LLM Naming and Versioning Explained: Aliases, Snapshots and Model IDs Decoded

LLM naming and versioning is the set of conventions AI providers use to label model families, tiers, releases and the exact API identifiers that select one fixed model. A model ID such as gpt-5.5-2026-04-23, claude-sonnet-4-6, mistral-small-2603 or NVIDIA-Nemotron-3-Super-120B-A12B-BF16 packs in a family name, a version, sometimes a date, and for open models the parameter count and number format. This page explains how to read those strings, which IDs stay fixed and which move, and how nine providers document their schemes. It is written for developers, buyers and anyone who needs to know exactly which model an application is calling.

Last checked: 2026-11-15. Providers are listed alphabetically; order is not a ranking.

This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider.

Key takeaways

• A display name is not an API ID. "Claude Sonnet 4.5", "GPT-5.5" and "Mistral Large 4" are marketing names; the string sent to the API (for example claude-sonnet-4-5-20250929) is what selects the model.
• Snapshots are pinned, aliases move. A snapshot or dated ID refers to one fixed release; an alias such as gemini-flash-latest, mistral-large-latest or xAI's <modelname>-latest can be repointed to a newer release.
• The same string shape can mean different things. Anthropic documents that claude-sonnet-4-5 is an alias while claude-sonnet-4-6 is a fixed snapshot, so you must read each provider's own rules.
• Dates are encoded differently. OpenAI appends YYYY-MM-DD, Mistral uses YYMM, Cohere uses MM-YYYY, DeepSeek uses MMDD in release names, and Google preview IDs use MM-YYYY.
• Open model names carry architecture facts. On Hugging Face, "120B-A12B" means 120 billion total and about 12 billion active parameters, and suffixes such as BF16 or NVFP4 give the weight precision.

What is an LLM model name made of?

Most model identifiers combine the same building blocks, in a provider-specific order. Reading them one by one removes most of the confusion.

As of 2026-11-15, from the provider pages cited in Sources.

What is the difference between an alias and a snapshot?

A snapshot is a fixed model release. OpenAI's model pages say "Snapshots let you lock in a specific version of the model so that performance and behavior remain consistent." An alias is a convenience name that resolves to a snapshot and can be repointed. Using an alias means an application can change behaviour on the day the provider moves the pointer, without any code change.

• Pinned (snapshot) IDs: best for production workloads that need reproducible outputs, regression testing and audit evidence of which model produced a result.
• Moving (alias) IDs: convenient for prototypes and for teams that want new releases automatically, at the cost of unannounced behaviour change between test and production.
• Infrastructure still changes: Anthropic documents that model weights are fixed for a given ID but "the serving infrastructure around the model can change over time", naming the request router, safety classifiers and sampling logic. Pinning an ID fixes the model, not every part of the service.
• Retirement ends both: a pinned snapshot is fixed only until its retirement date. Anthropic states that every model ID "has its own distinct deprecation and retirement schedule". See the AI Model Retirement Guide.

How do the main providers name and version models?

The table uses identical fields for every provider. Where a provider's fetched documentation does not describe a point, the cell says so.

As of 2026-11-15. Sources: each provider's model documentation listed under Sources.

Provider notes, in alphabetical order

Alibaba Cloud (Qwen)

Model Studio API names combine family, version and tier, for example qwen3.7-plus, qwen3.8-flash and qwen3.8-omni-flash. The models page fetched does not define stable, latest or snapshot semantics, so the behaviour of undated names is not stated by Alibaba Cloud there. Open-weight Qwen repositories on Hugging Face use size names such as Qwen/Qwen3.8-27B, which the card describes as a dense model with a vision encoder.

Anthropic

Anthropic changed format with the 4.6 generation. Earlier IDs embed a snapshot date (claude-haiku-4-5-20251001); 4.6 and later IDs are dateless (claude-sonnet-4-6, claude-opus-5-5), and major releases omit the minor segment (claude-opus-5). Anthropic states that a 4.6-generation dateless ID "is not an alias. It is the snapshot" and that updated versions ship under a new model ID. Cloud platforms add prefixes: Amazon Bedrock uses anthropic.claude-sonnet-4-6, and Google Cloud uses @ before the date for older models (claude-haiku-4-5@20251001). The Models API returns a line field, and the docs advise reading it rather than inferring the family from the ID.

Cohere

Cohere IDs end in a month and year, such as command-a-03-2025 and command-a-reasoning-08-2025; the docs describe command-r-08-2024 as "delivered in August 2024". Embedding and rerank models use version numbers instead (embed-v5.0-pro, rerank-v5.0-fast). The two documented aliases, command-r and command-r-plus, point to 2024 snapshots and are deprecated as of 15 September 2025.

DeepSeek

DeepSeek keeps API names stable and changes the model behind them, recording each change in its update log. deepseek-chat moved through DeepSeek-V2, V2.5, V3, V3.1, V3.2 and V4-Flash before its scheduled discontinuation on 24 July 2026. On 10 September 2026 DeepSeek added deepseek-flash pointing to V4.1 Flash, retired deepseek-v4-flash and routed it temporarily to the new model; on 17 October 2026 deepseek-flash was updated to DeepSeek-V4.1-Flash-1017 under the same name. Release names carry a month-day code (0324, 0528, 1017).

