What Is GPT-5? OpenAI's Latest Model
Photo: Shantanu Kumar / Pexels
GPT-5 is OpenAI's flagship large language model, released on August 7, 2025 during a livestream event (OpenAI, "Introducing GPT-5"). It was the company's first unified model — a single system meant to fold the fast GPT-4o "instant" line and the slower o-series "thinking" line into one product, with an internal router deciding on the fly how hard to think about each prompt. The rollout did not go smoothly, and the year that followed reshaped how OpenAI ships models at all.
This piece explains what GPT-5 actually is, how the router works, why the launch triggered one of the loudest user revolts in ChatGPT's history, how the model stacks up against Claude, Gemini, and Grok, and what changed as OpenAI pushed out a rapid string of point releases — 5.1 through 5.6 — through mid-2026.
What "unified model" actually means
Before GPT-5, ChatGPT users picked a model from a dropdown: GPT-4o for quick, conversational answers, or one of the o-series models (o1, o3) when they wanted the system to reason step by step before answering. Those were architecturally separate systems, each trained and priced differently, and most users had no idea which one to pick for a given task.
GPT-5 replaced that dropdown with a single entry point. Behind the scenes, OpenAI describes it as "a smart, efficient model that answers most questions, a deeper reasoning model... for harder problems, and a real-time router that quickly decides which model to use based on conversation type, complexity, tool needs, and explicit intent" (OpenAI, "Introducing GPT-5"). Say "think hard about this" and the router routes you to the deep-reasoning path even for a question it would otherwise answer instantly.
Think of it like a hospital triage nurse instead of a phone menu. A phone menu makes you choose your own department before you know what's wrong with you ("press 1 for cardiology, press 2 for orthopedics"). A triage nurse looks at what you're actually presenting with and sends you to the right specialist without asking you to self-diagnose first. GPT-5's router is that nurse: it reads the prompt, estimates how much reasoning it needs, and picks a processing path automatically.
The router itself is a learned system, not a fixed rule table. OpenAI has said it is "continuously trained on real signals, including when users switch models, preference rates for responses, and measured correctness" — meaning the routing logic itself improves the more people use it and correct its decisions.
The GPT-5 lineup and how the router picks a path
GPT-5 shipped in three sizes plus a premium reasoning tier, a structure OpenAI has kept (with renumbering) through every point release since:
- GPT-5 — the flagship, used by default in ChatGPT Plus, Pro, Team, and Enterprise.
- GPT-5 mini — smaller and cheaper, used as an automatic fallback when free-tier rate limits are hit, and available standalone via the API for cost-sensitive workloads.
- GPT-5 nano — the smallest model in the family, built for low-latency, high-volume API tasks like classification or short completions.
- GPT-5 Pro — a higher-effort reasoning tier exposed only on the $200/month ChatGPT Pro plan, which runs much longer internal reasoning chains for research-grade problems.
How does the GPT-5 router decide when to "think"?
The router looks at prompt complexity, whether tools (code execution, browsing, file analysis) are likely needed, and any explicit signal from the user, such as "think carefully" or "just give me a quick answer." A simple rewrite request stays on the fast path; a multi-step coding or proof problem gets routed into extended reasoning, the same behavior that used to require manually picking o1 or o3.
That routing layer is also why GPT-5's benchmark numbers can look inconsistent across independent write-ups. If a benchmark runner disables extended reasoning or caps "thinking tokens," they are effectively testing GPT-5's fast path only, not its full capability — a distinction that mattered a lot in the disputes over Grok 4's claimed benchmark parity later in 2025.
Benchmark results at launch
According to OpenAI's own GPT-5 system card, the model posted strong scores on some of the hardest public evaluations available at the time:
- SWE-bench Verified (real-world GitHub bug-fixing tasks): 74.9%
- Aider Polyglot (multi-language code editing): 88%
- AIME 2025 (advanced competition math, no tools): 94.6%, a new state of the art at launch
- GPQA Diamond (PhD-level science multiple choice), with GPT-5 Pro's extended reasoning: 88.4% without tools
Those numbers put GPT-5 ahead of o3 on most tasks while, notably, costing dramatically less per token to run — a combination that had not happened together in a single OpenAI release before.
