GPT-5
GPT-5 was less a single model than a transition in OpenAI’s product architecture: a routed family combining fast chat models, deeper reasoning models, small-model fallbacks, and later specialized agent/coding variants. Its strongest verified contributions were in reasoning-token-based inference, tool-using agent workflows, long-context execution, and professional/coding benchmarks; its most contested areas were opaque architecture, benchmark interpretation, hidden reasoning, and the economics of high-effort inference.
Coverage note: verified through May 11, 2026.
1. What “GPT-5” meant
OpenAI launched GPT-5 on August 7, 2025, as both a ChatGPT product layer and an API model family. In ChatGPT, GPT-5 was explicitly a Routed Model System: a fast “main” model, a deeper “thinking” model, a real-time router, and smaller fallback models after usage limits. In the API, OpenAI exposed gpt-5, gpt-5-mini, and gpt-5-nano, with gpt-5 representing the reasoning model that powered maximum-performance GPT-5 behavior rather than the non-reasoning chat model used in some ChatGPT paths. OpenAI
That distinction matters. The public brand “GPT-5” bundled several operational ideas that had previously been separated across GPT-4o, o-series reasoning models, and smaller utility models: automatic routing, Reasoning Tokens, test-time scaling, tool-use policies, and cost/latency tiering. OpenAI’s GPT-5 system card described the family as a unification of GPT-4o-style general models and o3/o4-style reasoning models, with model progressions from GPT-4o to gpt-5-main, o3 to gpt-5-thinking, o4-mini to gpt-5-thinking-mini, and o3 Pro to gpt-5-thinking-pro. OpenAI
“Standard” is used in this article as a convenience label for the non-mini, non-Pro flagship line: gpt-5, gpt-5.4, gpt-5.5, and their ChatGPT equivalents. OpenAI’s official naming more often used “GPT-5,” “Thinking,” “Instant,” “Pro,” “Mini,” “Nano,” and “Codex,” rather than a formal “Standard” SKU.
2. Family chronology and variant taxonomy
| Variant / release | Approximate role | Key properties | Evidence status |
|---|---|---|---|
| GPT-5 / Standard, August 2025 | Flagship reasoning/API model and ChatGPT default system | 400k-token API context, reasoning_effort, verbosity, tool support, strong code/agent benchmarks |
Official OpenAI launch and developer docs OpenAI |
| GPT-5 Mini / Nano, August 2025 | Cost/latency optimized models | API sizes for performance-cost-latency tradeoffs; Mini also used as fallback after ChatGPT limits | Official OpenAI system card and API launch OpenAI |
| GPT-5 Thinking | Deliberative reasoning path | Higher test-time reasoning, selectable in ChatGPT paid tiers; API effort controls | Official system card and reasoning docs OpenAI |
| GPT-5 Pro / Thinking-Pro | Maximum-quality path | Parallel or scaled test-time compute; external experts preferred GPT-5 Pro over GPT-5 Thinking in OpenAI’s evals | Official OpenAI announcement and system card OpenAI |
| GPT-5.1 Instant / Thinking, November 2025 | Conversational + reasoning refresh | More conversational Instant model; Thinking more adaptive to task difficulty | Official release note OpenAI |
| GPT-5.2 Instant / Thinking / Pro, December 2025 | Enterprise and agentic-workload refresh | Better long context, tool calling, vision, and agentic workflows | Official release note OpenAI |
| GPT-5.3 Instant, March 2026 | Default everyday ChatGPT refinement | Smoother conversation, better web-search answers, fewer dead ends | Official release note OpenAI |
| GPT-5.3-Codex, February 2026 | Agentic coding branch | Faster long-running coding, higher SWE-Bench Pro / Terminal-Bench results, Codex integration | Official release note OpenAI |
| GPT-5.4 Thinking / Pro, March 2026 | General-purpose agent/professional model | 1M context, native computer use, tool search, professional-work benchmarks | Official release note OpenAI |
| GPT-5.4 Mini / Nano, March 2026 | Small subagent / fallback models | Mini for ChatGPT/API/Codex; Nano API-only; used in subagent patterns | Official release note OpenAI |
| GPT-5.5 Thinking / Pro, April 2026 | Frontier professional/research/coding model | Strongest GPT-5-line benchmark posture; API gpt-5.5 and gpt-5.5-pro; High preparedness classification in cyber and bio/chem |
Official release and system card OpenAI |
| GPT-5.5 Instant, May 2026 | Default ChatGPT chat model | New default model for all ChatGPT users; reduced hallucinated claims versus GPT-5.3 Instant in OpenAI’s high-stakes evals | Official release note OpenAI |
The pattern was not a single monolithic GPT-5 release followed by a pause. It was a release train. By May 2026, GPT-5 had become a product family spanning everyday chat, explicit reasoning, professional work, coding agents, small helper models, and Pro-grade test-time compute.
