- Post History
- Subscribe to RSS Feed
- Mark as New
- Mark as Read
- Bookmark
- Subscribe
- Printer Friendly Page
- Report Inappropriate Content
2 hours ago - edited 2 hours ago
What is this update?
ServiceNow's third-party model provider integrations now support new model versions from Microsoft Azure OpenAI and Anthropic Claude via AWS, released March 12, 2026. Customers are not automatically migrated — this article explains what's new, how to upgrade, and how to validate your custom skills before switching.
On March 12, 2026, ServiceNow added support for the latest model versions from Microsoft Azure OpenAI and Anthropic Claude via AWS across our third-party model provider integrations. These new models deliver meaningful improvements in reasoning, speed, coding quality, and efficiency — and are now available for customers on Zurich Patch 7 (ZP7) who have the latest Now Assist Admin Console and Generative AI Controller apps installed.
Previous model versions remain available and no customer is switched automatically. Upgrading is a deliberate, customer-initiated action for this upgrade. We recommend testing in a sub-production environment before switching production workloads.
📋 In this article:
✓ New models released March 12, 2026
✓ Key benefits by model
✓ How to upgrade to the new versions
✓ How to test OOTB and custom skills before switching
✓ What to do if quality appears to regress
Five new model versions are now available across Microsoft Azure OpenAI and Anthropic Claude via AWS providers. See the Third-Party LLM Model Card for full technical specifications, supported features, and regional availability.
GPT-5.2 is OpenAI's latest flagship update in the GPT-5 family, delivering stronger reasoning, higher accuracy, and improved multimodal understanding compared to earlier GPT-5 versions. GPT-5-mini is a compact, lower-latency variant optimized for well-defined tasks and high-volume workloads.
GPT-5.2 — Key Benefits
- Significantly reduced hallucination rates compared to prior GPT-5 versions
- Improved performance on math, coding, and multimodal reasoning benchmarks
- Adaptive reasoning depth — scales computational effort based on task complexity
- More measured, grounded response style with improved instruction-following
GPT-5-mini — Key Benefits
- Faster, more efficient alternative to GPT-5.2 for high-volume tasks
- Retains the same instruction-following and safety tuning as the full GPT-5 family
- Ideal for well-scoped, lower-complexity skill prompts where latency matters
Claude Haiku 4.5 is Anthropic's fastest and most efficient model in the Claude 4 family. It delivers near-frontier performance — roughly on par with Claude Sonnet 4, and is the first Haiku model to support extended thinking.
Claude Haiku 4.5 — Key Benefits
- Up to 4–5× faster than Claude Sonnet 4.5, optimized for low-latency, high-throughput workloads
- First Haiku model with extended thinking — enables complex multi-step reasoning at speed
- Near-frontier coding quality: achieves ~90% of Sonnet 4.5's performance on agentic coding benchmarks
- First Haiku model with context awareness — tracks remaining context window across long sessions
- Strong alignment results — lower rate of misaligned behaviors than its predecessor Haiku 3.5
Claude Sonnet 4.5 is Anthropic's best model for complex agents and coding, delivering significant improvements across the entire development lifecycle. It balances frontier-level intelligence with practical speed, making it the recommended choice for most production workloads requiring depth of reasoning.
Claude Sonnet 4.5 — Key Benefits
- State-of-the-art coding benchmark performance, with advanced planning and system design capabilities
- Significantly better on complex agents, multi-step reasoning, and ambiguous instructions than prior Sonnet versions
- Improved security engineering: more robust vulnerability detection and secure coding practices
- Concise, direct communication style optimized for agentic workflow momentum
- Extended thinking support for deep reasoning on the most complex tasks
Claude Opus is Anthropic's most capable model, designed for the highest-complexity reasoning tasks. In the ServiceNow platform, it is available exclusively for Build Agent use cases.
⚙️ Availability Note: Claude Opus is supported for Build Agent workflows only. It is not available for general OOTB skills or custom skill assignments. Refer to the Third-Party LLM Model Card for full availability details.
Customers are not switched automatically. Previous model versions remain fully available. To upgrade to the new versions, follow the steps in the platform documentation below.
Before upgrading, confirm your environment meets all of the following requirements:
- Zurich Patch 7 (ZP7) — Instance must be on ZP7 or later
- Now Assist Admin Console — Latest version required
- Generative AI Controller — Latest version required
Follow the ServiceNow documentation to select and activate the new model version for your third-party provider:
👉 Manage model versions — ServiceNow Docs
We strongly recommend upgrading and testing in a sub-production instance first before enabling new models in production. See the testing guidance in the next section.
