MODEL DOSSIER — CHATGPT / OPENAI

What it is

ChatGPT is the product that made this industry exist. OpenAI's November 2022 launch turned large language models from a research curiosity into a consumer category, and three and a half years later the company still holds the position that launch bought: the default. The biggest user base, the broadest product surface, the deepest brand penetration — when a normal person says "AI," they mean ChatGPT.

That default status is the core of the buying decision. OpenAI's offer is not "the best model for your specific task" — on any given workload, a competitor may beat it. The offer is breadth: one subscription that talks, sees, listens, generates images and video, browses, codes, runs agents, remembers you, and plugs into an ecosystem of custom bots and integrations larger than everyone else's combined. The trade-off, documented below, is churn — products, APIs, and model names turn over faster here than anywhere else in the industry.

---

The family

OpenAI's model naming has been through several regimes; here is the current state, cut to what matters. As of July 2026: ⚑ unverified

The GPT-5.6 family shipped publicly on July 9, 2026 — days ago, as this dossier is written — replacing the GPT-5.3/5.4/5.5 lineup as the recommended tier. ⚑ unverified Expect the older models to be deprecated on OpenAI's usual aggressive schedule.

Where did the o-series go? If you last paid attention in 2024–2025, you remember two parallel product lines: GPT-4-class general models and the o-series (o1, o3, o4-mini) "reasoning" models that thought step-by-step before answering. That split is over. GPT-5, launched in 2025, folded reasoning into the main line: a router decides per-request whether to answer fast or think hard, and the "Thinking" and "Pro" variants are the o-series' descendants under unified branding. ⚑ unverified The concept survives — reasoning effort is now a dial, not a separate model family — but nobody should be building on o-series model IDs in 2026.

Two practical notes on the table:

The three-tier structure mirrors the industry pattern. Sol/Terra/Luna maps to Anthropic's Opus/Sonnet/Haiku and Google's Pro/Flash split: a frontier model for the hardest work, a mid-tier default, a cheap volume tier. Cross-vendor comparison shopping now happens tier-to-tier, and at the frontier tier OpenAI's $5/$30 undercuts Anthropic's Fable 5 ($10/$50) while landing near Opus 4.8 ($5/$25). ⚑ unverified

Model turnover is fast and enforced. GPT-5.1 models were removed from ChatGPT in March 2026, roughly four months after the 5.1 era. ⚑ unverified If you build on OpenAI, version-migration is a recurring maintenance line item, not a rare event.

How the lineup got here

A compressed lineage, because the naming history genuinely confuses buyers. GPT-3.5 (2022) launched ChatGPT. GPT-4 (March 2023) set the frontier bar for two years. GPT-4o (2024) merged text, vision, and audio into one "omni" model and became the workhorse. Then came the fork: o1 (late 2024) introduced chain-of-thought reasoning as a separate product line, followed by o3 and o4-mini in 2025 — models that traded latency for depth on math, code, and science. GPT-5 (August 2025) ended the fork by putting a router in front of both behaviors: one product decides per-query whether to answer instantly or reason at length. ⚑ unverified The 5.x point releases since (5.1 through 5.5, at a cadence of roughly one every couple of months) refined that architecture, and GPT-5.6 (July 2026) reorganized the family into the named Sol/Terra/Luna tiers. ⚑ unverified

Two takeaways from the history. First, OpenAI iterates in public and re-brands aggressively — any tutorial, benchmark, or integration more than six months old likely references a model that no longer exists. Second, the unified-reasoning architecture is the durable idea: every major lab has now converged on "one model, variable thinking depth," and OpenAI got there by folding a whole product line into the flagship.

---

ChatGPT — the product tiers

ChatGPT's subscription ladder has grown to seven rungs. ⚑ unverified

Read the ladder honestly and two things stand out. First, OpenAI is segmenting hard in both directions — an $8 tier with ads-supported free below it, and a $100 rung inserted between $20 and $200 — which is what a company does when it needs revenue per user to rise. Second, the $20 Plus tier remains the best-known deal in consumer AI: for most individuals it is the right amount of ChatGPT.