3. Architecture and training: what is known, and what is not
3.1 Publicly verified architecture: system-level routing, not disclosed model internals
OpenAI has not publicly disclosed GPT-5’s parameter count, dense-vs-sparse structure, exact transformer architecture, training-token mixture, or training compute. The strongest public architectural claim is not “GPT-5 is MoE,” but “ChatGPT’s GPT-5 is a routed system.” The GPT-5 system card says ChatGPT used a smart efficient model, a deeper reasoning model, and a real-time router trained on signals such as model-switching behavior, preference rates, and measured correctness; when users reached limits, a mini version handled remaining queries. OpenAI
This makes GPT-5 an important case study for Model Routing: product intelligence comes from model selection, effort selection, tool access, context persistence, and fallback behavior as much as from a base model. The API surface reinforced that separation: OpenAI said GPT-5 in ChatGPT was a system of reasoning, non-reasoning, and router models, while GPT-5 in the API was the reasoning model used for maximum performance. OpenAI
3.2 Rumored MoE base
The Mixture of Experts claim remains unverified. Rumors and analyst inferences circulated around a sparse/MoE base because the GPT-5 family combined strong benchmark performance with multiple latency/cost tiers and explicit routing behavior, but none of OpenAI’s GPT-5 or GPT-5.5 primary materials confirm a MoE architecture. The safe reading is: GPT-5 clearly used system-level routing; whether any underlying base model was itself a sparse expert model is not established by public evidence.
This distinction is not pedantic. A product router chooses among models or modes at the application layer; a MoE transformer routes token representations among experts inside a forward pass. The first is documented for GPT-5; the second is not.
3.3 Reasoning-token integration
GPT-5 made Reasoning Tokens a mainstream deployment primitive. OpenAI’s reasoning-model documentation explains that GPT-5-class reasoning models may generate internal reasoning tokens before producing visible output; those tokens are not exposed through the API, but they consume context and are billed as output. Developers can control effort with settings such as minimal, low, medium, and high, later extended in GPT-5.5-class docs to include higher effort modes. OpenAI Developers
The practical implication is that GPT-5 quality was partly a function of hidden test-time computation. “Thinking” was not just a prompting style; it was an inference budget. GPT-5 Pro extended this idea further: OpenAI described GPT-5 Pro as using scaled parallel test-time compute, with expert preference results favoring Pro over GPT-5 Thinking in difficult prompts. OpenAI
This created a new operational problem for engineering teams: the relevant unit of cost shifted from “tokens in, tokens out” toward “tokens plus hidden reasoning plus tools plus retries per solved task.” That shift explains why OpenAI’s own developer guidance repeatedly emphasized calibrating reasoning_effort, using the Responses API for reasoning persistence, and defining stopping criteria for agents. OpenAI Developers
3.4 Agentic-tool-use training
GPT-5 was trained and deployed as an Agentic Tool Use model. OpenAI’s developer launch said it trained GPT-5 on real-world coding tasks with early testers, highlighted performance in agentic coding products such as Cursor, Windsurf, GitHub Copilot, and Codex CLI, and introduced custom tools, verbosity control, parallel tool calling, and tool-use preambles. OpenAI
