Amazon Nova, Cohere Command and Microsoft MAI are the in-house model families of three companies that sell to enterprises through cloud platforms, and each publishes a different set of facts about its models, prices and data handling. This page compares the three families and one representative general-purpose text model from each (Amazon Nova 2 Lite, Cohere Command A+ and Microsoft MAI Thinking 1) with identical fields: IDs, release dates, context and output limits, reasoning controls, tools, open weights, prices, where each runs, model cards, provider-reported benchmarks, lifecycle dates and data handling. It is written for architects, platform teams and buyers who already use AWS, Azure or a private cloud and want the documented facts in one place. Providers are listed alphabetically (Amazon, Cohere, Microsoft) in every table; the order is not a ranking and the page gives no verdict.
Last checked: 2026-11-15.
This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider. Model details change, so confirm against each provider's current documentation.
Key takeaways
- Three different routes to market. Nova 2 Lite runs on Amazon Bedrock through cross-region inference profiles; Command A+ runs on Cohere's platform and partner clouds and is also published as open weights; MAI Thinking 1 is in public preview on Microsoft Foundry.
- Only Command A+ has open weights. Cohere publishes it under Apache 2.0 on Hugging Face; Amazon and Microsoft list no open weights for Nova 2 Lite or MAI Thinking 1 on the pages checked.
- Context windows range from 128K to 1M tokens. Nova 2 Lite lists 1M tokens with 64K output, MAI Thinking 1 lists 256K, and Command A+ lists 128K input with 64K output.
- Per-token prices are hard to find for all three. The Bedrock pricing page did not render Nova prices, Microsoft states no MAI Thinking 1 price, and Cohere's model page says Command A+ is free on trial and production keys up to rate limits.
- Nova 2 Lite has the nearest lifecycle date. Its Bedrock model card gives an end of life no sooner than 2 December 2026; Cohere lists no retirement for the Command A family and Microsoft states none for MAI Thinking 1.
How the models were selected
Each family has several models, so this page names one general-purpose text model per provider and covers the rest of each family in the next section. Amazon: Amazon's Nova models page shows Nova 2 Lite, Nova 2.5 Sonic (speech) and Nova Multimodal Embeddings, and describes Nova 2 Lite as "a fast, cost-effective reasoning model for most everyday tasks". Nova 2 Pro and Nova Premier are not shown on that page, so they are not covered. Cohere: Command A+ (command-a-plus-05-2026) is the newest live Command model in Cohere's models table and is described on its model page as Cohere's first Mixture of Experts model. Microsoft: MAI Thinking 1 is the only general-purpose reasoning model on Microsoft AI's models page; the other MAI models cover code, cyber defence, images, speech and transcription.
For each provider's full catalogue see Amazon Nova Models, Cohere Models and Microsoft MAI and Phi Models. For frontier models from other providers see Frontier AI Models Compared.
The three families at a glance
As of 2026-11-15. Sources: Amazon Nova models page, Cohere models documentation, Microsoft AI models page.
| Family | Models listed by the provider | Main channel |
|---|---|---|
| Amazon Nova | Nova 2 Lite (reasoning, text/image/video in), Nova 2.5 Sonic (speech), Nova Multimodal Embeddings | Amazon Bedrock; customisation also through Amazon SageMaker AI and Nova Forge |
| Cohere Command | Live: Command A+, Command A Reasoning, Command A, Command A Vision, Command A Translate, Command R7B, Command R (08-2024), Command R+ (08-2024) | Cohere platform and Model Vault; partner clouds; private deployment |
| Microsoft MAI | MAI Thinking, MAI Code, MAI Cyber, MAI Image, MAI Transcribe, MAI Voice, Custom Models | Microsoft Foundry; MAI Code in GitHub Copilot and VS Code; MAI Cyber in Microsoft's MDASH harness |
The three models at a glance
As of 2026-11-15. Sources: Bedrock Nova 2 Lite model card, Cohere Command A+ model page and Hugging Face card, Microsoft MAI Thinking 1 page and announcement.
