On August 7, 2025, OpenAI launched GPT-5, which it called its best AI system yet, and made it the default model in ChatGPT for signed-in users. GPT-5 replaced a confusing menu of models (GPT-4o, o3, o4-mini, GPT-4.1 and GPT-4.5) with one system that decides for itself when to answer quickly and when to reason at length. For developers, the API release brought three model sizes, aggressive pricing and new controls over reasoning depth, output length and tool formats, which together changed how most OpenAI-based applications are configured.

Key Facts#

  • Release date: August 7, 2025. API models gpt-5, gpt-5-mini and gpt-5-nano, with snapshots dated 2025-08-07.
  • Context: 400,000 tokens in total, split into up to 272,000 input tokens and 128,000 output tokens, according to Microsoft's Azure OpenAI documentation.
  • List pricing per million tokens: gpt-5 at $1.25 input and $10 output, gpt-5-mini at $0.25 and $2, and gpt-5-nano at $0.05 and $0.40.
  • Vendor-reported results: 74.9% on SWE-bench Verified, 88% on Aider Polyglot, 94.6% on AIME 2025 without tools, and 84.2% on MMMU.
  • Reliability claims: OpenAI says GPT-5's thinking responses are about 80% less likely to contain a factual error than o3's, and that deception in real ChatGPT traffic fell from 4.8% of o3 responses to 2.1%.
  • New API controls: a verbosity parameter, a minimal reasoning effort setting, free-form custom tools that accept raw text, and context-free grammars that constrain tool output.
  • ChatGPT rollout: Free, Plus, Pro and Team users on day one, with Enterprise and Edu a week later. GPT-5 pro replaced o3-pro for the most demanding tasks.

What Happened#

In ChatGPT, GPT-5 is a system rather than a single model. A fast model handles most questions, a deeper reasoning model ("GPT-5 thinking") takes the hard ones, and a real-time router picks between them based on the conversation, the task's complexity, the tools needed and explicit cues such as asking the model to think hard. When users hit their usage limits, smaller mini versions take over. OpenAI said it planned to fold these pieces into a single model in the future.

OpenAI called GPT-5 its strongest coding model to date, citing gains in complex front-end generation and in debugging larger repositories. It also emphasized efficiency: in OpenAI's evaluations, GPT-5 with thinking beat o3 while using 50% to 80% fewer output tokens across visual reasoning, agentic coding and graduate-level science problems. Because output tokens drive both cost and latency for reasoning models, that claim mattered as much as the headline benchmark scores.

The launch put unusual weight on reliability. Beyond the factual-error reductions, OpenAI described a test in which it removed the images from a visual reasoning benchmark and asked the models to answer anyway. o3 still gave confident answers about the missing images 86.7% of the time, compared with 9% for GPT-5. OpenAI also said GPT-5 was trained on Microsoft Azure AI supercomputers.

For developers, OpenAI released a set of new controls alongside the models. The verbosity parameter lets you ask for terse, balanced or expansive answers without rewriting prompts. The minimal reasoning effort runs GPT-5 with few or no reasoning tokens for fast, simple tasks. Custom tools let the model send raw text, such as a SQL query or a script, to your code instead of wrapping it in JSON, and context-free grammars can restrict that text to a precise syntax. OpenAI recommended the Responses API for the best results, while keeping Chat Completions supported.

Background#

By mid-2025, OpenAI's lineup had become hard to navigate. GPT-4o was the general default, GPT-4.1 served fast API workloads, and the reasoning models o3 and o4-mini handled harder problems. When it launched o3, OpenAI had already said it intended to merge the reasoning skills of the o-series with the conversational and tool-use strengths of its GPT models. GPT-5 delivered on that promise.

Competition was intense. Anthropic released Claude Opus 4.1 two days earlier, on August 5, 2025, xAI had launched Grok 4 in July, and Google had made Gemini 2.5 Pro generally available in June. GPT-5's pricing, with the flagship at $1.25 per million input tokens and $10 per million output tokens, matched Gemini 2.5 Pro's list price and sat well below Claude Opus 4 and Grok 4, a signal that OpenAI intended to compete on cost as well as capability.

Why It Matters for Developers#

GPT-5 simplified model selection. Instead of choosing between a fast model and a reasoning model, you pick a size and a reasoning effort. That is a better fit for configuration-driven design: a single deployment can serve classification at minimal effort and code review at high effort. Remember that the default effort for GPT-5 is medium, which adds latency you may not want on interactive paths, and that the usable input budget is 272,000 tokens, not the 400,000-token total.

In .NET, Microsoft.Extensions.AI covers the portable part of this through ChatOptions.Reasoning, which maps to OpenAI's reasoning effort. GPT-5-specific settings such as verbosity and the minimal effort level are not part of the cross-provider abstraction, so set them through ChatOptions.RawRepresentationFactory, which lets you supply the provider's own options object, or call the official OpenAI .NET library directly for those requests.

Custom tools and grammars deserve attention from teams that generate code or queries. Constraining a model to a grammar for your SQL dialect or a configuration format is a stronger guarantee than a prompt that asks nicely, and it pairs well with the techniques in our guides on structured outputs and function calling. Still, validate and parameterize anything that reaches a database.

Finally, lower hallucination rates are welcome but not a license to skip grounding. OpenAI's own figures show errors are reduced, not eliminated. For facts that matter, retrieval over your own data remains the safer design, as described in our RAG guide.

What's Next#

OpenAI iterated quickly on the GPT-5 family. GPT-5.1 followed on November 12, 2025, with a new default reasoning effort of none in the API, and GPT-5.2 arrived on December 11, 2025. The numbering continued into 2026 with GPT-5.4, GPT-5.5 and GPT-5.6, and OpenAI moved to a GPT-6 generation in September 2026.

The original gpt-5 models remain listed in OpenAI's SDK and in Azure's model catalog as of September 2026. If you still run them, the main migration risk is the change in default reasoning effort in later versions: code that relied on GPT-5's medium default will behave differently on newer models unless you set the effort explicitly. Pin your settings and test before switching.

Sources#