ModelRefs / GPT-5 - AI model implementation reference

GPT-5 - AI model implementation reference

GPT-5 is OpenAI's August 7, 2025 hosted reasoning model for coding, reasoning, and agentic tasks. The API model accepts text and image input and produces text, supports configurable reasoning effort and verbosity, and has a 400,000-token context window. It is now a previous-generation model, so implementation decisions should use the immutable gpt-5-2025-08-07 snapshot and compare it with current successors before new adoption.

Overview

This page helps you evaluate GPT-5, OpenAI's August 7, 2025 hosted reasoning model for coding, reasoning, agentic tool-use, and long-document analysis. OpenAI now classifies it as a previous model, so ModelRefs presents it as an implementation and migration reference rather than the provider's current flagship.

Use this page to review the immutable gpt-5-2025-08-07 snapshot, 400,000-token context window, text and image-input surface, reasoning-effort and verbosity controls, seven source-scoped benchmark records, current lifecycle status, and the OpenAI sources behind each implementation claim.

Seven provider-run benchmark records are admitted for scoped decision support. Fit signals remain provisional until tested on your workload. OpenAI does not disclose their physical execution dates or GPT-5's complete training corpus, parameters, compute, or serving architecture; ModelRefs preserves those fields as undisclosed and does not substitute publication dates or aggregate score tables.

Benchmark & Evaluation

ModelRefs currently has partial, narrow benchmark coverage for GPT-5. Treat the available benchmark evidence as one input to the decision, not a guarantee that GPT-5 is the strongest option for your workload, and evaluate it on representative workloads before selecting it.

  • Seven canonical provider-run records cover SWE-bench Verified, Aider Polyglot, GPQA Diamond, AIME 2025, tau2-bench telecom, BrowseComp Long Context, and LongFact Concepts for GPT-5 at high reasoning where the source specifies it.
  • These results are source-scoped and provider-reported rather than independent reproduction. OpenAI publishes methodology qualifications and a publication date, but not the physical execution dates; ModelRefs does not infer them.

Implementation considerations

  • Evaluate reasoning-effort settings explicitly: higher effort changes latency and cost, and task quality does not scale uniformly across workloads.
  • Pin a dated model snapshot for production and re-run task evaluations when the default snapshot advances.
  • Available as gpt-5 and the immutable gpt-5-2025-08-07 snapshot through the Responses API and Chat Completions API; current model documentation labels GPT-5 as a previous model.
  • The documented API price is $1.25 per million input tokens, $0.125 cached input, and $10 per million output tokens; rate limits and retention controls remain account- and endpoint-specific.

Architecture disclosure

  • OpenAI distinguishes the GPT-5 API reasoning model from ChatGPT's routed GPT-5 system and from gpt-5-chat-latest; benchmark claims must not be transferred between those surfaces.
  • The documented API surface supports text input/output and image input, with audio and video unsupported for this model. OpenAI does not publish parameter count, layer design, tokenizer details, training compute, or serving topology.

Prompts and code examples

Version-pinned Responses API request

import OpenAI from "openai";

const client = new OpenAI();
const response = await client.responses.create({
  model: "gpt-5-2025-08-07",
  input: "Review this deployment plan and return the three highest-risk assumptions.",
  reasoning: { effort: "medium" },
  text: { verbosity: "low" },
});

console.log(response.output_text);

Provider implementation source

Training disclosure

  • OpenAI publishes a September 30, 2024 knowledge cutoff but not a complete training-data inventory, parameter count, compute ledger, or reproducible training recipe.
  • OpenAI states that API customer data is not used to train models unless the customer explicitly opts in; this policy does not describe the original GPT-5 training corpus.
  • The current GPT-5 model page marks fine-tuning as unsupported for this snapshot.

History and source-reviewed changelog

  1. 2025-08-07 — GPT-5 released in the API

    OpenAI released GPT-5, GPT-5 Mini, and GPT-5 Nano and documented reasoning effort, verbosity, custom tools, pricing, and provider-run evaluations.

    Milestone source
  2. 2025-08-07 — Immutable GPT-5 snapshot published

    OpenAI documents gpt-5-2025-08-07 as the snapshot that locks behavior and performance for the GPT-5 API model.

    Milestone source
  3. 2026-08-09 — Lifecycle status reviewed

    OpenAI's current model documentation describes GPT-5 as a previous model and recommends a newer GPT-5 generation for new work; ModelRefs retains GPT-5 for migration and implementation-reference decisions.

    Milestone source

Risks and limitations

  • Hosted-model behavior, quotas, pricing, and data controls can change without a client-side version pin unless a dated snapshot is used.
  • Provider-reported capabilities require task-specific evaluation before production reliance.

Source coverage

ModelRefs verified the GPT-5 launch, model documentation, system card, implementation guidance, API example, data controls, and seven source-scoped provider evaluations. OpenAI does not publish the evaluation execution dates, so ModelRefs preserves them as undisclosed rather than substituting the August 7, 2025 publication date.

Connected ModelRefs evidence

Sources

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to GPT-5 - AI model implementation reference.