know.2nth.ai › Agents › GPT · OpenAI
agents · GPT · Skill Leaf

The most-used closed frontier model.

GPT is OpenAI's flagship model family. As of November 2026 the frontier is GPT-6.2 Astra (launched 17 November 2026 for demanding professional work), with GPT-6.2 Sol for everyday use and cheaper GPT-6.1 Terra and GPT-6 Luna tiers on the API. Reasoning is built in — the model decides when to "think" rather than making you pick a separate model. o3 remains as a dedicated reasoning model; the older GPT-4.x line is legacy. ~1M context across the lineup, native multimodal (vision + audio), leading function-calling reliability, hosted via OpenAI Platform and Azure OpenAI Service. The most-used closed frontier model family in the world by deployed count.

GPT-6.2 Astra · current frontier Closed model ~1M context Built-in reasoning Native multimodal

OpenAI's flagship family.

GPT (Generative Pre-trained Transformer) is the model family from OpenAI, the AI lab founded in 2015 and primarily known to most people through ChatGPT. The first GPT shipped in 2018; the family has been on near-annual major release cadence since GPT-2 (2019), with GPT-3 (2020), GPT-3.5 (2022), GPT-4 (2023), GPT-4o (2024), and GPT-5 (2025) marking the major generations. The o-series reasoning line (o1 in late 2024, o3 in early 2025) launched separately as the "thinking" complement, before that capability was folded into the GPT-5.x flagship.

Where Claude has historically led on instruction-following and Gemini on raw context length, GPT has historically led on three things: function-calling reliability (the most consistent structured tool dispatch in the field since 2023), multimodal reach (image, audio, video input across the lineup, plus DALL-E and image generation), and API ecosystem depth (more SDKs, more tool integrations, more community examples than any competing family). Those three properties make it the default choice when ecosystem maturity is the deciding factor.

Distribution: GPT is available via the OpenAI Platform directly, via Azure OpenAI Service (including South Africa North, where only gpt-4o on Regional PTU runs in country — see where LLM inference runs in SA), via ChatGPT (the consumer product), and through every major aggregator (OpenRouter, LiteLLM, the LangChain ecosystem). Microsoft co-developed the platform and is the second-largest deployment surface after OpenAI Platform itself.

Naming · how the family lines up

OpenAI's naming has moved on quickly, and its own pages do not agree on every label. The current API set in OpenAI's models documentation is gpt-6.2-astra (frontier), gpt-6.2-sol (everyday), gpt-6.1-terra and gpt-6-luna (cheaper tiers). OpenAI's 30 July 2026 announcement used GPT-5.6 Sol / Terra / Luna names with different prices, so confirm the live model list before pinning an ID. o3 remains as a dedicated reasoning model. The older GPT-4.x (4o, 4.1) and chat-only GPT-3.5/GPT-4 are legacy or deprecated — check the API model list before pinning anything older.

One flagship line, cost-tiered.

The lineup is now segmented by price tier. GPT-6.2 Astra is the broadest-capability model; GPT-6.2 Sol sits below it as the everyday model; GPT-6.1 Terra and GPT-6 Luna are the cheaper API tiers; o3 remains for dedicated reasoning; GPT-4.x is the cheap legacy floor. Most production stacks route across these price points.

Line · current flagship

GPT-6.2 Astra

The frontier model, with reasoning built in — it allocates "thinking" when the task needs it instead of you switching models. Best for: demanding professional work, agentic workflows, code generation, multimodal, long-context reasoning. gpt-6.2-astra is the top of the line at $8 / $40 per million tokens. ~1M context.

Line · cost-tiered

GPT-6.2 Sol · Terra, Luna

The everyday and volume tiers. gpt-6.2-sol is the everyday model at $2 / $10; gpt-6.1-terra and gpt-6-luna are the cheaper tiers for routing, extraction and high-volume backend work.

Line · reasoning

o3

The dedicated reasoning model — the survivor of the o-series after reasoning was folded into GPT-5.x. It replaced the expensive o1 at a fraction of the cost. Reach for it when you want an explicit reasoning SKU; for most work, GPT-5.x's built-in reasoning is enough.

Capability · cross-line

All lines ship

Function calling, structured outputs (JSON-schema), vision (image input), ~1M context on the flagship, streaming, prompt caching, Batch API. Realtime voice / audio is via the separate Realtime API; image generation is a separate model family.

Reasoning, now built into the flagship

OpenAI's o-series (o1 in late 2024, o3 in early 2025) pioneered explicit "think-before-answering" reasoning as a separate model line. By late 2026 that capability is built into the GPT-6 family — the model decides when to spend compute on reasoning, the same convergence Anthropic made with adaptive thinking and Google with Gemini's thinking models. o3 survives as a dedicated reasoning SKU, but you no longer have to route hard work to a separate model by default. The OpenAI Agents SDK still lets you pin a specific model per agent when you want that control.

Where GPT is the right pick — and where it isn't.

Honest cross-family positioning. GPT's strengths sit on ecosystem reach, function-calling, and multimodal range; its weaknesses are around the "confidently wrong" failure mode and (for some teams) the OpenAI brand-risk profile. Claude leads on instruction-following; Gemini on context length and search grounding; open-weights on cost and residency.