Before upgrading your production instance, we encourage customers to validate skill quality in sub-production. Testing scope depends on whether you use OOTB skills, custom skills, or both.
ServiceNow has pre-tested all OOTB skills, AI agents, and agentic workflows with the new model versions. No customer action is required for OOTB skills.
Skill performance depends on the combination of the underlying model and the instructions (prompts) provided to it. ServiceNow owns and maintains all OOTB skills, allowing us to thoroughly test and validate their behavior with upgraded models before release.
Customers who wish to evaluate third-party model provider performance across select OOTB and custom skills can do so using automated evaluations in Now Assist Skill Kit for select OOTB skills and custom skills. See this walkthrough video for guidance.
Custom skills use customer-authored prompts, which may behave differently with a new model. ServiceNow has limited visibility into these prompts, so customers should validate them before switching.
Before upgrading, test all active custom skills that use a third-party model provider — using either full evaluations or spot-checks — and republish if tuning is needed.
|
1
|
Upgrade your instance and applications Ensure your sub-production instance is on ZP7 with the latest Now Assist Admin Console and Generative AI Controller apps before proceeding. |
|
2
|
Clone your prompt Open your impacted skill in Now Assist Skill Kit. In the Prompt Editor, locate your current published prompt and clone it using the "Clone Prompt" icon. In the Clone Prompt form, select the appropriate third-party provider and model version from the dropdowns. Save and finalize the new prompt. |
|
3
|
Assess skill quality Run quality evaluations on both your current prompt (legacy model) and the cloned prompt (new model). See Evaluate a Prompt in the platform documentation for guidance on running full evaluations. This is the preferred method. If you have previously run evaluations on this skill, this step may be skipped. |
|
4
|
Tune custom prompts if needed If overall quality has improved, no action is needed. If regressions are identified or responses do not meet expectations, tune prompts as needed and reevaluate. Iterate until outcomes are acceptable. |
|
5
|
Finalize and publish your skill Once quality meets your standards, finalize and publish the upgraded prompt. Published changes will be reflected in production. See Finalize and publish a skill for details. |
Short answer: ServiceNow has pre-tested all OOTB skills, AI agents, and agentic workflows with the new models. Custom skills, however, use customer-authored prompts, which may behave differently and should be validated after the upgrade.
Skill performance depends on the combination of the underlying model and the instructions (prompts) provided to it. ServiceNow owns and maintains all OOTB skills, allowing us to thoroughly test and validate their behavior with upgraded models. Custom skills are built and owned by customers using their own prompts — ServiceNow has limited visibility and ability to test these. While new models are expected to deliver higher quality across both OOTB and custom skills, untested prompt-model combinations may lead to unexpected behaviors. We recommend validating custom skills after the upgrade to ensure prompt quality and performance remain consistent.
Short answer: Tune prompts, retest, and publish once satisfied.
If evaluations or spot-checks indicate regressions, iterate on prompt tuning and reevaluate until outcomes meet expectations. When quality is acceptable, finalize and publish.
If degradations persist after reasonable tuning efforts, escalate through your standard support channels.
Yes. Customers can evaluate third-party model providers for select OOTB skills and custom skills before upgrading. This allows you to compare model output quality across providers in your own environment.
👉 Watch the third-party model provider evaluation walkthrough
| Model | Provider | Best for | Available |
|---|---|---|---|
| GPT-5.2 | Microsoft Azure OpenAI | Complex reasoning, high-quality output, reduced hallucinations | March 2026 |
| GPT-5-mini | Microsoft Azure OpenAI | High-volume, lower-complexity tasks requiring speed and efficiency | March 2026 |
| Claude Haiku 4.5 | Anthropic Claude via AWS | Real-time, low-latency workloads; high-throughput agentic subtasks | March 2026 |
| Claude Sonnet 4.5 | Anthropic Claude via AWS | Complex agents, coding, multi-step reasoning, ambiguous instructions | March 2026 |
| Claude Opus | Anthropic Claude via AWS | Build Agent only — highest-complexity reasoning tasks | March 2026 |
Summary: Five new model versions released March 12, 2026 across Microsoft Azure OpenAI and Anthropic Claude via AWS providers. All previous versions remain available. Customer upgrade is opt-in. Claude Opus is available for Build Agent use only.
📎 Related Resources
- Third-Party LLM Model Card — Full model specifications and availability
- Manage model versions — How to upgrade to the latest model
- Evaluating third-party model providers — Walkthrough video
