Table of Contents
- Family Evolution Timeline
- Two Product Lines: Closed API for Revenue, Open GPT-OSS for Ecosystem
- Architecture: Why One System Can Handle Many Tasks
- GPT-5.6: Sol, Terra, Luna—How to Choose
- Sub-lines & Ecosystem: A Table of All GPT Products
- Position Against Competitors
- What This Means for Agent Developers
- Overall
- References
🌏 中文版
On November 30, 2022, ChatGPT went live—1 million users in five days, 100 million in two months, the fastest-growing consumer product in history. Four years later, in July 2026, the GPT-5.6 Sol/Terra/Luna three-tier lineup launched, covering every price band from $0.20 to $30, with OpenAI at 1B+ users and 2M+ enterprise customers. This is the third family deep-dive in the "AI 模型家族" series, tracing GPT's full evolution from GPT-1 to GPT-5.6.
For how to read the benchmark numbers cited here, see our AI model evaluation sources guide. This is part of the AI Model Landscape Overview series.
Family Evolution Timeline
| Version | Date | Key Facts |
|---|---|---|
| GPT-1 | 2018-06 | 117M params, first generative pre-trained model |
| GPT-2 | 2019-02 | 1.5B params, delayed release as "too dangerous" (later fully open-sourced) |
| GPT-3 | 2020-05 | 175B params, few-shot learning, semi-open API (application required) |
| GPT-3.5 | 2022-03 | ~175B, RLHF fine-tuned, ChatGPT's base |
| ChatGPT | 2022-11 | ~175B, consumer AI revolution, 100M users in 2 months |
| GPT-4 | 2023-03 | ~1T (MoE), multimodal, 90th percentile on bar exam |
| GPT-4 Turbo | 2023-11 | ~1T, 128K context, faster and cheaper |
| GPT-4o | 2024-05 | Native multimodal (text + vision + speech) |
| o1 | 2024-09 | First reasoning model, "think before speaking" |
| o3 / o4-mini | 2025-04 | Reasoning model upgrades |
| GPT-4.1 | 2025-04 | API-only, 1M context |
| GPT-5 | 2025-08 | Unified routing: fast model + reasoning model + real-time router |
| GPT-5.5 | 2026-04 | Efficiency & agent capability boost, 1M context |
| GPT-OSS | 2025-08 | OpenAI's first open weights: 120B / 20B, Apache 2.0 |
| GPT-5.6 Sol/Terra/Luna | 2026-07 | Three-tier product line, 1.05M context |
Eight years, multiple milestones. GPT's through-line is clear: from research paper to consumer product to platform ecosystem. GPT-3 proved scaling laws, ChatGPT created consumer demand, GPT-4 set the quality bar, and everything after GPT-5 focused on agent capability and efficiency. 2025 added a side branch—weights released for the first time (GPT-OSS).
Two Product Lines: Closed API for Revenue, Open GPT-OSS for Ecosystem
To understand GPT's 2026 layout, split it into two lines—the opposite direction from Qwen (open-first, recently closing): GPT started closed and only recently added open:
Closed API Line (OpenAI official API / ChatGPT / Azure OpenAI Service): All flagships after GPT-5 are API-only, no downloadable weights. In 2026 this line split from a single GPT-5 into three SKUs—Sol (flagship) / Terra (balanced) / Luna (efficient)—covering every price band. This line owns revenue—1B+ users and 2M+ enterprise customers all run here.
Open-Weight Line (GPT-OSS): Only released in August 2025, 120B and 20B, Apache 2.0. This line owns ecosystem—giving developers who need self-host, fine-tuning, and data sovereignty a path, but explicitly not the frontier model.
The turning point: OpenAI's 2019 capped-profit conversion, IPO rumors in 2026, ever-increasing commercial pressure. GPT-OSS looks more like a strategic response (to Llama, Qwen, DeepSeek open offensives) than a return to roots—the closed-flagship main line has never loosened.
Architecture: Why One System Can Handle Many Tasks
Unified Routing Architecture: One System, Not One Model
GPT-5 (August 2025) marked OpenAI's shift from "single model" to "unified system":
GPT-5 = fast model + reasoning model + real-time router
The router decides which model to use based on conversation type, complexity, tool needs, and explicit user intent (e.g., "think carefully"). Conceptually similar to Anthropic's adaptive thinking, but GPT-5's implementation is more explicit—two independent models behind the scenes, while Claude adjusts reasoning depth inside one model.
Developers no longer pick a model; the system decides. The cost: routing is opaque, sometimes the wrong layer is chosen, and cost controllability suffers.
GPT-5.6 Inference Efficiency: The Model Discounts Itself
Two price cuts after GPT-5.6 launched in July 2026, driven by efficiency gains: GPT-5.6 Sol optimized its own inference kernels (end-to-end serving cost -20%); speculative decoding efficiency +15%. The model discounts itself.