FamilyStrengthsWatch out for
GPT (OpenAI)Largest API ecosystem, best function-calling, broadest multimodal (vision/audio/image-gen), built-in reasoning on GPT-5.x"Confidently wrong" failure mode more than Claude; brand-risk for some enterprise; Realtime / image-gen / transcription billed separately
Claude (Anthropic)Best instruction-following, adaptive thinking, MCP-native, top tier (Fable 5) leads on long-horizon agentic work; Bedrock af-south-1 endpoint with in-country logs (inference itself routes globally via cross-Region inference — not strict residency)Closed; USD billing; tighter rate limits than OpenAI on free tier
Gemini (Google)Largest context window (1M+, Gemini 3.1 Pro reportedly to ~2M), native search grounding, voice via Live API, JHB Vertex regionLess consistent on instruction-following; weaker community / ecosystem than OpenAI
Llama (Meta)Open weights, runs locally, largest community fine-tune ecosystemFrontier gap; instruction-following lags closed frontier; Meta licence has commercial restrictions over 700M MAU
Gemma (Google)Open weights, multimodal, frontier-lab safety tuning, runs locallySmaller fine-tune ecosystem than Llama; not as code-strong as Qwen

Why function-calling matters more than benchmark numbers

Most production agent failures aren't reasoning failures — they're tool-use failures. The model picks the wrong function, calls it twice, returns malformed arguments, or skips a required tool. GPT has had the most reliable function-calling in the field since 2023. Claude has caught up on quality but still has more variance on edge cases; Gemini lags both. For agents where structured tool dispatch is the load-bearing capability, GPT's function-calling consistency is a real production advantage.

Tiered, with three meaningful cost levers.

OpenAI is USD-billed at frontier-tier rates for the flagship; the cheaper tiers are dramatically cheaper. The production cost levers to wire in: tier routing (don't run everything on Astra), prompt caching (cached input billed at ~10% of standard on the GPT-6 family), and Batch API (50% off for non-urgent workloads). Always check openai.com/api/pricing for current numbers — rates below were checked against OpenAI's models documentation in November 2026.

List rates, per million tokens (input / output):

  • GPT-6.2 Astra — $8 / $40. The frontier model with built-in reasoning, launched 17 November 2026. Reserve it for the hard cases.
  • GPT-6.2 Sol — $2 / $10. The everyday model, a quarter of Astra's rate.
  • GPT-6.1 Terra — $1 / $5. The balanced cheaper tier.
  • GPT-6 Luna — $0.10 / $0.50. The high-volume tier for routing and extraction.
  • o3 — ~$2/M input; the dedicated reasoning model, an ~87% cut from the o1 it replaced.
  • Prompt caching — cached input billed at ~10% of standard (e.g. GPT-5.5 input $5 → $0.50 cached). Big saver for RAG / repeated context.
  • Batch API — 50% off input + output for non-urgent workloads.

Note: data-residency (regional-processing) endpoints carry a ~10% uplift for models released on or after 5 March 2026 — relevant if you pin a region for POPIA.

The three-lever routing pattern

For SA studios watching FX, all three levers matter. Build a Luna- or Terra-driven router as the first agent in the chain. Most of the work (60-80%) doesn't need the Astra flagship; route routine traffic to the cheaper tier. Use prompt caching for any RAG or long-system-prompt workload. Push nightly classification / extraction / enrichment to Batch API. This pattern can cut OpenAI costs by 60-80% versus running everything on Astra. The OpenAI Agents SDK's per-agent model selection makes this trivial — sub-agents on different tiers are a one-liner.

When GPT is the right model.

Use GPT when

  • Function-calling reliability is load-bearing
  • You need broad multimodal range (vision + audio + image gen)
  • You're already on Azure or want Microsoft enterprise contracting
  • Reasoning workloads benefit from o-series explicit thinking
  • Ecosystem maturity matters — you want the largest community + integrations
  • You need DALL-E image generation alongside text
  • You're standardised on OpenAI / Azure for enterprise compliance posture

Where GPT lands in SA delivery work.

Enterprise · Azure OpenAI in South Africa North

Azure OpenAI Service in South Africa North (Johannesburg) hosts GPT models with SA data residency, Microsoft enterprise contracting, and POPIA-compliant data handling. For SA banks, insurers, and telcos with existing Microsoft / Azure relationships, this is the structurally clean answer for OpenAI workloads. Caveats: not every model lands in South Africa North at launch — newest GPT releases sometimes lag US East regions by weeks or months. Plan for either a North Europe / UK South fallback or accept the lag if residency is non-negotiable.

Studio · OpenAI Platform direct

For studios without enterprise residency requirements, OpenAI Platform is simpler and cheaper than Azure. Tracing, evals, prompt management on the same Platform dashboard. New models / features land on Platform first, Azure second. The pragmatic SA path: start on Platform for prototypes and pilots; move to Azure SA North if a client requires it.

FX exposure mitigations

OpenAI is USD-billed. Same playbook as the Claude leaf: tier routing (Luna or Terra default, Astra only for hard cases), prompt caching for RAG workloads, Batch API for nightly. For high-volume routing tasks, also consider hybrid local+cloud where Ollama-Gemma 3 handles the 60-80% of work that doesn't structurally need a frontier closed model.

Where GPT links in the tree.

Primary sources.