Google (Gemini)

Google defines four version types. Stable codes such as gemini-3.6-flash "usually don't change". Preview codes such as gemini-2.5-flash-preview-09-2025 may be used for production and are deprecated with at least two weeks' notice. Latest aliases such as gemini-flash-latest change with each release, with two weeks' email notice. Experimental models are typically not suitable for production. Product names and codes can differ: "Nano Banana 2" is served as gemini-3.1-flash-image.

Mistral AI

Mistral version strings use a two-digit year and month, so 26.03 means March 2026, and API IDs append the short date (mistral-small-2506, voxtral-mini-2507). Some IDs carry prefixes such as labs- or open-. Mistral Large 4 is offered as mistral-large-4, mistral-large-4-0 and mistral-large-latest, version 26.10. Documentation page slugs such as mistral-small-4-0-26-03 differ from API IDs; the overview maps the slug mistral-small-3-2-25-06 to the API ID mistral-small-2506.

NVIDIA (Nemotron)

Nemotron repository names spell out architecture. The card for NVIDIA-Nemotron-3-Super-120B-A12B-BF16 describes the model as "120B (12B active)", and BF16 is the weight precision of that checkpoint. Sibling repositories include an NVFP4 quantised version, a Base pre-trained checkpoint and an MTPv2 checkpoint with an updated multi-token prediction head.

OpenAI

OpenAI model pages list an undated alias and dated snapshots, for example gpt-5.5 and gpt-5.5-2026-04-23. Current flagship names add a tier word after the version (gpt-6-astra, gpt-6.1-sol, gpt-6.1-terra, gpt-6-luna), and specialised lines use their own prefixes (gpt-image-2.6-sunburst, gpt-realtime-2.1, gpt-transcribe, gpt-6-cyber).

xAI

xAI documents three patterns: <modelname> is aliased to the latest stable version, <modelname>-latest to the latest version, and <modelname>-<date> refers directly to a specific release. Media models use their own lines (grok-imagine-image-2.0, grok-imagine-video-1.5).

How do you read open-weight model names?

Open-weight repositories on Hugging Face usually encode facts that closed APIs do not publish.

As of 2026-11-15. Not every publisher follows these tokens; read each card. See Mixture of Experts Explained and Open-Weight vs Closed Models Explained.

Why do model versions matter for security and governance?

• Change control: a moving alias is an untracked change to a production system. Record the resolved ID with every request and treat a model switch as a change. See AI Model Change Management.
• Inventory: register the exact IDs each application and agent uses, including cloud-specific forms such as Bedrock prefixes. See the AI Agent Registry vs Model Registry.
• Retirement planning: dated and dateless IDs each carry their own schedule, and DeepSeek routed retired names to a new model temporarily. Track provider deprecation pages.
• Evidence: model cards and system cards describe a named release, so the ID in your logs must match the card you relied on. See Model Cards and System Cards Explained.

Common failure patterns

• Assuming a dateless ID is evergreen: Anthropic calls this "a common misconception"; adopting a newer model requires changing the ID.
• Assuming a dateless ID is fixed: the opposite error. OpenAI's undated gpt-5.5, Google's -latest and DeepSeek's API names can resolve to a different release.
• Treating the marketing name as the version: Google's "Nano Banana 2" is gemini-3.1-flash-image; Mistral's documentation page slugs differ from the API IDs listed beside them.
• Parsing IDs in code: formats change between generations (Anthropic dropped dates and the Bedrock -v1 suffix). Use provider fields such as Anthropic's line where available.
• Ignoring routed retirements: a retired name that still answers may be served by a different model.

Practitioner checklist

• Pin dated or snapshot IDs in production configuration; reserve aliases for development.
• Log the model ID returned in each API response, not only the one requested.
• Keep an inventory of model IDs per application, agent and cloud platform.
• Subscribe to each provider's deprecation and change log pages and review them monthly.
• Run a regression evaluation before moving to a new snapshot or when an alias is repointed.
• For open models, record total and active parameters, precision and checkpoint type from the card.
• Match each deployed ID to the model card or system card that governs it.

Standards and references

• Anthropic: Model IDs and versioning (accessed 2026-11-15)
• Anthropic: Models overview (accessed 2026-11-15)
• Alibaba Cloud: Model Studio models (accessed 2026-11-15)
• Cohere: Models overview (accessed 2026-11-15)
• DeepSeek: API change log (accessed 2026-11-15)
• Google: Gemini models and version name patterns (accessed 2026-11-15)
• Mistral AI: Models overview and Mistral Large 4 model card (accessed 2026-11-15)
• NVIDIA: Nemotron 3 Super 120B-A12B model card (accessed 2026-11-15)
• OpenAI: GPT-5.5 model page (snapshots) (accessed 2026-11-15)
• xAI: Models and aliases (accessed 2026-11-15)

Related NHI Mgmt Group resources: AI Platforms Guide · AI Model Retirement Guide · Model Pricing Explained · Knowledge Cutoff Explained