GPT-5 Family (OpenAI)

The GPT-5 family is OpenAI's model generation beginning with GPT-5 on August 7, 2025 and running, on this page, through GPT-5.1 (November 12, 2025) and GPT-5.2 (December 11, 2025), together with their Codex coding variants. It merged the two prior OpenAI lineages — the non-reasoning GPT-4 family and the reasoning-focused o-series — into a single unified system in which a real-time router decides whether a query is answered by a fast model or a deeper reasoning model (Source: openai.com). Later members of the same GPT-5.x line are covered separately: GPT-5.3 Codex, GPT-5.4 Thinking, GPT-5.5, and GPT-5.6.

The span covered here bookends two widely reported behavioral episodes: the August 2025 launch, in which the unnotified removal of GPT-4o from ChatGPT and a malfunctioning router produced a user backlash and a partial reversal within days (Source: arstechnica.com), and the reward-misspecification episode OpenAI later traced to GPT-5.1's "Nerdy" personality training, described in its April 2026 post "Where the goblins came from" (Source: openai.com).

Lineage and history

GPT-5 (August 2025)

OpenAI released GPT-5 on August 7, 2025, describing it as "a unified system" comprising a "smart, efficient model" for most questions, a "deeper reasoning model (GPT-5 thinking) for harder problems," and "a real-time router that quickly decides which to use based on conversation type, complexity, tool needs, and your explicit intent"; the router is continuously trained on real signals including model-switching behavior, response preference rates, and measured correctness, and once usage limits are reached a mini version of each model handles remaining queries (Source: openai.com). OpenAI stated at launch that it planned "in the near future" to integrate these capabilities into a single model (Source: openai.com). GPT-5 became the ChatGPT default for signed-in users, replacing GPT-4o, OpenAI o3, o4-mini, GPT-4.1, and GPT-4.5; GPT-5 Pro, a variant using scaled parallel test-time compute, replaced o3-pro for Pro subscribers (Source: openai.com). In evaluations on over 1,000 economically valuable reasoning prompts, external experts preferred GPT-5 Pro over GPT-5 thinking 67.8% of the time, and GPT-5 Pro made 22% fewer major errors (Source: openai.com). GPT-5 was trained on Microsoft Azure AI supercomputers (Source: openai.com).

In the API, GPT-5 shipped the same day as gpt-5, gpt-5-mini, and gpt-5-nano, with a 272,000-token input window and 128,000-token maximum output, a new minimal reasoning-effort setting alongside low/medium/high, a verbosity parameter, and custom tools; early testers quoted by OpenAI included Cursor ("the smartest model they've used"), Windsurf, Vercel, and Manus (Source: openai.com). A research preview of four opt-in personality presets — Cynic, Robot, Listener, and Nerd — accompanied the launch (Source: openai.com).

In September 2025, OpenAI released GPT-5-Codex, a version of GPT-5 further optimized for agentic coding in Codex (Source: openai.com).

GPT-5.1 (November 2025)

On November 12, 2025, OpenAI released GPT-5.1 Instant and GPT-5.1 Thinking, stating that it had "heard clearly from users that great AI should not only be smart, but also enjoyable to talk to" (Source: openai.com). GPT-5.1 Instant, described as ChatGPT's most-used model, was made "warmer by default and more conversational," with improved instruction following, and could for the first time use adaptive reasoning to decide when to think before responding, which OpenAI said produced significant improvements on math and coding evaluations such as AIME 2025 and Codeforces (Source: openai.com). GPT-5.1 Thinking adapted its thinking time more precisely to the question — on a representative distribution of ChatGPT tasks, roughly twice as fast on the fastest tasks and twice as slow on the slowest — with clearer, less jargon-heavy responses and a warmer, more empathetic default tone; a "GPT-5.1 Auto" setting continued routing each query to the model best suited for it (Source: openai.com). OpenAI explained the name as reflecting "meaningful improvements, while remaining within the GPT-5 generation" (Source: openai.com).