Why the launch backlash happened
The technical story behind the August 2025 backlash is more specific than "people didn't like change." According to Altman's own account, a bug in the real-time router meant it was frequently misrouting prompts to the wrong internal model for a stretch of launch day, which made GPT-5 feel noticeably less capable than it actually was. Altman wrote that "GPT-5 will seem smarter starting today" once the fix shipped, because "the autoswitcher broke and was out of commission for a chunk of the day" (Openxcell, citing Altman's public statements).
That was compounded by a separate decision: OpenAI removed GPT-4o and the o-series models from the model picker on the same day, with no opt-out. Users who had built workflows, custom instructions, or simply a communication style around GPT-4o's specific voice lost access instantly, with no warning period. Altman later described this as a mistake in itself, separate from the router bug, writing that he underestimated "how much of an attachment some people have to specific AI models" and that "suddenly deprecating old models that users depended on in their workflows was a mistake" (Platformer, "Three big lessons from the GPT-5 backlash").
OpenAI's response was to reverse course within days: GPT-4o came back as a selectable option for Plus subscribers, and reasoning-mode rate limits were raised. It is a case study now cited across the industry for a specific lesson — users treat a familiar model's personality as a relationship, not just a tool, and swapping it out without consent reads as a loss, regardless of whether the replacement scores higher on benchmarks.
GPT-5 vs the competition
Here is how GPT-5's original launch specs compared to its closest rivals at the time.
GPT-5 launch-era comparison across major frontier models
| Model | Reasoning mode | Code (SWE-bench) | Input $/M tokens | Output $/M tokens | Context window |
|---|---|---|---|---|---|
| GPT-5 | Auto-routed | 74.9% | $1.25 | $10 | 400K |
| GPT-5 mini | Light, on demand | ~60% | $0.25 | $2 | 400K |
| Claude 4 Opus | Extended thinking, manual | ~73% | $15 | $75 | 200K |
| Gemini 2.5 Pro | Thinking mode, manual | ~64% | $1.25 | $10 | 1M+ |
| Grok 4 | Heavy mode, manual | ~73% (self-reported) | $3 | $15 | 256K |
| DeepSeek R1 | Reasoning, always-on | ~57% | $0.55 | $2.19 | 128K |
Table: pricing and benchmark figures reflect model states at their respective 2025 launches; see OpenRouter's GPT-5 listing for current live pricing, which has since shifted as OpenAI moved its default lineup to the 5.x series.
Independent usage in the months after launch surfaced patterns that raw benchmark scores didn't capture:
- GPT-5 vs Claude 4 — Many engineers reported Claude was more consistent on multi-file refactors and long agentic coding sessions, while GPT-5 pulled ahead on competition math and pure logic puzzles. See our Anthropic deep dive for how Claude approaches reasoning differently.
- GPT-5 vs Gemini 2.5 Pro — Gemini's edge was raw context length (1M+ tokens versus GPT-5's 400K), making it the better pick for ingesting very long documents or codebases in one shot, even though it trailed on the hardest coding and reasoning benchmarks. More in our Gemini 2.5 Pro explainer.
- GPT-5 vs Grok 4 — xAI publicly claimed state-of-the-art results for Grok 4; independent replication was mixed and depended heavily on whether external tools were permitted during evaluation, a recurring sticking point in cross-lab benchmark disputes.
- GPT-5 vs DeepSeek R1 — R1 remained far cheaper and open-weight, useful for teams that need to self-host or fine-tune, but it lagged GPT-5 on most frontier benchmarks and lacked GPT-5's native multimodal and tool-calling stack.
Is GPT-5 better than Claude Opus 4?
Neither model is a categorical winner. GPT-5 scores higher on pure math and logic benchmarks like AIME 2025, while Claude Opus 4 has generally been rated more reliable for long, multi-file coding sessions and agentic tool use in independent developer testing. The right choice depends on whether your workload looks more like "solve this problem" or "manage this multi-step project."
What actually got better beyond the benchmarks
Numbers on a leaderboard rarely tell you what changed for someone typing into ChatGPT every day. OpenAI's own release materials and independent testing pointed to several concrete, user-facing shifts:
- Fewer hallucinations. OpenAI reported a meaningfully lower factual-error rate compared to GPT-4o, particularly when the reasoning path was engaged on fact-heavy prompts.
- Better instruction following. GPT-5 follows multi-step system prompts and formatting constraints (word counts, specific structures, tone rules) far more reliably than GPT-4o did, which matters most for anyone building products on top of the API rather than chatting casually.
