OpenAI: GPT-5

GPT-5 is OpenAI’s most capable general-purpose model to date, positioned as a single system rather than a model you have to choose between. In ChatGPT it replaces the previous menu of GPT-4o, o3 and o4-mini with one entry point; the system decides in real time whether a prompt needs a fast answer or extended reasoning. Via the API it is exposed with a 400,000-token context window and developer controls for reasoning effort and verbosity.

The headline change over the GPT-4 generation is reliability under pressure. OpenAI frames GPT-5 as "PhD-level" across many domains, but the more useful framing for buyers is consistency: fewer confident-but-wrong answers, better instruction-following on long multi-step tasks, and noticeably stronger performance on coding, mathematics, scientific reasoning and health questions. It is the first OpenAI model to make extended reasoning a routed default rather than an opt-in mode.

Under the hood GPT-5 ships as three coordinated parts. gpt-5-main is a fast, efficient model that handles the majority of turns. gpt-5-thinking is a deeper model that spends more test-time compute on hard problems — competition mathematics, multi-step proofs, complex debugging. A lightweight router, trained on real usage signals such as whether users regenerate or switch models, decides which to invoke and how much reasoning to apply.

This matters commercially because it decouples cost from capability. Easy questions are answered cheaply by the main model; only genuinely hard requests pay for extended reasoning. Developers can bias this behaviour through the API with a reasoning-effort parameter, and end users can nudge it in natural language ("think hard about this"). When the routed budget is exhausted, the system falls back to the smaller mini variants to keep latency predictable.

On public evaluations GPT-5 (with reasoning enabled) sits at or near the top of most leaderboards reported at launch. The figures below are OpenAI’s published results; independent reproductions vary by harness and prompt, so treat them as directional rather than absolute. The clearest real-world signal is the SWE-bench Verified score: a measure of resolving genuine GitHub issues end to end, which correlates well with day-to-day coding usefulness.

Mathematics and science scores are unusually high because GPT-5 can spend test-time compute and, where permitted, call tools such as a code interpreter. On AIME 2025 it reaches the high-90s with tools; on graduate-level science (GPQA Diamond) it lands in the high-80s. These are the kinds of tasks where the routed reasoning model, rather than the fast model, is doing the work.

Coding is where GPT-5 is most differentiated for professional users. It is tuned for long agentic runs: reading a repository, planning a change, editing across multiple files, running tests and iterating until they pass. OpenAI highlights front-end generation and "vibe coding" — turning a short description into a working interface — alongside more rigorous tasks like bug triage and refactoring under test coverage.

In practice this makes GPT-5 a strong default for IDE assistants and autonomous coding agents that need to stay coherent over dozens of steps. The trade-off is cost and latency on the deepest tasks: extended reasoning runs consume more output tokens, which is why the routing system and the mini/nano variants exist. Teams typically reserve full GPT-5 reasoning for hard tickets and route routine completions to a cheaper tier.

GPT-5 accepts text and image input and is designed for cross-modal tasks — reading a screenshot of a dashboard, interpreting a diagram, or grounding a coding task in a UI mock-up. The 400,000-token API context window (with large output budgets) lets it hold entire codebases, long contracts or multi-document research bundles in a single request, which is the practical enabler for retrieval-augmented (RAG) and long-document workflows.

For agent builders the large window reduces the amount of external memory plumbing required: more of the working state can live in the prompt, and the model can reason over it directly. As always, effective use of long context depends on good retrieval and ordering — a 400k window is an upper bound, not a guarantee that every token is weighted equally.

OpenAI’s GPT-5 system card documents a shift from hard refusals to "safe completions": instead of declining a request outright, the model aims to give the most helpful answer that stays inside safety boundaries, and to explain the limit when a request is dual-use. The card reports meaningful reductions in hallucination rates relative to o3 and GPT-4o, lower sycophancy (less tendency to simply agree with the user), and reduced deceptive behaviour in evaluations designed to elicit it.

GPT-5 is also treated as a high-capability model under OpenAI’s Preparedness Framework, with additional safeguards in sensitive domains such as biology and chemistry. The honest caveats remain: it still hallucinates, particularly on obscure facts; reasoning depth is bounded by the routed compute budget; and the headline benchmark numbers were produced with reasoning and, in some cases, tools enabled, so a cheap, low-effort call will not reproduce them. Buyers should evaluate on their own tasks before relying on it in high-stakes settings.

GPT-5 is available in ChatGPT (including a routed experience for free users, with higher limits on paid tiers) and via the OpenAI API and aggregators such as OpenRouter. API pricing is roughly $1.25 per million input tokens and $10.00 per million output tokens for the full model, with the gpt-5-mini and gpt-5-nano variants costing materially less for high-volume or latency-sensitive workloads.

Because output tokens dominate cost on reasoning-heavy runs, the economics favour routing: send routine traffic to mini/nano and reserve full GPT-5 for tasks that genuinely benefit from extended thinking. For most production systems this tiered approach keeps quality high on hard requests while holding average cost per request close to that of a mid-range model.