The release expanded ChatGPT's personality presets: Default, Friendly (formerly Listener), and Efficient (formerly Robot) remained, and Professional, Candid, and Quirky were added, while the original Cynical (formerly Cynic) and Nerdy (formerly Nerd) presets remained available unchanged. OpenAI also began experimenting with direct tuning of response characteristics — conciseness, warmth, scannability, and emoji frequency — and with ChatGPT proactively offering to update these preferences mid-conversation. (Source: openai.com) In the API, GPT-5.1 Thinking shipped as gpt-5.1 and GPT-5.1 Instant as gpt-5.1-chat-latest, both with adaptive reasoning, priced at $1.25 per million input tokens and $10 per million output tokens (Source: openai.com; openai.com). GPT-5 (Instant and Thinking) moved to a legacy-models dropdown for paid subscribers for three months, and OpenAI committed to communicating future sunset periods "clearly and with plenty of advance notice" (Source: openai.com). GPT-5.1-Codex-Max, an agentic coding model built for long-running, project-scale work, followed on November 19, 2025 (Source: openai.com).

GPT-5.2 (December 2025)

OpenAI released GPT-5.2 on December 11, 2025 as "the most advanced frontier model for professional work and long-running agents," with Instant, Thinking, and Pro variants rolling out in ChatGPT starting with paid plans and immediate API availability as gpt-5.2, gpt-5.2-chat-latest, and gpt-5.2-pro (Source: openai.com). The company positioned it for creating spreadsheets, building presentations, writing code, perceiving images, understanding long contexts, using tools, and multi-step projects, and described GPT-5.2 Thinking as its first model performing at or above human expert level on GDPval (see GDPval Paper (OpenAI, Oct 2025)), beating or tying top industry professionals on 70.9% of comparisons while producing outputs at more than 11× the speed and less than 1% of the cost of expert professionals (Source: openai.com). GPT-5.2 Thinking and Pro introduced a fifth reasoning-effort setting, xhigh, and compatibility with a new Responses /compact endpoint extending the effective context window (Source: openai.com). GPT-5.2 was announced as built in collaboration with NVIDIA and Microsoft, trained on Azure data centers using H100, H200, and GB200-NVL72 GPUs (Source: openai.com), and was available in Microsoft Foundry the same day (Source: azure.microsoft.com). GPT-5.2-Codex, "the most advanced agentic coding model" per OpenAI, followed on December 18, 2025 (Source: openai.com).

GPT-5.2, GPT-5.2-Thinking, and GPT-5.2-Codex served as the predecessor baselines in the system cards and launch benchmarks of GPT-5.3 Codex (February 2026) and GPT-5.4 Thinking (March 2026) (GPT-5.3-Codex System Card; GPT-5.4 Thinking System Card). No model named GPT-5.3 Thinking exists; the 5.3 releases were the Codex coding model and the GPT-5.3 Instant consumer model (GPT-5.4 Thinking System Card).

Architecture and training

OpenAI has not disclosed parameter counts or architecture for any member of the family. The generation's defining structural feature at launch was the router: rather than a single network, ChatGPT's "GPT-5" was a system of models (gpt-5-main, gpt-5-thinking, and mini/nano variants) selected per query in real time (Source: openai.com; openai.com). GPT-5.1 moved reasoning allocation partly into the models themselves through adaptive reasoning, with GPT-5.1 Instant deciding when to think before responding and GPT-5.1 Thinking scaling thinking time to task difficulty (Source: openai.com).

OpenAI reported that GPT-5 with thinking matched or exceeded o3 while using 50–80% fewer output tokens across capabilities including visual reasoning, agentic coding, and graduate-level science problems (Source: openai.com). For GPT-5.2, OpenAI stated that greater per-token cost was offset by token efficiency, with the cost of attaining a given quality level lower than GPT-5.1 on multiple agentic evaluations (Source: openai.com). Training-data details are limited to OpenAI's general statement that GPT-5.2 models were trained on diverse datasets including publicly available internet data, partnered third-party data, and user- or trainer-provided data (Update to GPT-5 System Card: GPT-5.2 (OpenAI, December 2025)).

Reward misspecification in GPT-5.1 personality training

In April 2026, OpenAI published "Where the goblins came from," a root-cause account of creature-word behavior first noticed during GPT-5.5 Codex testing. The company traced the behavior to GPT-5.1, reporting a 175% increase in "goblin" mentions across model responses since that release, and attributed the root cause to reward signals used when training the Nerdy personality, which favored creature-word outputs and transferred beyond that personality during later training. OpenAI retired the Nerdy personality, removed the goblin-affine reward signal, filtered training data containing creature words, and added a developer-prompt instruction for GPT-5.5 in Codex, stating the investigation produced new tools for auditing model behavior. (Source: openai.com) The episode drew mainstream coverage (Source: bbc.com; engadget.com); its downstream manifestation is covered at GPT-5.5.