| Field | Nova 2 Lite (Amazon) | Command A+ (Cohere) | MAI Thinking 1 (Microsoft) |
|---|---|---|---|
| Model ID | amazon.nova-2-lite-v1:0; profiles us., eu., jp., global. | command-a-plus-05-2026 | Not stated on the pages checked; Foundry catalogue name MAI-Thinking-1 |
| Status | Active | Live | Public preview on Foundry |
| Input modalities | Text, image, video | Text, image | Not stated by Microsoft on the pages checked |
| Output modalities | Text | Text | Not stated by Microsoft on the pages checked |
| Knowledge cutoff | October 2025 | 1 April 2025 | Not stated by Microsoft on the pages checked |
| Reasoning control | Off by default; reasoningConfig effort low, medium, high | thinking can be enabled or disabled, with token budgets | Adjustable reasoning effort; reasoning trace can be inspected |
| Architecture and parameters | Not stated by Amazon | Sparse Mixture of Experts, 218B total, 25B active; 128 experts, 8 active plus 1 shared | Sparse Mixture of Experts, about 35B active and about 1T total |
| Open weights / licence | No / proprietary | Yes / Apache 2.0 | No open weights listed on the pages checked |
| Languages | Not stated on the model card | 48, including all official EU languages | Not stated on the pages checked |
Command A+ is the only one of the three that can be downloaded and run on your own hardware; Cohere's model page lists one B200 or two H100 GPUs at W4A4 quantisation. See Open-Weight vs Closed LLMs Explained and the Open-Weight Model Security Guide.
When was each model released?
| Model | Release | Version naming | Source |
|---|---|---|---|
| Nova 2 Lite | 2 December 2025 (Bedrock launch date) | Version suffix v1:0 in the model ID | Bedrock model card |
| Command A+ | 20 May 2026 | Month and year in the ID (05-2026) | Cohere Command A+ model page |
| MAI Thinking 1 | Announcement updated as of 12 August 2026; public preview on Foundry | Version number in the product name; no API snapshot ID published | Microsoft AI announcement |
Cohere's older Command models carry dated IDs too: command-r-03-2024, command-r-plus-04-2024 and their aliases were deprecated on 15 September 2025. For a dated list across providers see the LLM Release Tracker.
How much does each model cost?
As of 2026-11-15. Sources: Amazon Bedrock pricing page and Nova 2 Lite model card, Cohere Command A+ model page, Microsoft MAI Thinking 1 page and announcement.
| Item | Nova 2 Lite | Command A+ | MAI Thinking 1 |
|---|---|---|---|
| Per-token input and output price | Not rendered on the Bedrock pricing page checked; the model card points to that page | Model page: free on trial and production keys until rate limits are reached | Not stated by Microsoft; the announcement calls it cost-efficient without figures |
| Service tiers | Standard (default), Priority (price premium), Flex (lower cost); Reserved set at account level | Trial and production keys; Model Vault for production use | Not stated on the pages checked |
| Prompt caching | Explicit and implicit; minimum 1K tokens per checkpoint, 4 checkpoints, 5-minute TTL, up to 20K cached tokens | Not stated on the model page | Not stated on the pages checked |
| Self-hosting cost | Not applicable (no weights) | Your own hardware under Apache 2.0 | Not applicable (no weights listed) |
| Consumer or free access | Not stated | Trial keys | MAI Playground, personal non-commercial use only |
The absence of a published per-token price for all three means cost comparison has to be done in each cloud's own calculator or contract. Treat any figure you find outside the provider's own pricing page with care. See Model Pricing Explained and the LLM Pricing Tracker.
What is the context window of each model?
| Limit | Nova 2 Lite | Command A+ | MAI Thinking 1 |
|---|---|---|---|
| Context window | 1M tokens | 128,000 tokens | 256K tokens (about a 600-page document, per Microsoft) |
| Max output | 64K tokens | 64,000 tokens | Not stated by Microsoft on the pages checked |
| Token counting | Count tokens not supported for this model on bedrock-runtime | Not stated on the model page | Not stated on the pages checked |
Within the Command family, Command A and Command A Reasoning list 256K tokens of context, twice Command A+, with 8K and 32K output respectively. See Context Windows Explained.
Capabilities and tools compared
| Capability | Nova 2 Lite | Command A+ | MAI Thinking 1 |
|---|---|---|---|
| APIs | Invoke and Converse on bedrock-runtime; Responses and Chat Completions not supported | Chat V2, Chat Completions, Chat V1 | Compatible with the Chat Completions API |
| Tool calling | Client-side tool calling | Tool use | Not stated on the pages checked |
| Built-in tools | Code interpreter, web grounding, remote MCP tools (Nova models page) | Not stated on the model page | Not stated on the pages checked |
| Structured output | Not supported on bedrock-runtime | Supported | Not stated as an API feature |
| Platform features | Guardrails, abuse detection, prompt optimisation; Knowledge Bases, Agents and Flows not supported | Not stated on the model page | Foundry evaluation, observability, safety and deployment features |
| Customisation | Pre-training, supervised fine-tuning and reinforcement fine-tuning through Bedrock and SageMaker AI; Nova Forge | Open weights for your own fine-tuning | Microsoft lists Custom Models as a separate MAI offering |
Tool calling and remote MCP connections act with whatever credentials the application gives them. See Amazon Bedrock Guardrails Configuration and the Computer-Use and Browser Agents guide.
