GPT-7 'Bel' and the Collapse of Model Naming as Signal

01 When the Name Stopped Describing the Model

Consider a week of 2026 model news: a preview of GPT-7 under the pet name “Bel,” a mid-cycle upgrade to a model called Mythos landing at version 5.1, a numbered Mistral Large 4, and a coding product renamed after a person. None of those names tells a buyer what the model does, how it compares to its predecessor, whether it is a new architecture or a fine-tune, or what it costs. A generation ago, software versioning existed precisely to answer those questions: the identifier communicated compatibility and change magnitude. In 2026’s model market, the identifier communicates positioning, availability windows, and pricing tiers — everything except engineering meaning.

The collapse happened gradually, then suddenly. Version numbers stretched past the point of intuition (what does 4.1o-turbo-preview imply that 4.5 does not?), skipped digits appeared (3.5 to 4, then past some numbers entirely), and codenames migrated from internal project labels to consumer-facing brands. A large language model, an AI system trained on vast text to generate and analyze language, now ships under a name chosen by marketing for a launch, then sometimes renamed for general availability — a practice the release cycles of conventional software abandoned decades ago for good reason.

02 Mechanism: Versioning Schemes Versus Marketing Versions

The classical schemes are worth restating because each encodes a promise. Semantic versioning — major.minor.patch — promises that the major number moves only on breaking changes; date-based versioning promises freshness and lets buyers judge staleness at a glance; build identifiers promise reproducibility. Each scheme is a contract with the buyer’s operational planning: a database administrator can schedule an upgrade around a major-version bump because the number carries information. Software versioning persists in enterprise infrastructure precisely because it survived contact with procurement, audit, and change management.

Model naming answers to none of those masters. The buyer of a model subscription does not integrate a specific version into a controlled system — or rather, does not want to, because the vendor wants to improve the product continuously. So names are optimized for launch-day salience: a new word for a new flagship (a “Bel,” a “Mythos”) generates more coverage than a decimal increment, and an ambiguous name lets the vendor ship improvements under the same banner without the commitment semantics of a version. The mechanism is straightforward: naming is the marketing surface of continuous deployment, and marketing names are designed to be remembered, not parsed.

03 Evidence: A Year of Names That Encode Nothing

Count the names and the pattern hardens. Across 2024-2026, the major labs published dozens of distinct flagship-tier identifiers while shipping only a handful of genuine architecture generations — the name volume rose far faster than the underlying model-lineage count. Tally the styles and fewer than half of 2026’s headline launches carried a strict version number at all; codename-first and tier-only namings took the rest. Some vendors now maintain parallel naming systems for the same artifacts: an API model string, a product name, and a conversation shorthand, all drifting independently.

The launch-style tally is the sharper signal. Version-numbered launches dominated 2023, when numbering helped a new category look progressive. By 2026 the codename and adjective styles lead — a reversal that tracks the industry’s pivot from “look how new this technology is” to “look how distinct this brand is.” The numbers stopped being an asset when increments slowed and comparisons became unflattering; the words took over because words make no measurable promises.

Even the numbers that survive have lost their scale. A jump from version 4 to version 5 once implied an architecture change; in the model market it has labeled, variously, a new training run on similar architecture, a rebrand after a leadership change, and a bundle of fine-tunes. Buyers respond rationally: version numbers in model names are now treated as marketing ordinality — a hint at recency — and nothing more.

04 What the Confusion Costs Buyers

The confusion is billable. Enterprises building on model APIs maintain internal compatibility matrices — spreadsheets mapping vendor product names to deprecation dates, context windows, and behavior-change notes — that add real engineering headcount to every integration. Deprecation whiplash, where a named model the business depends on acquires a new name and changed behavior mid-contract, has become a standard line in AI procurement questionnaires. The multi-vendor routing layer, which abstracts model names behind a capability interface, is now a product category precisely because names became unreliable identifiers: buyers pay an abstraction tax to avoid re-plumbing integrations every launch cycle.

There is also an evaluation cost. Benchmark results attach to names, and when names migrate or fork, the public record of what was measured against what decays — a leaderboard entry for a snapshot that no longer ships is archaeology within months. For the buyer choosing a model for a regulated workflow, the question “exactly which artifact did we validate?” has become surprisingly hard to answer, and answering it wrong means re-validating under time pressure when the vendor renames or deprecates. Naming chaos does not just annoy; it transfers the vendor’s release-velocity problem onto the buyer’s compliance ledger.

05 How the Industry Is Re-Learning to Communicate Change

The correction is arriving from the edges. Snapshot and date identifiers — pinning a model as “the March snapshot” with a frozen behavior contract — have become the enterprise-grade option at every vendor that offers them, and API strings increasingly carry explicit date suffixes even when marketing names do not. Deprecation notices with migration guides, once rare, are now table stakes for landing enterprise deals. The pattern mirrors what cloud platforms learned a decade earlier: marketing names for the storefront, immutable identifiers for the plumbing, and a documented mapping between them.

Third-party infrastructure absorbs the rest. Router services, evaluation harnesses, and model registries maintain the crosswalks between names and artifacts, versioning on the buyer’s behalf. The industry has effectively outsourced semantic versioning to its tooling layer — an elegant-enough equilibrium, but one with a price and a dependency: the buyer now needs a middleman to know what it bought. When the identifier layer of a market becomes a product category, the identifiers have failed at their one job.

06 Limits: Why Naming Will Not Be Fixed Voluntarily

Do not expect a reform movement. The economics run the other way: salient names raise launch coverage, ambiguous names preserve pricing flexibility, and pet names humanize products in a crowded market. A vendor that unilaterally adopted strict semantic versioning would signal restraint while competitors signaled excitement — a classic coordination failure in which the disciplined player pays the advertising cost of everyone else’s chaos. The pressure on names is also structural: continuous deployment is genuinely how modern models improve, and a naming scheme built for discrete releases does not map cleanly onto a product that changes weekly.

There is a deeper ambiguity worth naming in turn: for a model, the version boundary is itself fuzzy. A fine-tune, a safety retrain, a quantization, and a system-prompt update all change behavior without changing weights, and vendors reasonably disagree about which crosses the versioning threshold. The classical schemes assume the vendor can enumerate its changes; the model development cycle makes the enumeration expensive and strategically awkward. Naming chaos is not just marketing culture — it is the visible end of a genuine measurement problem.

07 Legacy: Identifiers as Infrastructure

The lasting resolution is already visible in outline: two naming layers, permanent and parallel. The public layer will keep generating brands, pet names, and ordinals because that layer’s job is attention. The infrastructure layer — API strings, snapshots, registry hashes — will drift toward the immutability conventions that made modern software supply chains auditable, because enterprise buyers now demand the behavior contracts. The gap between the two layers is where the routing, registry, and evaluation vendors live, and their existence is the market’s confession that the public layer cannot be trusted alone.

For the buyer, the practical doctrine has crystallized: never integrate a brand. Pin artifacts, date-stamp validations, contract deprecation terms, and treat every launch-week name as a claim to be verified against a frozen snapshot. “Bel” is a lovely name, and it means nothing — which is precisely why the unglamorous identifier underneath it is becoming one of the most consequential lines in an enterprise AI contract. The version number is dead; long live the version control.