The prompting guide made the deployment thesis explicit: GPT-5 was trained for tool calling, instruction following, and long-context understanding, and OpenAI recommended the Responses API for agentic flows because reasoning can persist between tool calls. This is a different regime from single-shot chat completion: the model is expected to search, inspect, patch, run tests, retry, and summarize state across an execution loop. OpenAI Developers
GPT-5.5 continued that trajectory. Its system card described the model as designed for complex real-world work across code, online research, information analysis, documents, spreadsheets, and multi-tool workflows. It also stated that reasoning models are trained through reinforcement learning to reason through problems, try strategies, recognize mistakes, and produce long internal chains of thought. OpenAI Deployment Safety Hub
3.5 Commonalities and differences across the family
| Axis | Instant / Main | Thinking / Standard API | Pro | Mini / Nano | Codex branch |
|---|---|---|---|---|---|
| Primary optimization | Fast conversational utility | Higher accuracy via reasoning | Maximum quality via more test-time compute | Cost, latency, subtask delegation | Long-running software work |
| Reasoning budget | Lower or automatically routed | Configurable effort | Highest / parallelized | Lower-cost effort tradeoff | Dynamically allocated for coding tasks |
| Tool use | ChatGPT tools, web/search depending product | Built-in and custom tools, function calling, Responses API | Same, with higher compute budget | Useful as subagents or fallback | Terminal, IDE, repo, patch, test, review workflows |
| User-visible reasoning | Summaries or behavior, not raw chain of thought | Summaries / hidden reasoning tokens | Same | Same | Logs, diffs, test outputs, review artifacts |
| Best fit | Everyday chat and lightweight assistance | Research, coding, professional tasks | Hard expert tasks where cost matters less | Routing, high-volume work, narrow subtasks | Software engineering agents |
4. Empirical capabilities
4.1 Reasoning, math, and science
OpenAI reported that initial GPT-5 achieved 94.6% on AIME 2025 without tools, 85.7% on GPQA Diamond without tools in the developer benchmark table, and 24.8% on Humanity’s Last Exam without tools. GPT-5 Pro was reported at 88.4% on GPQA in the launch article’s Pro-oriented results. OpenAI
Independent benchmarker Artificial Analysis found GPT-5 high-effort scored 68 on its Intelligence Index, with medium close behind at 67, and characterized the improvement over o3 as meaningful but not comparable to the jumps from GPT-3 to GPT-4 or GPT-4o to o1. That judgment became central to the “incremental progress” critique: GPT-5 raised the frontier, but did not reset expectations in the way GPT-4 had. Artificial Analysis
By April 2026, GPT-5.5 moved the frontier again. OpenAI reported GPT-5.5 results including 93.6 on GPQA, 41.4 on HLE without tools and 52.2 with tools, FrontierMath tier results, GeneBench/BixBench improvements, and verified Lean proof work. Artificial Analysis later ranked GPT-5.5 xhigh/high at the top of its live Intelligence Index, while noting significant cost and hallucination tradeoffs. OpenAI+2Artificial Analysis+2
4.2 Coding
Initial GPT-5’s headline coding scores were strong: OpenAI reported 74.9% on SWE-bench Verified and 88% on Aider Polyglot, with fewer output tokens and fewer tool calls than o3 in some comparisons. The system-card appendix notes important methodology details: OpenAI used a fixed 477-task validated SWE-bench subset, withheld unit tests, averaged over four tries for pass@1 estimation, and warned that verbosity settings can affect performance. OpenAI