Where is each model available?
| Channel | Nova 2 Lite | Command A+ | MAI Thinking 1 |
|---|---|---|---|
| Provider's own service | Amazon Bedrock | Cohere platform; Model Vault | Microsoft Foundry (public preview); MAI Playground (personal, non-commercial) |
| Amazon Web Services | Geo cross-region (US, EU, JP) and global cross-region; in-region not supported | Cohere SDK documents Bedrock and SageMaker; per-model availability not stated on that page | Not stated by Microsoft on the pages checked |
| Microsoft Azure | Not stated by Amazon | Cohere SDK documents Azure (Foundry) | Foundry |
| Other clouds | Not stated by Amazon | Oracle Cloud Infrastructure (example calls command-a-plus-05-2026) | Not stated on the pages checked |
| Private or on-premises | Not stated by Amazon | Private deployment, isolated VPC and on-premises; open weights on Hugging Face | Not stated on the pages checked |
Nova 2 Lite's geo profiles route within a geography (for example us-east-1, us-west-2 and ca-central-1 for the US profile; Frankfurt, Stockholm and Ireland among the EU regions), while the global profile can route worldwide. Choose the profile that matches your data residency rules. See the Amazon Bedrock Security Guide, the Azure OpenAI and Azure AI Foundry Security Guide and the Data Sovereignty and Residency Guide.
Model cards and system cards
| Item | Nova 2 Lite | Command A+ | MAI Thinking 1 |
|---|---|---|---|
| Document | Nova 2 Lite service card (PDF), linked from the Bedrock model card | Hugging Face model card; no safety section on it | Foundry model card (aka.ms link) and a technical paper |
| System card | Not stated by Amazon on the pages checked | Not found on the pages checked | Not stated by Microsoft on the pages checked |
| Training data statement | Not stated on the model card | Not stated on the model card | Trained without distilling from third-party models (announcement) |
See Model Cards and System Cards Explained.
Benchmarks
No evaluation on the pages checked has a score reported by each provider for its own model, so there is no shared row.
| Provider and model | Evaluation and score as reported | Source and date |
|---|---|---|
| Amazon: Nova 2 Lite | Not stated on the Bedrock model card or Nova models page checked | Checked 2026-11-15 |
| Cohere: Command A+ | Not stated on the Cohere model page or Hugging Face card checked | Checked 2026-11-15 |
| Microsoft: MAI Thinking 1 | AIME 2025: 97.0%; AIME 2026: 94.5% | Microsoft AI announcement, 12 Aug 2026 |
Microsoft's announcement also makes comparative claims against Anthropic models without numeric scores in the text; those are not reproduced here. Independent leaderboard results (LMArena, Artificial Analysis, Epoch AI) were not checked for this page. See How to Read LLM Benchmarks.
Deprecations and retirements
| Model or item | Details | Source |
|---|---|---|
| amazon.nova-2-lite-v1:0 | Active; end of life no sooner than 2 December 2026; legacy period of at least 6 months | Bedrock model card |
| command-a-plus-05-2026 | Live; no scheduled retirement for the Command A family; deprecated models stay available until shutdown only to users active within 90 days of the announcement | Cohere models and deprecations pages |
| Older Command models | command-r-03-2024, command-r-plus-04-2024, command-r, command-r-plus, command and command-light deprecated 15 September 2025 | Cohere models page |
| MAI Thinking 1 | Public preview; no retirement date stated by Microsoft (2026-11-15); Azure preview terms may apply | Microsoft AI announcement; Microsoft Foundry data privacy page |
"No sooner than" is a floor, not a date, so teams on Nova 2 Lite should watch Bedrock's lifecycle notices. Track all three on the LLM Deprecation and Retirement Tracker.