- Less "AI voice." Writing output showed reduced reliance on stock phrasing and generic structure, with better tone-matching when a user specifies a style.
- Safe completions instead of hard refusals. GPT-4o-era models tended to flatly refuse borderline requests. GPT-5 shifted toward attempting a useful partial answer that stays within policy boundaries — a meaningful behavioral change from a strict "cannot help with that" pattern to a more graded response.
- Expert delegation inside one model. Rather than a user manually invoking a plugin or switching models, the router can call out to specialized subsystems — code execution, web browsing, file analysis — as part of a single conversation.
- Built-in multimodality. Image input and voice were native at launch. Video understanding was partial at first and expanded through the API in later releases.
Why does GPT-5 hallucinate less than GPT-4o?
The primary driver is the reasoning path itself: when the router sends a prompt into extended thinking, the model has more opportunity to check intermediate steps against known facts before committing to a final answer, rather than generating a single pass. This does not eliminate hallucinations, but it reduces them on tasks where verification is possible mid-generation, such as math or code.
Pricing — the reversal nobody expected
GPT-5 launched at $1.25 per million input tokens and $10 per million output tokens for the flagship — cheaper than Claude Opus 4's $15/$75 pricing, and roughly matching Gemini 2.5 Pro. GPT-5 mini, at $0.25/$2, undercut most competing "small" models outright.
This mattered structurally: it was the first time in the frontier-model race that OpenAI's flagship model was priced below Anthropic's flagship, rather than above it. That reversal set off a round of price adjustments across the industry through the rest of 2025, as competitors recalibrated against a cheaper, higher-scoring flagship.
That pricing did not stay fixed, though. As OpenAI iterated into the GPT-5.x line, per-token costs shifted with each release — GPT-5.5, for instance, launched at a notably higher $5.00 per million input tokens and $30.00 per million output tokens (OpenRouter, GPT-5.5 pricing), reflecting its expanded 1M-token context window and heavier agentic capabilities rather than a simple cost-of-compute increase. Anyone building against the API should treat "GPT-5 pricing" as a moving target tied to the specific dated model, not a fixed number.
From GPT-5 to GPT-5.5: what changed in the point releases
OpenAI did not stop at GPT-5. Over the following year it shipped a rapid cadence of point releases, each with a narrower, more specific improvement rather than a full architectural overhaul:
| Release | Approximate date | What it focused on |
|---|---|---|
| GPT-5.1 | November 13, 2025 | Stronger writing quality and data-science-oriented output |
| GPT-5.2 | December 11, 2025 | Iterative quality tuning; retained an August 2025 knowledge cutoff |
| GPT-5.3 | February 2026 | Incremental reasoning and coding improvements |
| GPT-5.4 | March 5, 2026 | Further coding and instruction-following refinement |
| GPT-5.5 | April 23, 2026 | Reframed as an "agent runtime" rather than a chat model — first 1M-token API context window, 82.7% on Terminal-Bench 2.0, 58.6% on SWE-bench Pro |
| GPT-5.6 ("Sol") | June 27, 2026 | Limited preview release, latest in the family as of this writing |
Table: dates and positioning drawn from public release coverage; see Wikipedia's GPT-5.5 entry and the OpenAI GPT-5.5 announcement for primary framing.
The GPT-5.5 shift is the most conceptually significant one in that table. OpenAI began describing it less as a better chatbot and more as a system "built to plan, call tools, inspect its own output, and keep iterating without constant user prompting" — a definition much closer to an autonomous coding agent than a Q&A assistant. On Terminal-Bench 2.0, a benchmark measuring command-line task completion, GPT-5.5 reached 82.7% versus Claude Opus 4.7's 69.4%; on BrowseComp, a web-research benchmark, GPT-5.5 Pro scored 90.1% versus Opus 4.7's 79.3%.
Why does OpenAI keep releasing incremental point versions instead of a "GPT-6"?
Shipping small, frequent updates lets OpenAI respond to competitive pressure from Anthropic and Google in weeks rather than the year-plus cycle a full architecture change requires, and it lets the company course-correct quickly if a release underperforms — exactly the lesson learned from the rocky GPT-5.0 launch. A jump to "GPT-6" is generally reserved for a genuine architectural change, not a capability tune-up.