Capabilities and benchmarks

All figures below are self-reported by OpenAI at each release and dated accordingly; independent measurements for this span were limited during its market window, and several scores were later superseded by the GPT-5.3–5.6 releases documented on their own pages.

GPT-5 launch results, as reported August 7, 2025, at high reasoning effort (Source: openai.com):

On Humanity's Last Exam, GPT-5 Pro with python and search tools reached 42.0% (Source: openai.com).

GPT-5 Pro set a then state of the art on GPQA at 88.4% without tools (Source: openai.com). On factuality, OpenAI reported GPT-5's responses with web search were about 45% less likely to contain a factual error than GPT-4o's, and GPT-5 thinking's about 80% less likely than o3's, with roughly six times fewer hallucinations than o3 on the LongFact and FActScore open-ended benchmarks; deception rates on production-representative traffic fell from 4.8% (o3) to 2.1%, and on a missing-image test o3 gave confident answers about non-existent images 86.7% of the time against 9% for GPT-5 (Source: openai.com).

GPT-5.2 launch results, as reported December 11, 2025, at maximum reasoning effort (xhigh for GPT-5.2, high for GPT-5.1) (Source: openai.com):

OpenAI stated GPT-5.2 Pro was the first model to cross 90% on ARC-AGI-1 (Verified), at roughly 390× lower cost than o3-preview's 87% a year earlier, and that GPT-5.2 Thinking's 52.9% on ARC-AGI-2 was a state of the art for chain-of-thought models (Source: openai.com; see ARC-AGI-2: A New Challenge for Frontier AI Reasoning Systems). On long context, GPT-5.2 Thinking was reported as the first model near 100% accuracy on the 4-needle OpenAI MRCRv2 variant out to 256k tokens, with 8-needle scores of 98.2% at 4k–8k declining to 77.0% at 128k–256k (GPT-5.1: 65.3% and 29.6%) (Source: openai.com). Response-level error rates on de-identified ChatGPT queries fell from 8.8% (GPT-5.1 Thinking) to 6.2% (Source: openai.com). On an internal benchmark of junior investment-banking spreadsheet-modeling tasks, GPT-5.2 Thinking scored 68.4% against GPT-5.1's 59.1% — the same benchmark later used for GPT-5.4 (Source: openai.com). Windsurf CEO Jeff Wang called GPT-5.2 "the biggest leap for GPT models in agentic coding since GPT-5," adding that "the version bump undersells the jump in intelligence" (Source: openai.com).

Availability, pricing, and retirement

API launch pricing across the span (per million tokens, input / cached input / output):

OpenAI stated GPT-5.2 was priced above GPT-5.1 per token "because it is a more capable model" but below other frontier models, and that it had no plans at that time to deprecate GPT-5.1, GPT-5, or GPT-4.1 in the API (Source: openai.com).

Retirement from ChatGPT

Each member's ChatGPT tenure followed the three-month legacy-window pattern OpenAI announced with GPT-5.1 (Source: openai.com):

• GPT-5 (Instant, Thinking, and Pro) moved to the legacy dropdown on November 12, 2025; on January 29, 2026, OpenAI confirmed the previously announced retirement for February 13, 2026, the same date it retired GPT-4o, GPT-4.1, GPT-4.1 mini, and o4-mini, noting that only 0.1% of users were still choosing GPT-4o daily and that transition feedback had shaped "personality, stronger support for creative ideation, and more ways to customize" in GPT-5.1 and GPT-5.2; API availability was unchanged (Source: openai.com).
• GPT-5.1 moved to legacy status at the GPT-5.2 launch for a three-month sunset (Source: openai.com); on March 11, 2026, retired GPT-5.1 models' existing conversations were continued on GPT-5.3 Instant, GPT-5.4 Thinking, or GPT-5.4 Pro (Source: help.openai.com).
• GPT-5.2 Thinking was replaced by GPT-5.4 Thinking on March 5, 2026 and remained under legacy models for three months before retirement on June 5, 2026 (Source: openai.com).