What the paid product actually includes, and where OpenAI genuinely leads:

• Multimodal breadth. Native image generation, Sora video generation, real-time voice conversation, screen and camera input — no competitor bundles this much modality in one $20 subscription.
• Custom GPTs and ecosystem. A store of user-built bots, connectors to consumer and workplace apps, and MCP support (adopted March 2025) for standardized tool integrations. ⚑ unverified
• Memory and personalization. ChatGPT's cross-conversation memory is the most developed in the category — genuinely useful, and worth reviewing in settings if you care what it retains.
• Agent features. Computer-using agents, deep research, and scheduled tasks have moved from labs demos into the mainstream tiers.

Worth dwelling on the multimodal point, because it is the moat. Image generation inside ChatGPT (the GPT-image line, successor to DALL-E) handles text rendering, instruction-following edits, and style control well enough that it displaced standalone image tools for most casual use. Sora, OpenAI's video model, ships both inside ChatGPT and as its own app, and while professional video work still goes to dedicated tools, nothing else puts usable video generation inside a general assistant subscription. Voice mode is a real-time conversation — interruptible, expressive, camera-aware on mobile — not a text-to-speech bolt-on. Individually, each of these has a stronger specialist competitor. Bundled at $20, they don't. ⚑ unverified

The custom GPT ecosystem deserves one honest paragraph too. The GPT Store hosts millions of user-built bots, and the long tail is junk — thin prompt-wrappers with SEO names. But the mechanism matters for teams: a custom GPT is the fastest way to package instructions, files, and tool connections into a shareable internal assistant with zero code, and inside Business/Enterprise workspaces that is how non-technical departments actually deploy AI. ⚑ unverified

---

The API and platform

OpenAI's developer platform is the industry's largest by adoption, and mid-2026 finds it in the middle of a significant consolidation.

The Responses API is the present. It merges Chat Completions with built-in tools — web search, file search, computer use, code execution — into one call. The older Assistants API is deprecated with a mid-2026 sunset. ⚑ unverified New builds should target Responses; anything on Assistants needs a migration plan now.

The Agents SDK is the agent framework. Open source, and notably model-agnostic — it can orchestrate Anthropic, Google, or open-weight models alongside OpenAI's. It is the recommended path for code-defined agent workflows.

AgentKit is a cautionary tale. Launched October 2025 as the "complete platform" for visual agent building, its Agent Builder and Evals components were announced for shutdown in June 2026, gone by November 30, 2026 — a roughly one-year product life. ⚑ unverified ChatKit (embeddable chat UI) survives. The lesson for buyers is structural, not incidental: OpenAI ships fast, learns in public, and retires products fast. Build on the primitives (models, Responses API, Agents SDK), not the scaffolding.

Platform strengths worth naming: the largest third-party library ecosystem (nearly every AI tool integrates OpenAI first), fine-tuning support across tiers, and availability through Microsoft Azure for enterprises that procure there. ⚑ unverified

The cost levers matter as much as the sticker prices. The Batch API halves the rate on anything that can tolerate up to a day's turnaround — classification backlogs, embedding runs, offline evaluation. ⚑ unverified Automatic prompt caching discounts repeated input prefixes, which is most of the bill in any chat or agent workload that resends history each turn. ⚑ unverified And because OpenAI's API shape is the de facto industry interface — most competitors, including open-weight hosts, expose OpenAI-compatible endpoints — code written against this API is the most portable in the business. That is a genuine hedge: the integration you build for GPT-5.6 can usually be pointed at a rival model by changing a base URL, which keeps OpenAI honest on price in a way that no contract clause does.

---

The enterprise angle

OpenAI's enterprise pitch is distribution and familiarity: your employees already use ChatGPT, so sanction it, govern it, and stop the shadow-IT version. That pitch works — ChatGPT Enterprise and Business seats are the most widely deployed corporate AI assistant. ⚑ unverified

The mechanics: Business at $25/seat is self-serve with training-data exclusion and admin controls; Enterprise adds SSO, compliance certifications, longer context, and custom terms, but carries a 150-seat minimum and an unpublished, negotiated price — reported contracts cluster around $45–75 per user per month. ⚑ unverified Procurement leverage exists; use it.

Two honest cautions for enterprise buyers. The Microsoft relationship cuts both ways: Azure OpenAI is a mature procurement path, but Microsoft Copilot is also a competing wrapper around the same models, and many organizations end up paying for both. And OpenAI's pace of deprecation — models removed months after launch, platform products sunset within a year — means enterprise deployments need version-pinning discipline and a standing migration budget that slower-moving vendors don't demand.