What people get wrong about GPT-5
"GPT-5 replaced GPT-4o entirely." It didn't stay that way. OpenAI walked back the full deprecation within days of the August 2025 backlash and restored GPT-4o as a selectable model for Plus subscribers. As of the GPT-5.x era, the "unified, single-model" framing OpenAI used at launch has also softened — some coverage describes OpenAI scrapping the automatic router in favor of explicit model settings users can pick themselves, closer to the old dropdown than the original one-model pitch.
"The router always picks the best possible answer." The router optimizes for a tradeoff between latency, cost, and estimated task difficulty, not an unconditional maximum-quality setting. A user who wants guaranteed deep reasoning on every message should say so explicitly or use GPT-5 Pro rather than assume the default path is always the most capable one.
"Higher benchmark scores mean better in practice." GPT-5's AIME and SWE-bench numbers led the field at launch, but independent developer reports consistently noted Claude's edge in long, multi-file agentic coding sessions despite Claude's lower raw SWE-bench score at the time. Benchmarks measure narrow, specific tasks; day-to-day coding work involves context management, tool orchestration, and error recovery that benchmarks capture only partially.
"GPT-5 pricing is fixed at $1.25/$10." That was the launch price for the original GPT-5 flagship. Every subsequent point release in the 5.x line has shipped with its own pricing, and GPT-4o, GPT-5, GPT-5 mini, and GPT-5 nano have since been removed from OpenAI's default pricing page entirely as the 5.x family became the default lineup.
When to use which GPT-5-family model
- General chat, writing, brainstorming — the Plus tier's default flagship model is plenty; you rarely need to force reasoning mode.
- Hard math, complex proofs, multi-step agentic tasks — use GPT-5 Pro, or the API with extended reasoning explicitly enabled.
- High-volume, cost-sensitive API workloads — GPT-5 mini or nano cut per-token costs by roughly 5 to 40 times versus the flagship, with an acceptable quality tradeoff for simpler tasks like classification or short completions.
- Very long documents or large codebases in a single context — Gemini 2.5 Pro's 1M+ token window, or GPT-5.5's newer 1M-token API context, are the better fits; the original GPT-5 flagship's 400K window is comparatively tight.
- Long, autonomous, multi-step coding or research agents — this is where the GPT-5.5 "agent runtime" framing and its Terminal-Bench and BrowseComp scores are most relevant, rather than the original GPT-5.0 numbers.
The takeaway
GPT-5's technology — the router, the reasoning path, the benchmark scores — was genuinely competitive at launch. What made it a case study wasn't the model; it was the rollout. A router bug that misrouted prompts on day one, combined with an abrupt, opt-out-free removal of a model people had built habits around, turned a strong technical release into a public-relations crisis that OpenAI had to walk back within a week. The deeper lesson that stuck through the 5.1-to-5.5 cadence is that OpenAI now treats model transitions as something users need consent and warning for, not just a backend swap, and it has pushed toward more frequent, narrower updates rather than another single, disruptive "unified" relaunch.
Frequently Asked Questions
Q: When did GPT-5 come out? A: OpenAI released GPT-5 on August 7, 2025, initially replacing GPT-4o and the o-series models as the default in ChatGPT, though GPT-4o was restored as a selectable option within days after user backlash.
Q: Is GPT-5 free? A: Free ChatGPT users got rate-limited GPT-5 access at launch (around 5 messages every few hours) before falling back to GPT-5 mini. Plus, at $20/month, gave unlimited access to the flagship model.
Q: Is GPT-5 better than Claude 4 Opus? A: GPT-5 led on math and pure reasoning benchmarks like AIME 2025, while Claude was widely reported as more consistent for long, multi-file coding and agentic tasks. They're best described as peers with different strengths rather than one being strictly better.
Q: How much does GPT-5 cost via the API? A: The original GPT-5 flagship launched at $1.25 per million input tokens and $10 per million output tokens, with GPT-5 mini at $0.25/$2. Later releases in the 5.x line, like GPT-5.5, launched at different price points ($5.00/$30.00), so pricing should be checked per specific dated model.
Q: What is GPT-5 Pro? A: A higher-effort reasoning tier available on the $200/month ChatGPT Pro plan. It runs longer internal reasoning chains than the standard flagship and targets research-grade problems where more "thinking time" improves accuracy.