Later GPT-5-family models shifted the coding emphasis from patch generation to long-running agentic work. GPT-5.3-Codex was introduced as OpenAI’s most capable agentic coding model at the time, with claims of faster performance, stronger SWE-Bench Pro and Terminal-Bench results, and availability across Codex app, CLI, IDE, and web surfaces. GPT-5.5 later reported 58.6 on SWE-Bench Pro, 82.7 on Terminal-Bench 2.0, and 73.1 on OpenAI’s internal Expert-SWE benchmark; OpenAI also noted evidence of possible memorization on SWE-Bench Pro, which is a direct warning against treating that benchmark as a clean measure of general coding ability. OpenAI
Reception from software developers was mixed. WIRED reported that some developers found GPT-5 useful for planning and complex one-shot tasks, while others preferred Claude-family models for code quality and criticized GPT-5 verbosity and benchmark framing. The same report quoted Princeton researcher Sayash Kapoor’s tests as finding GPT-5 cheaper but less accurate than Claude Opus 4.1 on a scientific-code-reproduction benchmark, illustrating the gap between benchmark leadership, task-specific reliability, and price-per-solved-outcome. WIRED
4.3 Multimodal and computer use
OpenAI reported initial GPT-5 at 84.2% on MMMU and 78.4% on MMMU-Pro, with visual reasoning gains on CharXiv when Python was enabled. GPT-5.4 then made computer use a more explicit family capability: OpenAI described GPT-5.4 as the first general-purpose model in the line with native computer-use capabilities in Codex and API, with reported results including OSWorld Verified 75%, WebArena Verified 67.3%, and MMMU-Pro 81.2. OpenAI
GPT-5.5 extended the same path: OpenAI reported OSWorld Verified 78.7, MMMU-Pro 81.2 without tools, and 83.2 with tools. These are not merely vision benchmarks; they indicate the model family’s movement toward Computer-Using Agents, where screenshots, GUI state, tool calls, and multi-step action plans become part of the evaluation target. OpenAI
4.4 Agentic and tool-use benchmarks
GPT-5’s agentic benchmark story was unusually central. OpenAI reported 96.7% on τ²-bench telecom, described improved tool instruction following, error handling, and the ability to chain dozens of tool calls in sequence or in parallel. It also reported long-context retrieval improvements, including BrowseComp Long Context performance on 128k–256k inputs. OpenAI
External evaluator METR assessed GPT-5 — specifically gpt-5-thinking in its report — on autonomy-relevant tasks before release. METR estimated a 50% task-completion time horizon of roughly 2 hours 17 minutes, compared with o3’s roughly 1 hour 30 minutes, and concluded that GPT-5 and incremental development from it were unlikely to pose catastrophic risk through AI R&D automation, rogue replication, or sabotage under the evaluated threat models. METR also cautioned that its work did not evaluate all broader risks. METR
GPT-5.5 later reported strong tool-use and professional-agent scores: OpenAI listed BrowseComp 84.4, BrowseComp Pro 90.1, Toolathlon 55.6, Tau2 Telecom 98, MCP Atlas 75.3, GDPval 84.9, FinanceAgent 60, and OfficeQA Pro 54.1. These numbers are important because they show the GPT-5 line being evaluated less as a text generator and more as a work-completion system. OpenAI
4.5 Representative benchmark table
| Domain | Initial GPT-5 result | Later GPT-5-family result | Interpretation |
|---|---|---|---|
| Math reasoning | AIME 2025: 94.6% without tools | GPT-5.5 FrontierMath and HLE improvements reported | Strong frontier reasoning, but effort and tool access matter OpenAI |
| Scientific QA | GPQA Diamond: 85.7% in developer table | GPT-5.5 GPQA: 93.6 reported | Continued improvement across 2026 family OpenAI |
| Coding patches | SWE-bench Verified: 74.9% | GPT-5.5 SWE-Bench Pro: 58.6 | Benchmarks differ; SWE-Bench Pro contamination/memorization caveats matter OpenAI |
| Coding agents | Aider Polyglot: 88% | GPT-5.3-Codex / GPT-5.5 Terminal-Bench gains | Shift from code diff to long-running agent execution OpenAI+2OpenAI+2 |