- 2026/07/30: Luna -80% ($1/$6 → $0.20/$1.20), Terra -20%
- 2026/08/21: Sol -20%+ (promo through 11/21)
This is OpenAI's 2026 core play: trade efficiency for price bands, pushing frontier quality down into more tiers.
Native Multimodality and GPT-OSS Architecture
GPT-4o (2024-05) established native multimodal—text, vision, speech handled inside one model, not via adapter plugins. That capability carries into the GPT-5 main line.
GPT-OSS is the first time OpenAI handed over weights. 120B and 20B both support function calling and structured outputs, deployable on self-owned infra, under clean Apache 2.0. Quality lags the closed flagship significantly, but it proves OpenAI now seriously farms the open ecosystem niche.
GPT-5.6: Sol, Terra, Luna—How to Choose
July 2026's three-tier naming serves distinct roles:
| Item | GPT-5.6 Sol | GPT-5.6 Terra | GPT-5.6 Luna |
|---|---|---|---|
| Positioning | Flagship reasoning, complex tasks | Daily balanced, cost-effective | High throughput, low cost |
| Input ($/MTok) | $5 | $2 | $0.20 |
| Output ($/MTok) | $30 | $12 | $1.20 |
| Long context input | $10 | $4 | $0.40 |
| Long context output | $45 | $18 | $1.80 |
| Cache hit | $0.50 | $0.20 | $0.02 |
| Batch API | $2.50 / $15 | $1 / $6 | $0.10 / $0.60 |
| Fast mode | $10 / $60 | $4 / $24 | $0.40 / $2.40 |
| Context | 1.05M | 1.05M | 1.05M |
| Max output | 128K | 128K | 128K |
| Reasoning effort | none → max | none → max | none → max |
| License | Closed API | Closed API | Closed API |
License Trap: Flagship Is Closed, Open Is Not Frontier
OpenAI's licensing is the most stratified in the series—a hard line between open and closed:
- Closed flagship (all GPT-5.6): API only, no weights, license = OpenAI Terms of Service. Strongest models are not self-hostable
- Open weights (GPT-OSS 120B / 20B): Apache 2.0, downloadable, fine-tunable, self-hostable—but only mid-large (120B) and small-mid (20B), not frontier
- Historical: GPT-2 fully open, GPT-3 semi-open (application), GPT-4 onward fully closed
The catch: "OpenAI open-source" ≠ "strongest model open-sourced." GPT-OSS is genuine Apache 2.0—cleaner than Llama 4's Community License—but 120B's active params and quality lag GPT-5.6 Sol significantly. If you need "Apache 2.0 + frontier quality," GPT-OSS can't deliver; look at DeepSeek (MIT) or Qwen (mostly Apache).
Another trap is API price volatility: OpenAI adjusted GPT pricing multiple times across 2025–2026 (mostly cuts, but long-context surcharges too); prices during a contract term aren't guaranteed long-term. Easily missed: beyond 272K input tokens, input doubles and output +50%. Long-document workloads (codebases, contracts, meeting transcripts) cost ~2× the short-context sticker price.
Performance Position
| Metric | GPT-5.6 Sol | Comparison |
|---|---|---|
| Artificial Analysis Coding Agent Index | 80 | Terra 77.4; Claude Fable 5 77.2; GPT-5.5 76.4; Claude Opus 4.8 72.5. Sol uses < half the output tokens of Fable 5, half the time, ~1/3 the cost |
| LiveBench Overall | 81.1% | Claude Fable 5 83.0%; GPT-5.5 80.2%; Claude Opus 5 80.1%; Terra 77.9% |
| Terminal-Bench 2.1 | 88.8% | Sol Ultra 91.9%; Claude Mythos 5 88%; GPT-5.5 85.6%; Fable 5 83.1% |
| BrowseComp | 90.4% | Sol Ultra 92.2%; Claude Mythos 5 88%; Terra 87.5%; Fable 5 84.3% |
| Agents' Last Exam | 53.6 | Claude Fable 5 40.5; Fable 5 (medium) 42.2—Sol at medium reasoning already beats Fable 5 at ~1/4 cost |
Direct comparison:
| Metric | GPT-5.6 Sol | Claude Fable 5 | DeepSeek V4 Pro |
|---|---|---|---|
| LiveBench overall | 81.1% | 83.0% | 77.4% |
| Coding Agent Index | 80 | 77.2 | — |
| BrowseComp | 90.4% | 84.3% | — |
| Terminal-Bench 2.1 | 88.8% | 83.1% | 82.7% |
| SWE-bench Verified | ~95% | 95% | 96.4% |
| Output pricing ($/MTok) | $30 | $50 | $0.87 |
| Context | 1.05M | 1M | 1M |
GPT-5.6 Sol clearly leads on agentic tasks (BrowseComp, Terminal-Bench, Agents' Last Exam), but DeepSeek V4 Pro matches or slightly beats SWE-bench Verified at ~1/34 the price. LiveBench's general ranking is slightly won by Claude Fable 5.