Data handling and security
| Topic | Amazon (Bedrock) | Cohere | Microsoft (Foundry) |
|---|---|---|---|
| Training on customer data | Bedrock FAQ: inputs and outputs not used to train Amazon Nova, Amazon Titan or third-party models | Enterprise customers can opt out under Data Controls; default not stated on the page | Prompts and completions not used to train generative AI foundation models without permission |
| Sharing with model providers | Not shared; providers have no access to Bedrock model deployment accounts | No prompts or generations received in private deployments | Not available to providers of Models sold by Azure |
| Retention | Content encrypted and stored at rest in the Region of use | SaaS logs deleted after 30 days; zero data retention for approved enterprise customers | Flagged content kept in a per-resource abuse monitoring store; modified abuse monitoring for approved customers |
| Encryption and location | TLS 1.2 required, 1.3 recommended; FIPS 140-3 endpoints | On-premises or isolated VPC options | AES-256 at rest, optional customer-managed keys; Standard, DataZone and Global deployment types |
| Assurance stated | Not stated on the pages checked | ISO 27001, ISO 42001, SOC 2 Type II | Not stated on the page checked |
Microsoft notes that previews may follow different privacy practices, which matters while MAI Thinking 1 is in public preview. Bedrock API keys, Foundry keys and Cohere keys are non-human identities with spend and data exposure. See LLM Data Retention and Training Policies Compared, Zero Data Retention for AI and the LLM API Key Security Guide.
How to choose a model: factual criteria
| Criterion | What the documentation shows |
|---|---|
| Need to self-host or run air-gapped | Command A+ (Apache 2.0 weights); Nova 2 Lite and MAI Thinking 1 list no weights |
| Video input | Nova 2 Lite lists text, image and video; Command A+ lists text and image; MAI Thinking 1 not stated |
| Long documents | 1M (Nova 2 Lite), 256K (MAI Thinking 1), 128K (Command A+; 256K on Command A) |
| Structured output on the native API | Supported on Command A+; not supported for Nova 2 Lite on bedrock-runtime; not stated for MAI Thinking 1 |
| Production status | Nova 2 Lite Active, Command A+ Live, MAI Thinking 1 public preview |
| Lifecycle floor | Nova 2 Lite: no sooner than 2 December 2026. Others: none stated |
| Cloud already in use | Check the availability rows above and that cloud's own catalogue |
Run your own evaluation on representative tasks, and get a written price from each channel, before deciding.
Practitioner checklist
- Record the exact model ID or inference profile (for example us.amazon.nova-2-lite-v1:0 or command-a-plus-05-2026) each application calls in the AI inventory; see the AI Register and AI Inventory Template.
- Pick the Nova inference profile (geo or global) that matches your residency rules, and record it.
- Set a review date before 2 December 2026 for any workload on Nova 2 Lite and watch for Bedrock lifecycle notices.
- Confirm in writing which preview terms apply before sending production data to MAI Thinking 1.
- For self-hosted Command A+, keep the weights, licence and model card version together and apply the controls in the Self-Hosted LLM Security Guide.
- Check Cohere's Data Controls toggle and request zero data retention if your policy needs it.
- Store every Bedrock, Foundry and Cohere key in a secrets manager, scope it to one project and set spend alerts.
Standards and references
- Amazon Bedrock: Amazon Nova 2 Lite model card (accessed 2026-11-15)
- Amazon: Amazon Nova models (accessed 2026-11-15)
- Amazon: Amazon Bedrock FAQs (accessed 2026-11-15)
- Amazon Bedrock: Data protection (accessed 2026-11-15)
- Cohere: Models overview (accessed 2026-11-15)
- Cohere: Command A+ (accessed 2026-11-15)
- Cohere Labs: command-a-plus-05-2026 model card on Hugging Face (accessed 2026-11-15)
- Cohere: Deprecations (accessed 2026-11-15)
- Cohere: Cohere works everywhere (accessed 2026-11-15)
- Cohere: Enterprise data commitments (accessed 2026-11-15)
- Microsoft AI: Models (accessed 2026-11-15)
- Microsoft AI: MAI Thinking 1 (accessed 2026-11-15)
- Microsoft AI: Introducing MAI Thinking 1 (accessed 2026-11-15)
- Microsoft Learn: Data, privacy and security for Models sold by Azure (accessed 2026-11-15)
Related NHI Mgmt Group resources: Amazon Nova 2, Cohere Command A, Microsoft MAI and Phi Models, Open-Weight Frontier Models Compared, Frontier AI Models Compared.