| Multimodal | MMMU: 84.2% | GPT-5.5 MMMU-Pro with tools: 83.2 | Strong visual reasoning, increasingly tool-mediated OpenAI |
| Agentic tools | τ²-bench telecom: 96.7% | GPT-5.5 Tau2 Telecom: 98; BrowseComp Pro: 90.1 | Tool-use reliability became a core family metric OpenAI |
| Autonomy horizon | METR: 50% horizon about 2h17m for GPT-5-thinking | METR’s time-horizon framework updated through May 2026 | Useful for long-horizon autonomy, not a complete safety measure METR |
5. Cost-quality landscape: GPT-5 vs Claude vs Gemini vs DeepSeek
The cost-quality story changed throughout the GPT-5 era. Initial GPT-5 API pricing was aggressive for a frontier model: $1.25 per million input tokens, $10 per million output tokens, with GPT-5 Mini at $0.25/$2 and Nano at $0.05/$0.40. GPT-5.5 later moved upward: official pricing listed GPT-5.5 at $5 input / $30 output per million tokens for standard short-context use, and GPT-5.5 Pro at $30 input / $180 output; long-context rates were higher. OpenAI
Competitors forced a sharper distinction between “best model,” “best coding model,” “best cheap model,” and “best production outcome per dollar.” Anthropic priced Claude Opus 4.7 at $5 input / $25 output per million tokens, making it close to GPT-5.5 on input and cheaper on output in standard pricing. Google’s Gemini 3.1 Pro Preview pricing was lower for input and output below 200k tokens, at $2 input / $12 output, rising to $4/$18 beyond 200k tokens. DeepSeek V4 Pro and Flash were far cheaper per token, with promotional DeepSeek V4 Pro prices listed at $0.435 input / $0.87 output per million tokens and Flash at $0.14/$0.28, though DeepSeek explicitly framed some prices as discounted and subject to change. Anthropic+2Google AI for Developers+2
Artificial Analysis captured the tradeoff well in 2026: GPT-5.5 led its Intelligence Index at high/xhigh effort, but Gemini 3.1 Pro and Claude Opus 4.7 were highly competitive, with Gemini often cheaper at similar index scores. The same analysis noted GPT-5.5’s strong AA-Omniscience accuracy but high hallucination rate compared with Claude and Gemini, meaning top aggregate intelligence did not imply best factual calibration in every setting. Artificial Analysis
| Provider/model class | Representative price posture | Strengths | Tradeoffs |
|---|---|---|---|
| GPT-5 initial | $1.25 input / $10 output per million tokens | Strong coding, reasoning, agentic tools; good initial cost-performance | Hidden reasoning tokens and tools complicate cost accounting OpenAI |
| GPT-5.5 | $5 input / $30 output; Pro much higher | Frontier aggregate intelligence, professional work, coding, tools | Expensive; high effort can be latency- and token-heavy OpenAI Developers |
| Claude Opus 4.7 | $5 input / $25 output | Strong coding/reasoning competitor; output cheaper than GPT-5.5 | Aggregate benchmark position varies by suite and effort setting Anthropic |
| Gemini 3.1 Pro Preview | $2/$12 under 200k; $4/$18 over 200k | Strong cost-quality ratio, large context | Preview status and model behavior vary by product surface Google AI for Developers |
| DeepSeek V4 Pro / Flash | Very low promotional and list prices | Extremely low token cost; 1M context listed | Lower aggregate frontier ranking in Artificial Analysis; discount volatility and deployment constraints DeepSeek API Docs |
The engineering conclusion is that GPT-5 was rarely the unconditional cheapest option by 2026. It was strongest when its agentic/tool-use reliability, reasoning quality, and OpenAI platform integration outweighed the price difference. For simple extraction, routing, classification, summarization, or subagent work, Mini/Nano, Gemini Flash-style models, or DeepSeek-class models often made more economic sense.
