OpenAI 2025-26 lineage

What this page covers

This page is a compact GPT version history for OpenAI’s models across 2025 and 2026, a period in which the company shipped new numbered releases at a pace that was hard to follow in real time. Read as a single timeline, the GPT version history shows a clear pattern: frequent point releases, an increasing split into task-specific and size-specific variants, and a steadily growing emphasis on coding.

From GPT-5 to the Code Red release

The line began with GPT-5 on 7 August 2025, the release that reset OpenAI’s numbering and set the template for what followed. GPT-5.1 arrived in November 2025 as a first refinement. The more consequential step came with GPT-5.2 on 11 December 2025, which shipped in Instant, Thinking and Pro forms alongside a dedicated Codex variant aimed at software work. That release is notable in this GPT version history for its timing as much as its contents: it appeared roughly three weeks after Google’s Gemini 3 Pro, and was reported to have been pushed out amid an internal “Code Red” at OpenAI, a phrase that captures how sharply competitive the frontier had become by the end of 2025. For readers piecing the sequence together, GPT-5.2 is the pivot: the release where the family split into task-specific variants and where competitive pressure visibly began to set the pace.

The coding turn

The coding focus deepened at the start of 2026. GPT-5.3-Codex, released on 5 February 2026, was a Codex-branded model built specifically for the agentic, tool-using style of programming that had become the industry’s most closely watched capability. The “Codex” name signals a model tuned to write, edit and run code inside a developer’s environment rather than simply to answer questions about it, and its arrival as a standalone release shows how much weight OpenAI was placing on that use case. It was also a sign of where revenue was heading, since coding assistants had become one of the most commercially valuable applications of large models, and a dedicated model let OpenAI compete for that work directly.

A laddered family

GPT-5.4 followed on 5 March 2026, initially as Thinking and Pro variants, with smaller mini and nano versions added on 17 March. The mini and nano releases came with a pointed detail: they were listed at four times the price of their GPT-5 equivalents, a reminder that “smaller” does not always mean “cheaper” once a lab reprices its range. That step completed a clear move within the GPT version history towards a laddered menu of models, running from full-size Thinking and Pro tiers down to compact mini and nano options, each pitched at a different balance of cost, speed and capability.

What it adds up to and what to watch

Taken together, these releases tell a coherent story. Over roughly seven months OpenAI moved from a single headline model to a structured family, layered its coding-specific Codex line on top, and iterated fast enough that the numbering alone, from 5 through 5.1, 5.2, 5.3 and 5.4, understates how much changed between versions. The competitive pressure was explicit, with Gemini 3 Pro named in the record as the trigger for at least one accelerated launch, and the pricing decisions on the mini and nano models show a company willing to charge more for newer small models rather than racing every tier to the bottom. Read end to end, the sequence is a study in how quickly a modern model line can branch, from one model in August 2025 to a graded family of Instant, Thinking, Pro, Codex, mini and nano options within seven months.

The cadence has continued: the line now runs to GPT-5.6, sold in Luna, Terra and Sol tiers, and on 30 July 2026 OpenAI cut Luna prices by 80 per cent to $0.20 and $1.20 per million tokens and Terra by 20 per cent to $2 and $12, with Sol unchanged at $5 and $30. For anyone tracking this GPT version history to decide which model to build on, the practical lesson is that OpenAI now ships often and splits finely, so the right choice is usually the most recent variant whose size and price fit the job. For how these releases sit against rival families, see our large language models hub and the broader AI section.