There is also a balance-sheet dimension enterprises should at least register. OpenAI has committed to compute buildouts measured in the hundreds of billions of dollars — the Stargate program and a lattice of chip, cloud, and data-center deals — against revenue that, while growing fast, does not yet cover that spend. ⚑ unverified That gap is the central exhibit in the broader is-this-a-bubble debate, and this desk covers it elsewhere. For a procurement decision the implication is narrower: OpenAI's aggressive monetization moves (ads, tier proliferation, price changes) are not random product churn — they are what the financing requires. Price your dependence accordingly.

---

Honest limits

Product churn is the tax on breadth. Assistants API deprecated, AgentKit half-retired within a year, model families turned over every few months, seven subscription tiers where there were three. None of this is fatal; all of it is maintenance cost that competitors with narrower, steadier lineups don't impose.

Depth on professional work is contested. On agentic coding specifically — the highest-value professional workload — Anthropic's Claude Code has set the pace, and developer sentiment has followed. OpenAI's Codex agent competes, and GPT-5.6 Sol is explicitly aimed at this gap, but "default consumer choice" and "default professional coding choice" are currently different companies. ⚑ unverified

The default position invites the least-common-denominator problem. Serving a billion casual users pushes the product toward agreeableness, safety-by-blandness, and engagement features (ads on the free tier being the newest). Power users routinely report that getting frontier-quality output requires paying for the top tiers and deliberately selecting the heavy models — the defaults are tuned for the mass market.

Trust history is mixed. OpenAI has repeatedly changed data-retention behavior, pricing, and product commitments with short notice, and its governance drama (2023's board crisis, the ongoing restructuring from capped-profit to for-profit) is a real consideration for anyone signing multi-year dependence on the platform. ⚑ unverified

---

When to choose ChatGPT / OpenAI

Choose it when:

• You want one subscription that does everything. Text, images, video, voice, browsing, custom bots, agents — for a generalist individual user, Plus at $20 is the broadest single purchase in AI. ⚑ unverified
• Ecosystem is the requirement. If your tooling, vendors, and integrations assume OpenAI — and much of the market does — going with the default is the low-friction path.
• You're deploying an assistant to a large non-technical workforce. Familiarity is adoption; nothing beats "it's ChatGPT, you already know it."
• Your workload is multimodal generation. Sora video, native image generation, and real-time voice have no equal bundled elsewhere.
• You want price-competitive frontier API access. GPT-5.6 Sol at $5/$30 is aggressive pricing for a flagship, and Terra/Luna undercut most mid-tier rivals. ⚑ unverified

Look elsewhere when:

• The core workload is agentic coding or long-document professional work — evaluate Claude head-to-head first.
• You need platform stability measured in years; OpenAI's deprecation velocity is the fastest in the industry.
• Your context regularly exceeds ~400K tokens; long-context work favors the million-token vendors. ⚑ unverified

---

How to start

As a user (10 minutes):

1. Free account at chatgpt.com; the free tier is a genuine trial, ads and caps included.
2. Test the breadth, not just chat: generate an image, use voice mode, run a deep-research query. Breadth is what you'd be paying for.
3. If it sticks, Plus at $20/mo is the sensible upgrade; ignore the $100/$200 tiers unless you hit reasoning-model limits weekly. ⚑ unverified

As a developer (30 minutes):

1. API key at platform.openai.com; pip install openai (or the TypeScript SDK).
2. Build on the Responses API, not Chat Completions patterns from old tutorials, and not the deprecated Assistants API. ⚑ unverified
3. Start with GPT-5.6 Terra as the price/quality default; promote hot paths to Sol only where measured quality demands it; push volume work to Luna. ⚑ unverified
4. For agent workflows, use the open-source Agents SDK — and note it can drive non-OpenAI models too, which keeps your architecture portable.
5. Pin model versions explicitly and calendar a quarterly deprecation review. On this platform, that is not paranoia; it is table stakes.

As a team: start with Business at $25/seat (training exclusion and admin controls included), pilot with a real department for a month, and only open the Enterprise conversation — with its 150-seat minimum and negotiated pricing — once usage data justifies it. ⚑ unverified

---

OpenAI changes models, prices, and product availability faster than any vendor in this stack; every figure marked ⚑ unverified should be checked against chatgpt.com/pricing and platform.openai.com immediately before publication.