6. Deployment patterns that emerged in 2025–2026
6.1 ChatGPT default routing
GPT-5’s most visible deployment pattern was default routing. OpenAI made GPT-5 the default in ChatGPT and said it replaced GPT-4o, o3, o4-mini, GPT-4.1, and GPT-4.5 for signed-in users, while paid users could explicitly choose GPT-5 Thinking or prompt the system to “think hard.” The system card’s router description explains why users experienced GPT-5 as a product behavior rather than a single model identity. OpenAI
That default-routing strategy created user-experience tension. Users wanted “the best model,” but the product increasingly optimized for a mixture of latency, quota, cost, task difficulty, and perceived helpfulness. The restoration of older models for some paid users after GPT-5 rollout backlash, reported by multiple outlets, reflected a real product lesson: model unification reduces picker complexity, but users still care about model personality, predictability, and continuity. TechRadar
6.2 API tier choices
API users converged on a routing architecture of their own. A typical GPT-5-era stack used a small or instant model for triage and cheap tasks, a standard reasoning model for normal complex work, and Pro/high-effort modes only for hard tasks where marginal quality justified the cost. OpenAI’s GPT-5.5 guide explicitly warned that higher effort is not automatically better and recommended reevaluating low and medium effort before escalating. OpenAI Developers
The API also pushed developers toward stateful execution. The reasoning docs describe preserving reasoning items across function calls, encrypted reasoning tokens for future turns, previous_response_id, and phase handling for long-running tool-heavy workflows. These features made GPT-5-family models more compatible with agent harnesses, but also made integrations more complex than classic stateless chat completions. OpenAI Developers
6.3 Agent harness compatibility
The GPT-5 line fit best in harnesses that could expose tools, preserve state, validate intermediate results, and impose stopping criteria. OpenAI’s own GPT-5 prompting guide advised using the Responses API for tool-calling flows, controlling “agentic eagerness,” lowering reasoning effort when latency mattered, and defining early-stop criteria for context gathering. GPT-5.5 prompt guidance later emphasized outcome-first prompts, retrieval budgets, validation rules, preambles, phase handling, and assistant-item replay for tool-heavy workflows. OpenAI Developers
The deployment pattern that emerged was Model-Orchestrated Software: the model is not merely asked to answer, but to operate inside a scaffold. The scaffold supplies tools, memory, permissions, tests, retrieval, UI state, and cost controls; the model supplies planning, synthesis, and adaptive execution. GPT-5 accelerated that pattern because its benchmark improvements were most convincing when tool use and long-horizon execution were part of the task.
7. Reception and active critiques
7.1 The incremental-progress critique
The central critique was that GPT-5 was impressive but not epochal. Artificial Analysis found GPT-5 set a new standard in its independent benchmark, but explicitly said the increase over o3 was not comparable to the GPT-3-to-GPT-4 or GPT-4o-to-o1 leaps. WIRED’s reporting similarly captured developer disappointment: some users found GPT-5 strong at planning but not obviously superior at code generation, and some saw it as an incremental, subtask-specific improvement rather than a universal jump. Artificial Analysis
This critique is partly about expectations. GPT-5 inherited the symbolic burden of “the next GPT number,” but the actual technical shift was system composition: routing, reasoning effort, tools, long-context, and agent loops. That is real progress, but it is less legible than a single chat model that suddenly answers everything better.
7.2 Reasoning visibility
GPT-5 deepened the tension around Chain-of-Thought Visibility. OpenAI documents internal reasoning tokens, reasoning summaries, and chain-of-thought monitoring, but raw reasoning tokens are not exposed to API users. They consume context and are billed as output, yet users cannot directly inspect them. OpenAI Developers
OpenAI’s safety materials argue that hidden chains of thought can be useful for monitoring deception or misalignment. The GPT-5 system card reported internal monitoring results where a chain-of-thought monitor flagged lower apparent deception rates for GPT-5 thinking than o3 in representative conversations, and external Apollo-style evals found lower deceptive behavior than o3 in specially eliciting trajectories. GPT-5.5’s system card similarly discussed chain-of-thought controllability and monitoring confidence. OpenAI
The unresolved issue is auditability. Hidden reasoning may improve safety monitoring and reduce user confusion, but it also prevents outside developers and researchers from checking whether the model’s visible rationale matches its internal deliberation. For high-stakes domains, that is not a solved problem.
