GPT-5

GPT-5 is OpenAI's multimodal large language model released on August 7, 2025, marking the start of the GPT-5 generation of large language models. It processes text and images, supports agentic tool use, and introduced a real-time routing architecture that selects between a fast inference path and a deeper reasoning path based on query complexity. As of April 2026, GPT-5 has been superseded within the same generation by GPT-5.4 and then GPT-5.5, though it remains available as a legacy API option and is historically significant as the model that defined the GPT-5 generation at launch.

At a glance

• Lab: OpenAI
• Released: August 7, 2025
• Modality: Text and multimodal (vision)
• Open weights: No (closed)
• Context window: 400,000 tokens
• Pricing at launch: $1.25 per million input tokens, $10 per million output tokens (standard API); $0.125 per million cached input tokens
• Distribution channels: OpenAI API (https://platform.openai.com), ChatGPT (web and mobile), Microsoft Azure OpenAI Service, Microsoft Copilot

Origins

The GPT series begins with GPT-1 in 2018, OpenAI's first demonstration of large-scale pretraining on text corpora followed by task-specific fine-tuning. GPT-2 (2019) and GPT-3 (2020) established the pattern of scaling transformer models on internet text, with GPT-3's 175 billion parameters setting the field's benchmark at the time. GPT-4 (March 2023) added multimodal vision and substantially extended performance on professional-domain tasks. GPT-4o (May 2024) restructured the architecture for end-to-end multimodal processing, treating vision and language as native rather than additive.

GPT-5 arrived fourteen months after GPT-4o and represented a more fundamental architectural departure than any prior point-to-point upgrade in the series. Rather than presenting a single model, OpenAI built GPT-5 as a compound system: a fast high-throughput model, a deeper reasoning model, and a real-time router that routes each query to the appropriate path based on conversation type, complexity, tool requirements, and explicit user intent. This architecture unified what had previously been two separate product lines. OpenAI had maintained the GPT series for general use and the o-series (o1 in September 2024, o3 in early 2025) for extended chain-of-thought reasoning tasks. GPT-5 collapsed that distinction by baking the reasoning capability directly into the product surface.

OpenAI CEO Sam Altman described GPT-5 at launch as offering "PhD-level" abilities across diverse tasks and called it "the best model in the world." The rollout was not clean. On launch day, the routing system malfunctioned, causing the model to perform noticeably worse than expected as the autoswitcher failed to route queries to the reasoning path. Altman publicly acknowledged the issue the following day, committed to router fixes, and restored GPT-4o access for Plus subscribers who wanted to return to the previous model while the issues were addressed. A subsequent update in mid-August adjusted GPT-5's personality, which users had described as less agreeable than GPT-4o, reducing sycophancy concerns that had followed OpenAI's prior model updates.

GPT-5 was natively multimodal, trained from scratch on multiple modalities simultaneously rather than adding vision as a post-hoc capability. Training used a combination of unsupervised pretraining, supervised fine-tuning, and reinforcement learning from human feedback.

Capabilities

GPT-5's core text generation capabilities covered question answering, summarization, instruction following, code generation, and extended multi-turn dialogue. The model showed particular strength on coding tasks: SWE-bench Verified performance of 74.9% at launch placed it narrowly ahead of Claude Opus 4.1's 74.5%, making it competitive with the best available models on real-world repository bug-fixing at the time of release.

Vision was native. GPT-5 processed images alongside text in the same context window, supporting chart interpretation, document understanding, and visual question answering. The 400,000-token context window was substantially larger than GPT-4o's, enabling analysis of longer documents and code repositories.

The internal routing architecture made extended reasoning available without requiring users to switch between model families. For straightforward queries, the fast model path returned low-latency responses. For complex problems in mathematics, logic, or coding, the reasoning path activated automatically. Users could also invoke the thinking mode explicitly. The GPT-5 Pro variant, available to Pro-tier ChatGPT subscribers, applied parallel test-time compute for the most demanding tasks.

Agentic capabilities were first-class from launch. GPT-5 supported function calling, structured outputs, and code execution through the API. The model could complete multi-step tasks such as navigating calendar systems, generating software applications, and producing research briefs by coordinating tool use across steps.

OpenAI did not publish a technical report with parameter counts or detailed training data information for GPT-5. The architecture beyond the routing system description remained undisclosed.

Benchmarks and standing

These benchmarks reflect GPT-5's position at launch in August 2025. The field has advanced since; GPT-5.5 and subsequent models now hold these positions or surpass them.

On GPQA Diamond, the graduate-level scientific reasoning benchmark, GPT-5 scored 89.4% at launch, compared to Claude Opus 4.1 at 80.9%. This was among the highest GPQA Diamond scores recorded at the time and reflected GPT-5's strength on knowledge-intensive tasks requiring domain depth.

On SWE-bench Verified, GPT-5 scored 74.9%, placing it roughly at parity with Claude Opus 4.1 (74.5%) and ahead of other available models. This metric gained weight through 2025 as enterprise software engineering workflows became a primary evaluative axis for frontier models.

GPT-5 substantially reduced hallucination rates relative to its predecessor. On healthcare accuracy tasks, GPT-5 hallucinated 1.6% of the time versus 12.9% for GPT-4o. On general question tasks, the error rate was 4.8%, compared to 22% for o3 on the same evaluation, though the comparison across different model types requires caution in interpretation.

The Artificial Analysis Intelligence Index placed GPT-5 at 45 at launch. By the time of GPT-5.5's release in April 2026, GPT-5.5 held a score of 60.24 on the same index, which reflects both GPT-5.5's improvements and evolution in the index methodology over that period.