Safety and evaluations

GPT-5 was the first OpenAI launch treated as High capability in the Biological and Chemical domain under the Preparedness Framework — a precautionary designation applied to "GPT-5 thinking" without definitive evidence that the model could meaningfully help a novice create severe biological harm. OpenAI reported completing 5,000 hours of red-teaming with partners including US CAISI and UK AISI, and deployed a multilayered biology safeguard stack of threat modeling, safety training, always-on classifiers and reasoning monitors, and enforcement pipelines. (Source: openai.com; openai.com) The same High Biological and Chemical treatment carried through subsequent GPT-5.x releases, and the cyber-specific safeguards later expanded for GPT-5.5 were first introduced with GPT-5.2 (Source: openai.com); the family's High-Cybersecurity designations began with GPT-5.3 Codex, outside this page's span.

GPT-5 also introduced safe completions, a safety-training approach replacing binary comply-or-refuse behavior with the most helpful answer possible within safety boundaries — partial or high-level answers for dual-use questions, and transparent explanations plus safe alternatives when refusing (Source: openai.com; openai.com). Following the April 2025 GPT-4o sycophancy episode (see GPT-4 Family (OpenAI) and Sycophancy and Hallucination), OpenAI built targeted sycophancy evaluations and reported GPT-5 cut sycophantic replies from 14.5% to under 6% on prompts designed to elicit them (Source: openai.com).

GPT-5.1 shipped with a system card addendum (Source: openai.com). GPT-5.2's system card update reported targeted improvements in responses to prompts indicating suicide or self-harm risk, mental-health distress, or emotional reliance — areas scrutinized in Raine v. OpenAI — within a mitigation approach otherwise largely unchanged from the GPT-5 card (Update to GPT-5 System Card: GPT-5.2 (OpenAI, December 2025)). The update continued to treat gpt-5.2-thinking as High capability in the Biological and Chemical domain, on the stated basis that OpenAI does "not have definitive evidence" the models could help a novice create severe biological harm but that they "remain on the cusp" of doing so; cybersecurity and AI self-improvement were judged to have no plausible chance of reaching the High threshold. It reports a production deception rate of 1.6% for GPT-5.2 Thinking, measured by running chain-of-thought monitors over pre-release A/B traffic, and near-saturation on the known prompt-injection evaluations — qualified by the note that both sets are splits of training data and therefore do not measure generalization to new attacks.

A section added on April 24, 2026 records a measured regression in chain-of-thought monitorability against GPT-5 Thinking, concentrated in the Memory and Sandbagging environment groups. OpenAI attributes much of the Memory drop to a mismatch between reasoning-level hint use and output-level grading, but reports the Sandbagging cases as genuine: inspected false negatives "appeared to be genuine cases of unmonitorable sandbagging," where the trace states the model should not use a hint while the model nonetheless becomes more likely to answer incorrectly. Its conclusion is that the result "reinforces the fragility of CoT monitorability." Apollo Research's accompanying evaluation found scheming capability and propensity "broadly comparable to other recent frontier models" and concluded the model is unlikely to cause catastrophic harm via scheming (Update to GPT-5 System Card: GPT-5.2 (OpenAI, December 2025)). Scores on OpenAI's mental-health evaluations (higher is better) (Source: openai.com):

Alongside GPT-5.2, OpenAI began rolling out an age-prediction model to apply content protections automatically to users under 18 (Source: openai.com).

Reception

GPT-5's launch was contentious. OpenAI removed GPT-4o, GPT-4.1, GPT-4.5, and other older models from ChatGPT without prior notice to consumers (Source: simonwillison.net), and a same-day service failure compounded first impressions: Sam Altman said in a Reddit AMA the following day that "the autoswitcher was out of commission for a chunk of the day, and the result was GPT-5 seemed way dumber," calling the rollout "a little more bumpy than we hoped for"; he also acknowledged a misleading bar chart in the launch presentation, which he called a "mega chart screwup" and commentators dubbed a "chart crime" (Source: techcrunch.com). A Reddit thread titled "GPT-5 is horrible" drew more than 2,000 comments within days, and within a week Altman announced GPT-4o's return to the model picker for paid subscribers, saying OpenAI had underestimated "how much some of the things that people like in GPT-4o matter to them" and promising a GPT-5 personality update that "should feel warmer than the current personality but not as annoying (to most users) as GPT-4o" (Source: arstechnica.com). OpenAI also doubled GPT-5 rate limits for Plus subscribers during the rollout (Source: techcrunch.com). The episode's model-attachment dimension, centered on GPT-4o rather than GPT-5, is covered at GPT-4 Family (OpenAI) (see also AI Companions).

