The API Families
Lesson 3 of 3 in Model Families.
The other half of the map never hands you weights at all. Three Closed weights API families dominate it — GPT from OpenAI, Claude from Anthropic, Gemini from Google — each licensed as proprietary under API terms and consumed as managed endpoints. With no weights to inspect, the vendor’s docs pages and model cards are not just documentation: they are the entire knowable surface of the product. Reading them precisely — including noticing what they do not say — is the skill of this lesson.
Start with how each vendor organizes its own docs, because the organization is the durable signal. OpenAI’s models page arranges the GPT family as flagship models (GPT-6 Astra, GPT-5.6 Sol, GPT-5.6 Terra, GPT-5.6 Luna) plus specialized tiers — security-focused Daybreak models, a life-sciences model, and image, realtime and voice, speech, and transcription lines, with a realtime model already marked deprecated on the page. The docs call the flagship GPT-6 Astra “Our most capable model, built for the hardest end-to-end work”, listing a “1.05M” context window and “128K tokens” max output — as stated on the page — and note that “All latest OpenAI models support text and image input, text output, multilingual capabilities, and vision.” (Source: developers.openai.com/api/docs/models.)
Claude gets a one-line family definition straight from Anthropic’s docs: “Claude is a family of state-of-the-art large language models developed by Anthropic”. The page lists a current lineup — Claude Fable 5.1, Claude Opus 5, Claude Sonnet 5, Claude Haiku 4.5 — and, instructively, a legacy list of models that remain available. Documented specs, as stated: a “1M tokens” context window and 128K max output for Claude Fable 5.1, Claude Opus 5, and Claude Sonnet 5 (200K and 64K for Claude Haiku 4.5), “Adaptive” thinking steered by an effort parameter, and the blanket claim that “All current models support text and image input, text output, multilingual capabilities, vision, and tool use.” (Source: platform.claude.com/docs/en/about-claude/models/overview.)
Gemini is the widest catalog of the three: Google’s docs list a Gemini 3 family and a Gemini 2.5 family of text models alongside dedicated image, video and music, and specialized lines (computer use, deep research, embeddings, robotics). The documented trait: Gemini 3.8 Flash is “Our most intelligent Flash model, engineered for long-horizon software engineering, autonomous agents”, and the family carries dedicated Live API voice models (“Default Live API model for most low-latency voice agent experiences without reasoning delays”), including an “Extended Thinking” Live variant. (Source: ai.google.dev/gemini-api/docs/models.)
| Family | Developer | Weights & license | How the docs organize it | Documented trait (as stated) |
|---|---|---|---|---|
GPT | OpenAI | Closed-weights; proprietary (API terms) | Flagship tier (GPT-6 Astra, GPT-5.6 Sol / Terra / Luna) plus specialized tiers: security (Daybreak), life sciences, image, realtime and voice, speech, transcription — deprecations marked in place | “Our most capable model, built for the hardest end-to-end work” (GPT-6 Astra); “1.05M” context window and “128K tokens” max output, as stated |
Claude | Anthropic | Closed-weights; proprietary (API terms) | Current lineup (Claude Fable 5.1, Claude Opus 5, Claude Sonnet 5, Claude Haiku 4.5) plus an explicit legacy list that remains available | “1M tokens” context and 128K max output for the top three (200K / 64K for Claude Haiku 4.5); “Adaptive” thinking steered by an effort parameter |
Gemini | Closed-weights; proprietary (API terms) | Gemini 3 and Gemini 2.5 families plus image, video/music, and specialized lines (computer use, research, embeddings, robotics) | “Our most intelligent Flash model, engineered for long-horizon software engineering, autonomous agents” (Gemini 3.8 Flash); dedicated Live voice models incl. an “Extended Thinking” variant |
Now the epistemology. With an open-weights family you can, in principle, verify claims yourself — load the weights, count the parameters, inspect the tokenizer. With a closed family, what the vendor documents is what you can know, and the boundary is sharp.
On the knowable side: the lineup and its tiers, documented context windows and output caps (quoted as stated — note how this site writes “1.05M” and “1M tokens” exactly as the pages do), documented input and output modalities, thinking or effort controls the docs describe, Deprecation status and legacy lists, and the usage policies you agree to. This is the legacy of the model-card idea — Mitchell et al. proposed in 2018 (arXiv:1810.03993) that trained models should ship with standardized reporting, and today’s vendor docs pages are that idea grown into a product interface.
On the unknowable side: parameter counts, architecture, training-data composition, and anything else the vendor chooses not to state. Third-party estimates of these circulate constantly; treat them as rumors with confidence intervals, and never let one into a design document as a fact. The professional habit is precise agnosticism: cite what the page says, date it, and say “not documented” out loud when that is the truth.
Reading a closed family without inside knowledge
Four techniques turn vendor docs into a reliable instrument:
Read the tiering, not the names. Names churn; the shape persists. All three vendors’ pages express the same structure — a flagship tier, faster and cheaper tiers, and specialized lines — and that structure is what you map workloads onto. When a new release lands, the first question is not “how good is it?” but “which slot in the tiering does it fill?”
Read the legacy and deprecation tables first. They are the vendor telling you, in writing, the expected half-life of a pinned model id. A family whose docs keep legacy models available is making a different operational promise than one that removes them — and both promises change, which is why the read gets a date.
Quote limits as stated. Docs express the same concept differently across vendors — one page writes a context window as “1.05M”, another as “1M tokens”; output caps may be listed per model or per tier. Normalizing these numbers in your head is how transcription errors are born; quote them with their units and their source.
Snapshot what you read. The pages are living documents with no changelog you control. A dated copy — or a fact-table row with a source URL, as this site keeps — is the only way to later prove what the interface said when you decided. The Release Log is this site’s running diff; keep an equivalent for the families you build on.
In production
Choosing a family is never just a quality call — it cascades into your tokenizer (and therefore your token meter), your serving options, your fine-tuning paths, and whose terms govern the workload.
AWS
On AWS, the family decision selects the serving menu: a closed API family is consumable only as a managed endpoint, under the vendor’s API terms and deprecation calendar, through whatever catalog availability exists; an open-weights family can additionally be self-hosted on GPU instances or imported as custom weights, with the license you accepted at download governing what you may build. Either way, the family’s tokenizer defines the unit your budgets and quotas are denominated in.
Azure
Azure’s model catalog puts both worlds on one screen — partnered API families next to open-weights families deployable to managed endpoints or your own compute — which makes the cascade visible: pick a family and you have picked a cost model (per-token API versus hosting you operate), a fine-tuning surface (what the vendor or the license permits), and a compliance surface (API terms, community-license conditions, or both).
Google Cloud
On Google Cloud, the same fork runs between first-party API families and the open-weights entries in the model garden: API families rotate on the vendor’s deprecation schedule, while open families rotate when you decide — at the price of operating them yourself. Whichever side you pick, keep the family-level facts (license lineage, documented limits, release cadence) in a dated registry the whole team cites, exactly as this domain does.
Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.