as a trigger to refresh the Article 30 record and the employee guidance, because "we told people not to" is a weak answer to a supervisory authority asking how you control the flow.

The AI Act's literacy obligation is directly relevant. Article 4 of the EU AI Act, applicable since February 2025, requires organisations deploying AI systems to ensure a sufficient level of AI literacy among staff using them. A workforce quietly using unlimited consumer AI is a workforce using AI systems, whether or not you deployed them. Structured training on what may and may not go into which tool is both an Article 4 measure and your best practical defence against shadow AI, since most misuse is uninformed rather than malicious.

Residency and procurement favour the governed path. Microsoft's EU Data Boundary commitments for M365 and the ability to pin Azure OpenAI deployments to EU regions, including Sweden Central, give you answers to the residency questions that public-sector and regulated buyers in Sweden routinely ask. A free consumer service offers no such commitments and cannot be brought into an upphandling requirement set at all. For organisations subject to NIS2-driven scrutiny, the ability to show identity-integrated, logged, policy-controlled AI usage is the difference between an audit finding and an audit pass.

A 30-60-90 day response plan

Days 1 to 30: see the problem. Turn on visibility of AI service traffic through your secure web gateway or Defender for Cloud Apps. Measure actual consumer-AI usage before writing policy about it. Confirm Copilot Chat is enabled, pinned and discoverable for every licensed user.
Days 31 to 60: publish the lanes. Ship a one-page AI usage policy built on block, channel, provide. Pair it with short role-based training that satisfies Article 4 and actually explains why customer data does not belong in consumer tools. Configure Purview DLP for the block lane.
Days 61 to 90: provide something better. Pilot paid Copilot seats in two or three roles with measurable workflows. Stand up one Azure OpenAI application for a high-volume internal use case in an EU region. Re-measure consumer-AI traffic and compare against the day-30 baseline.

Takeaways

OpenAI made free ChatGPT text chats unlimited on 6 August 2026, with GPT-5.6 Luna cutting factual errors by roughly 62 percent versus the prior default. The free tier is now a credible daily work tool.
Every argument for shadow AI just got stronger, and every friction that used to limit it got weaker. Assume usage in your organisation rises from this week.
Bans alone will not hold. Answer with a portfolio: Copilot Chat as the free governed baseline, paid Copilot seats where tenant grounding pays, Azure OpenAI for specific high-value workflows.
Use the moment for compliance hygiene: refresh GDPR records, deliver AI Act Article 4 literacy training, and put DLP controls behind the policy rather than relying on trust.
Measure behaviour, not policy publication. Falling consumer-AI traffic is the only metric that proves your governed alternative is winning.