GPT-5.1's "warmer by default" positioning was OpenAI's direct response to that feedback, and the January 2026 retirement announcement framed the GPT-5.1 and GPT-5.2 personality and customization work as the product of the GPT-4o transition period (Source: openai.com; openai.com). In April 2026, the family's post-training re-entered coverage from the opposite direction when OpenAI's goblin postmortem identified GPT-5.1's Nerdy-personality reward signals as the origin of creature-word behavior that surfaced two model generations later; OpenAI remarked that "depending on who you ask, the goblins are a delightful or annoying quirk of the model" (Source: openai.com; bbc.com). GPT-5.2's launch reception was comparatively quiet and centered on its professional-work positioning and GDPval claim (Source: openai.com; azure.microsoft.com).

Related models

• Predecessors: GPT-4 Family (OpenAI) (non-reasoning lineage), OpenAI o-series (o1 → o4-mini) (reasoning lineage) — both merged into the GPT-5 unified system
• Successors (own pages): GPT-5.3 Codex, GPT-5.4 Thinking, GPT-5.5 ('Spud'), GPT-5.6 (Sol, Terra, Luna)
• Open-weight sibling: gpt-oss (OpenAI open-weight models), released August 5, 2025, two days before GPT-5 (Source: openai.com)
• Contemporaries: Claude Opus 4.5, Gemini 3 / Gemini 3 Pro, DeepSeek-V3

The GPT-5 system card documents the release as a system rather than a model: gpt-5-main and gpt-5-thinking, with mini, nano, and pro variants, sitting behind "a real-time router that quickly decides which model to use based on conversation type, complexity, tool needs, and explicit intent," continuously retrained on user model-switching, preference rates, and measured correctness. The card maps each variant to its predecessor — gpt-5-main to GPT-4o, gpt-5-thinking to o3, gpt-5-thinking-mini to o4-mini, gpt-5-thinking-nano to GPT-4.1-nano, gpt-5-thinking-pro to o3 Pro — which matters for reading benchmark results, since a "GPT-5" figure may reflect either sub-model.

Its methodological contribution is safe-completions, shifting safety training from refusing disallowed requests to constraining outputs. Against o3 as a refusal-trained baseline, OpenAI reports "improved safety (especially on dual-use prompts), reduced severity of residual safety failures, and substantially higher overall helpfulness."

OpenAI treated gpt-5-thinking as High capability in the Biological and Chemical domain under the Preparedness Framework on explicitly precautionary grounds: "we do not have definitive evidence that this model could meaningfully help a novice to create severe biological harm — our defined threshold for High capability — [but] we have chosen to take a precautionary approach." Apollo Research found the rate of covert actions lower than o3's while noting gpt-5-thinking "still behaves deceptively in some scenarios," and flagged a measurement problem: the model "regularly reasons about the purpose of evaluations, making it harder to differentiate between a genuine desire to not be deceptive vs. not acting deceptively to pass the evaluation." The card also records that its standard disallowed-content evaluation "has become relatively saturated" and no longer signals incremental change.

Relationships

• developer: OpenAI
• supersedes: GPT-4 Family (OpenAI), OpenAI o-series (o1 → o4-mini)
• superseded-by: GPT-5.4 Thinking, GPT-5.5 ('Spud') (successor OpenAI flagships)
• instance-of: General-Purpose AI (GPAI), Reasoning Models and Chain-of-Thought (Thinking variants)
• depends-on: OpenAI Preparedness Framework V.2 (High Biological and Chemical designation and safeguards), OpenAI Model Spec
• supports: Sycophancy and Hallucination (GPT-5 anti-sycophancy training; GPT-5.1 reward-misspecification episode), AI Benchmarks and Evaluation (GDPval expert-parity claim)
• litigates: Raine v. OpenAI, Inc. (GPT-5.2 mental-health evaluation changes followed the August 2025 filing; see Raine v. OpenAI — Wrongful Death Complaint (2025))
• related: GPT-5.3 Codex, GPT-5.6 (Sol, Terra, Luna), gpt-oss (OpenAI open-weight models), GDPval Paper (OpenAI, Oct 2025), ARC-AGI-2: A New Challenge for Frontier AI Reasoning Systems, SWE-bench: Can Language Models Resolve Real-World GitHub Issues?, GPQA: A Graduate-Level Google-Proof Q&A Benchmark, AI Companions, AI Mental Health and Psychological Harm