On LMArena head-to-head human preference evaluations, GPT-5 entered the leaderboard near the top of its cohort at launch but was overtaken as competitors released updates through late 2025 and early 2026.

Access and pricing

GPT-5 launched simultaneously through the API and the ChatGPT consumer product on August 7, 2025.

The OpenAI API provided programmatic access at $1.25 per million input tokens and $10 per million output tokens, with a 90% cache discount bringing cached input to $0.125 per million tokens. Three API-accessible sizes were available at launch: gpt-5, gpt-5-mini, and gpt-5-nano. The pricing was notably lower than Anthropic's Claude Opus 4.1, which launched at $15 per million input tokens and $75 per million output tokens, and prompted coverage suggesting the pricing could pressure competitors.

On ChatGPT, GPT-5 became the default model for all users at launch, including free-tier users, replacing GPT-4o as the standard experience. Plus subscribers at $20 per month received higher usage limits. Pro subscribers at $200 per month received unlimited GPT-5 access and the GPT-5 Pro variant with extended reasoning via parallel test-time compute. Following the bumpy initial rollout, OpenAI restored GPT-4o access for Plus subscribers as an alternative.

Microsoft Copilot and Microsoft Azure OpenAI Service provided additional distribution channels, with Azure serving as the primary path for enterprise deployments requiring data-residency or compliance configurations.

As of April 2026, GPT-5 remains available via the API as a legacy option, though OpenAI directs new users toward GPT-5.5.

Comparison

At launch in August 2025, GPT-5's direct competitors were:

• Claude Opus 4.1 (Anthropic). The immediate peer at launch. Benchmark performance between GPT-5 and Claude Opus 4.1 was close across most axes: GPT-5 led on GPQA Diamond (89.4% to 80.9%) and had the edge on hallucination reduction; Claude Opus 4.1 matched GPT-5 on SWE-bench Verified (74.5% to 74.9%). Anthropic's pricing at launch was substantially higher: $15 per million input tokens versus GPT-5's $1.25, which made the two models difficult to compare on pure cost terms and drove significant press coverage around OpenAI's aggressive pricing strategy.
• o3 (OpenAI). OpenAI's own reasoning-specialist model, released earlier in 2025. GPT-5's integrated routing architecture was partly designed to reduce the need for users to choose between GPT-series and o-series models. At launch, o3 remained available and held advantages on certain hard reasoning tasks, but GPT-5's combined architecture covered the majority of o3's use cases while also handling general tasks more fluidly.
• Gemini 2.5 Pro (Google DeepMind). A comparable-generation model from Google. Gemini 2.5 Pro's pricing was similar to GPT-5 at standard usage volumes but increased at high query volumes, which made GPT-5 more cost-competitive for API-heavy deployments. Gemini 2.5 Pro had the strategic advantage of integration with Google Search and Workspace, a distribution position GPT-5 could not match directly.
• GPT-5.5 (OpenAI, April 2026). GPT-5's successor within the same generation. GPT-5.5 improved substantially on the Intelligence Index (60.24 vs GPT-5's 45 at launch) and raised SWE-bench Verified, GPQA Diamond, and AIME 2025 scores to new highs. Pricing moved upward from GPT-5's launch rates: $5 per million input tokens for standard GPT-5.5, reflecting both the expanded capability and the market's demonstrated willingness to pay for frontier performance.

Outlook

GPT-5 is a closed generation. Its successors GPT-5.4 (March 2026) and GPT-5.5 (April 2026) have taken over its position, and OpenAI's stated trajectory points toward GPT-6 as the next generation boundary, expected in late 2026 or early 2027 if the quarterly cadence of the GPT-5 generation continues.

For users currently on GPT-5 via the legacy API, the practical question is migration timing. GPT-5.5 offers meaningfully higher benchmark performance across most categories. The pricing difference between GPT-5's $1.25 input tokens and GPT-5.5's $5 input tokens is real, and use cases where cost-per-token is the binding constraint may stay on GPT-5 for longer, particularly for high-volume workloads where the capability delta does not justify the cost increase.

The router architecture GPT-5 introduced has become a template. The approach of routing between fast and reasoning paths within a single model surface rather than requiring users to choose between separate model families is now reflected in GPT-5.5 and is likely to persist into GPT-6.

Whether OpenAI maintains GPT-5 API access through the GPT-6 generation or deprecates it when GPT-6 ships has not been publicly announced. GPT-4o remained available for more than a year after GPT-5 launched, suggesting OpenAI maintains legacy endpoints with reasonable longevity.

Sources

• TechCrunch: OpenAI's GPT-5 is here. Launch coverage with pricing, benchmark scores, and capability summary, August 7, 2025.
• TechCrunch: OpenAI priced GPT-5 so low, it may spark a price war. Pricing analysis and competitive comparison against Anthropic and Google, August 8, 2025.
• TechCrunch: Sam Altman addresses 'bumpy' GPT-5 rollout. Router malfunction details and corrective measures, August 8, 2025.
• Wikipedia: GPT-5. Model architecture, variants, training methodology, and successor relationships.
• Artificial Analysis: GPT-5 model profile. Intelligence Index score (45), context window, pricing, and performance data; notes recommendation of GPT-5.1 as successor.
• OpenAI: GPT-5.5 launch announcement. Generation context: GPT-5.5 Intelligence Index at 60.24 vs GPT-5's 45 at launch, April 2026.
• The Frontier Lab Exodus. Background on the competitive landscape and OpenAI's position entering the GPT-5 generation.