7.3 Cost and latency
GPT-5 made “thinking time” a product knob, but also a cost risk. Reasoning effort can improve hard-task success, but it increases hidden output tokens, latency, and sometimes tool calls. OpenAI’s own developer and prompt guidance repeatedly emphasized matching effort to task difficulty, using lower effort where possible, and defining stop conditions. OpenAI
GPT-5.5 intensified this critique. It led or nearly led many frontier comparisons, but official pricing was materially higher than initial GPT-5 and competitive alternatives. Artificial Analysis described GPT-5.5 as frontier-leading while also highlighting cost comparisons against Claude and Gemini; official OpenAI pricing showed GPT-5.5 Pro at much higher rates than standard GPT-5.5. Artificial Analysis
The practical critique is not simply “GPT-5 is expensive.” It is that GPT-5-class models require outcome-level accounting: retries, hidden reasoning, tool calls, context-window waste, human review time, and failed-agent loops all matter.
7.4 Benchmark validity and benchmark gaming
GPT-5’s benchmark story was strong but not clean. OpenAI disclosed that its SWE-bench Verified number used a fixed internally validated 477-task subset, not all 500 tasks, and that verbosity settings can affect performance. GPT-5.5’s release also noted evidence of memorization on SWE-Bench Pro. These disclosures are valuable, but they weaken simplistic leaderboard comparisons. OpenAI
The broader issue is that GPT-5-family models are increasingly evaluated inside scaffolds: tools, test loops, web search, code execution, MCP servers, browser state, or agent harnesses. A score may reflect the base model, the scaffold, the prompt, the tool budget, and the grading pipeline. For a wiki entry, the honest claim is: GPT-5 was empirically strong across many benchmark families, but benchmark comparability degraded as agentic evaluation became more realistic.
7.5 Safety, dual use, and professional autonomy
GPT-5.5’s system card classified biological/chemical and cybersecurity capabilities as High under OpenAI’s Preparedness Framework, though below Critical. External cyber evaluations found improved vulnerability-research capability and uplift to cyber operators, while also noting limits and missing full critical exploit outcomes in evaluated settings. OpenAI Deployment Safety Hub
METR’s GPT-5 evaluation was cautiously reassuring for the specific threat models it tested, but its scope was limited. It did not claim to settle broader risks such as labor displacement, persuasion, institutional misuse, data leakage, or long-term autonomy escalation. METR
8. Does GPT-5 mark the last major numbered model-family release?
The evidence is unresolved. Before GPT-5, Sam Altman publicly framed the roadmap as a simplification: GPT-4.5 would be the last non-chain-of-thought model, and GPT-5 would unify o-series reasoning models with GPT-series models into a cohesive system. Reuters reported that OpenAI would not release o3 as a standalone product and would integrate it into GPT-5 instead. Reuters
The GPT-5 launch partially fulfilled that roadmap. OpenAI’s system card explicitly described a unified GPT-5 system and said that, in the near future, OpenAI planned to integrate the capabilities of the fast model, reasoning model, and router into a single model. OpenAI
But the naming pattern did not end. OpenAI continued with GPT-5.1, GPT-5.2, GPT-5.3 Instant, GPT-5.3-Codex, GPT-5.4, GPT-5.4 Mini/Nano, GPT-5.5, and GPT-5.5 Instant. As of May 11, 2026, there was no official evidence that GPT-5 would be the last major numbered family. The better interpretation is narrower: GPT-5 marked the end of the clean separation between “GPT chat models” and “o-series reasoning models,” not necessarily the end of numbered GPT releases.