Sub-lines & Ecosystem: A Table of All GPT Products
In the closed-source camp GPT ranks second in family breadth (only behind Qwen):
| Sub-line | Representative | Positioning |
|---|---|---|
| Flagship reasoning | GPT-5.6 / Sol / Pro | Unified routing, highest intelligence |
| Lightweight | GPT-5.6 mini / nano | High speed, low price, embedded |
| Open weights | GPT-OSS 120B / 20B | Apache 2.0, self-hostable |
| Reasoning specialist | o-series (o4 / o5) | Deep thinking, test-time compute |
| Image | DALL·E 3 / 4 | Text-to-image |
| Speech | Whisper (open) / GPT-4o-realtime | Transcription / voice chat |
| Embeddings | text-embedding-3 | RAG backbone |
| Dev tools | Codex / Assistants API / Function Calling | Agent infrastructure |
| Consumer product | ChatGPT | World's largest AI application |
Two observations:
Unified routing is GPT-5's core bet. Unlike Claude's four explicit tiers (Fable/Opus/Sonnet/Haiku) for developers to pick, GPT-5 uses one router to auto-allocate compute per task difficulty. Pro: no manual model selection. Con: cost controllability drops—you don't know which layer served each request.
Breadth second only to Qwen. Codex (coding agent), Whisper (open speech), DALL·E (image), embedding line—each standalone. Highest ecosystem completeness in the closed camp. Final reminder: model suffixes (Sol/Terra/Luna/mini/nano) are service tiers, not guarantees of downloadable weights—check the license field for self-host, not the suffix.
Position Against Competitors
Placing GPT in the 2026 landscape:
- vs Claude: GPT-5.6 leads on agentic tasks (BrowseComp, Terminal-Bench, Agents' Last Exam); Claude is more stable on coding reliability and the overall LiveBench (83.0%). The two closed titans
- vs DeepSeek V4: DeepSeek matches or slightly beats SWE-bench Verified at ~1/34 price, MIT self-hostable. GPT's edge is ecosystem maturity (Codex, Assistants API) and global availability
- vs Gemini 3.1: Gemini leads on scientific reasoning and native multimodality; GPT is stronger on agentic tasks; general capability is split
- vs Open source (Llama 4 / Qwen / Kimi / Mistral): GPT quality leads, but price is 20–50× open source. GPT-OSS offers an open path but not frontier quality
What This Means for Agent Developers
- Strongest agentic performance → GPT-5.6 Sol (max): Coding Agent Index 80, BrowseComp 90.4%, Agents' Last Exam 53.6, multi-category leader
- Daily coding balanced for quality & cost → GPT-5.6 Terra, $12 output, Coding Agent Index 77.4, sufficient for most tasks
- High-throughput classification / summarization → GPT-5.6 Luna, $1.20 output, 1.05M context, for bulk lightweight requests
- Chasing lowest SWE-bench score → DeepSeek V4 Pro, $0.87 output, SWE-bench Verified 96.4%, ~1/34 Sol's cost
- Self-host + Apache 2.0 → GPT-OSS 120B/20B self-hostable, but not frontier; frontier self-host look at Qwen or DeepSeek
- Citing benchmarks → GPT's naming matrix (gen × Sol/Terra/Luna × mini/nano × closed/open) is the most error-prone in the series. Always write the full model name plus date, otherwise you're comparing different events
Overall
GPT's story is "from research paper to global infrastructure." OpenAI proved scaling laws with GPT-3, created consumer demand with ChatGPT, set the quality bar with GPT-4, and redefined model usage with GPT-5's unified routing. By GPT-5.6, the three-tier lineup (Sol/Terra/Luna) covers every band from $0.20 to $30, with 1.05M context and 128K output spanning most scenarios.
For agent developers, GPT-5.6's lead on agentic tasks (BrowseComp, Terminal-Bench, Agents' Last Exam) is real. But pricing is the highest tier—Sol's $30/MTok output is 1.2× Claude's and ~34× DeepSeek's. The pragmatic move: Sol for the hardest agent tasks, Terra for daily coding, Luna for high-throughput work.
The licensing stratification is worth watching: flagship fully closed, GPT-OSS genuinely Apache 2.0 but not frontier. On GPT, "open-source" is currently real at only one tier, and that tier is not the strongest model.
References
- GPT-5.6: Frontier intelligence that scales with your ambition — OpenAI
- Advancing the price-performance frontier with GPT-5.6 — OpenAI
- Introducing GPT-5 — OpenAI
- OpenAI API Pricing
- OpenAI — Wikipedia
- GPT-5 History Timeline & OpenAI Evolution
- The Complete History of OpenAI
- OpenAI Pricing in 2026 — What You Actually Pay Per Token — StackWrite
- LiveBench Leaderboard
- Artificial Analysis Coding Agent Index
- AI Model Evaluation Sources Guide — this site
- AI Model Landscape Overview — this site
Loading...