OpenAI released GPT-5 on August 7, 2025. The company says the new family of models combines quick conversational replies with heavier reasoning in one system. The rollout moves ChatGPT away from a menu of separate models and toward a single, adaptive experience for millions of users.
A single family, four sizes
GPT-5 arrives as a family rather than a single product. The lineup includes the full-capability GPT-5, a higher-capacity GPT-5 Pro for extended reasoning, and two smaller options called GPT-5 Mini and GPT-5 Nano. Mini and Nano trade some depth for speed and lower cost. Pro adds parallel compute during inference so it can handle longer, multi-step problems.
OpenAI is making GPT-5 the default model in ChatGPT for free and paid accounts. Free users will see usage limits and may be routed to Mini when they hit caps. Plus subscribers get higher limits. Pro subscribers at $200 per month receive unlimited or prioritized access to GPT-5 Pro. Team, Enterprise and Edu customers will gain broader Pro access in the days after launch.
Automatic routing, fewer choices
ChatGPT now decides which version of GPT-5 to use for each request. A routing system chooses quick replies or the deeper “thinking” mode when tasks need more processing. The aim is to remove the friction of asking users to pick a model for each task. The chat interface also offers four preset conversational styles, labeled to help users set tone without extra prompts.
On the chat platform OpenAI will phase out older selectable models for most users. Enterprise API customers can still call legacy models for now. That gives developers time to update systems that relied on prior model behavior.
How it performs on tests
OpenAI shared benchmark results showing gains in coding and complex reasoning. On SWE-Bench Verified, which tests real-world coding tasks, GPT-5 scored 74.9% on first attempts. That placed it ahead of some recent competitors on that metric. A GPT-5 Pro configuration with tools reached about 42% on Humanity’s Last Exam, a difficult composite reasoning test. On a PhD-level science benchmark, GPQA Diamond, the Pro setup recorded near-top performance. Health-focused tests also showed much lower rates of fabricated answers when the model used its thinking mode.
Benchmarks are mixed across workloads. In simulated web navigation tasks, GPT-5 posted strong results in some sections and weaker results in others. Those variations suggest the system gains are real but not uniform.
Fewer fabrications and safer responses
OpenAI reports that GPT-5 reduces the rate of fabricated answers compared with recent reasoning models. When thinking mode is enabled, hallucinations decline substantially compared with earlier versions. The company also introduced safer completion behavior so the system explains limits and avoids abrupt refusals. The model is designed to reject clearly malicious requests more reliably while permitting benign or borderline queries to receive useful guidance within safety boundaries.
New developer controls and bigger context
The GPT-5 API brings controls that let developers tune speed and depth. Teams can set reasoning effort levels to choose between rapid responses and deeper analysis. A verbosity option sets brief, standard or detailed reply lengths. Tools can accept free-form strings where appropriate. Developers can also constrain outputs with grammar or regular expressions and enable brief preambles that describe planned tool calls before they run.
OpenAI lists base API prices at $1.25 per million input tokens and $10 per million output tokens for the main tier, with lower unit costs for Mini and Nano tiers. The context window spans 256,000 tokens, allowing much larger documents and longer conversations without aggressive chunking.
What this is not
GPT-5 advances reasoning and task automation, but it does not meet broad definitions of artificial general intelligence. The model does not learn continuously while deployed and it lacks persistent autonomy. Those limitations mean GPT-5 is a significant capability upgrade rather than a shift to fully autonomous intelligence.
Early reactions and migration issues
Early enterprise testers report faster development cycles, more reliable code generation, and cleaner design choices in generated interfaces. Several developer tools and platforms are already running trials. At the same time some ChatGPT users are upset that they can no longer pick older models they had tuned to specific workflows. OpenAI says legacy models remain available via the API for now, which should ease migration for businesses.
An open and commercial approach
The GPT-5 launch comes alongside an open-weight release called gpt-oss intended for local use and customization. That creates a two-track approach: a more capable, centralized family available inside ChatGPT and a free open model for offline experimentation and lower-cost deployments.
Bottom line
GPT-5 packages multiple performance tiers into a single user experience and adds developer tools to tune cost, latency and accuracy. It improves coding and multi-step reasoning and lowers fabrication rates in many settings. The release simplifies the chat surface for most users while keeping options open for developers who need older models or different trade-offs.
Notes: This post was edited/created using GenAI tools.
Read next:
• Study Tests Threat-Based Prompts, Finds Results Unpredictable and Mostly Ineffective
• Where Are The Youngest And Oldest Business Owners In The US? [Infographic]