Q: Did GPT-4o get discontinued? A: OpenAI removed GPT-4o from the default ChatGPT picker at GPT-5's launch, then restored it as a selectable option for Plus subscribers within about a week after significant user backlash on Reddit and elsewhere.
Q: Why did people dislike GPT-5 when it launched? A: Two separate issues combined: a bug in the real-time router misrouted prompts to weaker internal models for part of launch day, and OpenAI simultaneously removed GPT-4o and the o-series models with no opt-out, which upset users who had built workflows around those specific models' behavior.
Q: What is GPT-5.5 and how is it different from GPT-5? A: GPT-5.5, released April 23, 2026, is positioned as an "agent runtime" rather than a chat model, built to plan multi-step tasks and call tools with less user prompting. It introduced OpenAI's first 1M-token API context window and posted 82.7% on Terminal-Bench 2.0, ahead of Claude Opus 4.7's 69.4%.
Q: Does GPT-5 still exist, or has it been fully replaced? A: As of mid-2026, OpenAI's default lineup has shifted to the GPT-5.x family (5.4, 5.5, and the limited-preview 5.6), and the original dated GPT-5, GPT-5 mini, and GPT-5 nano models were removed from OpenAI's default pricing page, though they may remain accessible for legacy API integrations.
Q: What is the GPT-5 router and why did it break at launch? A: The router is a learned system that reads each prompt and decides in real time whether to send it to the fast-answer model or the deeper reasoning model, based on complexity, tool needs, and explicit user signals. At launch, a bug caused it to misroute prompts for part of the first day, which made the model perform below its actual capability until OpenAI fixed it.
Sources
- OpenAI, "Introducing GPT-5" (August 2025)
- OpenAI, "Introducing GPT-5.5" (April 2026)
- OpenAI GPT-5 System Card (arXiv:2601.03267)
- OpenRouter, GPT-5 pricing and benchmarks
- OpenRouter, GPT-5.5 pricing and benchmarks
- Openxcell, "Sam Altman on GPT-5 Backlash"
- Platformer, "Three big lessons from the GPT-5 backlash"
- Yahoo/TechCrunch, coverage of GPT-4o restoration
- Wikipedia, "GPT-5.5"
Read next:
Follow new articles
Add SciHub101 to your feed reader to see new science and technology stories.
About the Author
SciHub101 is an independent science and technology publication founded and edited by Buğra Sözeri, a software engineer and app developer based in Türkiye. AI tools assist with drafting and research. Source documentation in the existing archive is uneven, and articles are being reviewed and corrected. The Editorial Team byline identifies editorial responsibility; it does not certify specialist or independent human review. Where we compare products, anything built by our own team is labelled as such — see the editorial guidelines and About page for how corrections and disclosures work.
Related Articles
What Is Replit Agent? Build Apps with AI
Replit Agent explained: build and deploy full-stack apps from a single prompt, how Agent 3 works, pricing, and how it compares to Bolt, v0, and Lovable.
What Is Bolt.new? StackBlitz AI Code Tool
Bolt.new explained: StackBlitz's AI app builder that runs Node.js in the browser via WebContainers, its $40M ARR run, and how it stacks up against Replit.
How Does ChatGPT Actually Work? A Plain-Language Guide to LLMs
ChatGPT predicts the next word based on patterns it learned from billions of pages of text. Sounds simple — but it produces essays, code, and reasoning that look uncannily human. Here's what's really happening.
How to Write AI Prompts That Actually Work
The quality of an AI model's output is determined largely by the quality of the input it receives. A prompt is not a search query — it is the only communication channel between the user and the model. Every word specifies or leaves ambiguous something about the desired output. This guide covers the principles behind effective prompting, the techniques with the most research evidence behind them, and the common mistakes that produce frustrating, generic results.
Why AI Models Give Different Answers to the Same Question
Send ChatGPT the same message twice within five minutes and you will likely receive two distinct responses. The wording differs, the examples change, sometimes the conclusions shift. This is not a bug, and it is not the model lying — it is the intended consequence of probabilistic sampling, a deliberate design choice that trades consistency for coverage. Understanding why requires a look at how language models generate text one token at a time.
How Large Language Models Actually Reason — Inside ChatGPT, Claude, and Gemini (2026)
Do LLMs actually reason or just predict text? A deep technical guide to transformer attention, chain-of-thought, test-time compute, and frontier model benchmarks in 2026.