A future GPT-6 could still appear. But GPT-5 made the family number less informative than the deployment mode: Instant vs Thinking, Standard vs Pro, Mini/Nano vs flagship, chat vs Codex, and low-effort vs high-effort inference. In that sense, GPT-5 may be remembered less as “the fifth GPT” than as the point where frontier models became Adaptive AI Systems rather than single callable networks.
References
[OpenAI GPT-5 announcement] — OpenAI’s launch article for GPT-5, including headline benchmark claims, ChatGPT rollout, GPT-5 Pro, and token-efficiency claims. OpenAI
[OpenAI GPT-5 system card] — OpenAI’s system card describing GPT-5 as a unified system, with main, thinking, mini, router, and Pro variants. OpenAI
[OpenAI GPT-5 for developers] — OpenAI’s API/developer launch article for GPT-5, including API model sizes, pricing, tool support, coding results, and agentic-task claims. OpenAI
[OpenAI reasoning model API docs] — OpenAI documentation on reasoning tokens, reasoning effort, hidden internal reasoning, summaries, and tool-call state handling. OpenAI Developers
[OpenAI GPT-5.5 announcement] — OpenAI’s GPT-5.5 release with coding, professional-work, multimodal, tool-use, academic, and cybersecurity benchmark tables. OpenAI
[OpenAI GPT-5.5 system card] — OpenAI’s GPT-5.5 safety and capability card, including training description, chain-of-thought monitoring, cyber/bio preparedness, and external evaluations. OpenAI Deployment Safety Hub
[Artificial Analysis GPT-5 benchmarks] — Independent GPT-5 benchmark comparison across effort levels, including intelligence-index position and the incremental-progress assessment. Artificial Analysis
[Artificial Analysis GPT-5.5 comparison] — Independent GPT-5.5 analysis comparing intelligence, cost, and hallucination behavior against Claude and Gemini. Artificial Analysis
[Artificial Analysis LLM Leaderboard] — Live leaderboard snapshot used for GPT-5.5, Claude, Gemini, and DeepSeek cost-quality comparisons. Artificial Analysis
[METR GPT-5 evaluation] — METR’s pre-release evaluation of GPT-5-thinking on autonomy-relevant threat models and time-horizon tasks. METR
[OpenAI GPT-5 prompting guide] — OpenAI Cookbook guidance on GPT-5 agentic workflows, tool-calling behavior, reasoning effort, verbosity, and prompt migration. OpenAI Developers
[OpenAI GPT-5.5 prompt guidance] — OpenAI API prompt guidance for GPT-5.5 and related GPT-5 family models, including outcome-first prompts, phase handling, and production-agent patterns. OpenAI Developers
[Anthropic Claude Opus 4.7 announcement] — Anthropic source for Claude Opus 4.7 positioning and pricing. Anthropic
[Google Gemini API pricing] — Google source for Gemini 3.1 Pro Preview context and price tiers. Google AI for Developers
[DeepSeek API pricing] — DeepSeek source for V4 Flash/Pro pricing, context, and output limits. DeepSeek API Docs
Companion entries
Core theory: Reasoning Tokens, Test-Time Compute, Chain-of-Thought Visibility, Model Routing, Mixture of Experts, Adaptive AI Systems
Architecture and training: Routed Model System, Agentic Tool Use, Reinforcement Learning for Reasoning Models, Computer-Using Agents, Hidden State in LLM APIs
Practice: Responses API, Cost-Aware Model Routing, Agent Harness Design, Prompt Migration, Codex, Subagent Architectures
Evaluation: SWE-bench, Terminal-Bench, METR Time Horizon, GDPval, Humanity’s Last Exam, BrowseComp, Tau Bench, OSWorld
Critiques and risks: Benchmark Contamination, Reasoning Visibility, AI Cost Curves, Dual-Use Cybersecurity Evaluation, Model Release Reception, Frontier Model Safety Cards