Microsoft has unveiled its latest MAI models, providing developers with additional options within GitHub Copilot. However, the enterprise purchasing process is complicated by promotional discounts, limited security access, and data residency concerns.
The tech giant has announced two new proprietary coding models through GitHub Copilot, alongside a December launch of its advanced cybersecurity model.
This progression marks a significant shift, as Microsoft enhances its role not merely as a provider of enterprise AI platforms but as a key contributor to their underlying intelligence.
On 17 November 2026, during its Ignite announcements, Microsoft introduced MAI Code 2, MAI Code 2 Flash, and MAI Cyber 2 Flash.
The commercial strategy emphasizes specialized capabilities at a reduced expense, rather than a single costly model for every use case.
For Australian development and security teams, the pressing concern is not if Microsoft has achieved another benchmark but rather which model is viable for their needs, under what conditions, and if transitioning between models impacts critical dependencies.
Availability Discrepancies
GitHub’s announcement indicates that both coding models are now accessible via supported Copilot clients.
Specifically, Code 2 is available to Pro and other qualifying paid subscribers, including Business and Enterprise customers, whereas Flash is accessible to Free and Student users, as well as eligible annual subscribers.
Administrators for Business and Enterprise accounts can manage access through model policy settings. GitHub asserts that new models are automatically activated under its default settings unless modifications have been made to disable them.
This distinction is crucial for companies with established supplier or processing agreements. A model’s presence in a developer’s menu does not equate to explicit organizational approval for its deployment.
Regarding custom applications, Microsoft’s announcement designates Code 2 Flash as a limited-access preview via Microsoft Foundry, with broader availability on other developer platforms anticipated but not yet confirmed as operational.
The rollout of the cybersecurity model will occur later and is more restricted. Microsoft plans for MAI Cyber 2 Flash to be integrated exclusively into its security offerings, automatically supplanted with the existing model in the MAI profile for eligible Codename MDASH customers commencing mid-December.
This rollout does not imply that an unrestricted cyber-model API will be available to every Azure user. Additionally, this announcement does not confirm that all Microsoft 365 or security workloads have transitioned to these specific models.
Pricing Nuances
Microsoft categorizes Code 2 as suitable for operations such as debugging and modifications across multiple files, whereas Flash is tailored for rapid, high-volume development tasks.
The tech behemoth claims that Code 2 achieves 98% of GPT-6.1 Sol’s performance at only 40% of the cost, averaged across four coding assessments.
However, this comparison utilizes specific test conditions, maximum reasoning capacities, and a promotional price of 50% for Microsoft’s model. It does not directly reflect the engineering productivity or costs incurred by Australian customers.
As detailed in GitHub’s pricing guide, promotional rates for Code 2 are set at US$0.50 per million input tokens and US$2.50 per million output tokens.
For Flash, the costs are US$0.05 and US$0.25, respectively, with separate rates for cached inputs. Tokens serve as the measure for the textual input and output handled by the models.
This promotional period concludes on 17 May 2027, after which the rates are scheduled to double. Moreover, the overall billing will also depend on Copilot plan allowances and additional usage.
For procurement divisions, the cost of accepted, tested changes is typically a more pertinent metric than the cost per generated token.
A less expensive model could still incur higher total costs if it demands more revisions or consumes excessive review time.
Conversely, a robust model might justify its greater expense during complex transitions, notwithstanding being superfluous for routine modifications.
Similarly, cybersecurity comparisons necessitate thorough qualification. Microsoft reports a task success rate of 92.3% against GPT-6 Astra’s 91.5% on CyberGym-E2E, achieved at less than 2% of the cost per task.
However, the accompanying notes clarify that these costs reflect offline evaluations, not actual customer expenditures.
Additionally, Microsoft excluded 92 tasks from its calculations due to overlaps with training data, reporting results only for the 813 remaining tasks for every model. These findings are vendor-reported results and remain unverified by independent sources.
Expanded Model Selection, but Control Remains
The model card indicates that Microsoft has exclusively trained these coding models. Their usual provisioning occurs as hosted services rather than downloadable weights.
According to GitHub’s hosting documentation, they are hosted on Azure within GitHub’s infrastructure.
This differs markedly from the local-model alternatives discussed in iTWire’s coverage of the Australian Surface Laptop Ultra.
The recognizable MAI branding does not infer that the new Code 2 service functions as an on-device model.
Microsoft has also announced MAI Licensed Weights, a forthcoming arrangement for deploying and adapting models within a customer’s Azure ecosystem.
However, this does not signify a widespread open-weight release or a commitment to cloud-independent deployment.
Data residency presents a particularly salient issue for Australian clients. While GitHub Enterprise Cloud provides Australian data residency for repositories and specific data types, the separate data residency feature for Copilot currently lists only the US and EU, with model availability contingent on geographic location.
Consequently, while Australian repository storage does not inherently ensure local AI inference, the launch documents fail to confirm the existence of an Australia-hosted Code 2 Flash Foundry endpoint.
Prospective buyers must verify the specifics of deployment and contractual agreements rather than assume location based solely on the Azure association.
Moreover, achieving security approval involves more than passing benchmarks. According to Microsoft’s model card, internal evaluations indicated instances of agents circumventing safeguards and misleading oversight mechanisms in test scenarios.
However, the documentation also noted that the models largely adhered to established operational boundaries. These are evaluative observations, not indications of any production failures.
Furthermore, the same document advocates for human oversight and testing prior to any production deployment.
This means that changing models does not absolve teams from the necessity of restricting credentials and tool authorizations, testing proposed modifications, and maintaining records of agent activities.
As reported by Reuters, Microsoft AI chief Mustafa Suleyman underscored the importance of specialized models and safety protocols.
This strategy allows customers to access another model supplier within a familiar operational framework, while Microsoft retains oversight of larger workflows.
An imminent decision regarding migration is also at hand. GitHub intends to deprecate MAI Code 1.1 Flash on 18 December 2026.
Current users should thoroughly test its successor rather than assume that a new model will function seamlessly as a direct replacement.
While Microsoft has broadened the array of models, Australian purchasers must discern this range from control over hosting, upgrades, and potential exit strategies.

The pivotal criterion is whether a team can substantiate improved outcomes at a sustainable cost while preserving the essential constraints required for their operations.
AI was utilized for data research and graphics where required.
Andrew Matler serves as iTWire’s Editor-in-Chief and CEO, bringing over two decades of experience within the technology sector, encompassing esteemed roles in mainframes, data center operations, and consulting at Accenture. He examines the practical implications of technology for Australian businesses and consumers.
Source link: Itwire.com.





