GPT-5 Review 2026: We Tested OpenAI's Most Powerful Model for 7 Months

What is GPT-5?

GPT-5 is OpenAI's flagship large language model, launched on August 7, 2025. It is the successor to GPT-4o and represents OpenAI's biggest architectural shift in years — moving from a single monolithic model to a router-based system that dispatches queries across three specialized sub-models depending on task complexity.

We have been using GPT-5 since launch day, running it through everything from long-form content generation and code debugging to agentic browser automation and multimodal analysis. After nearly eight months of daily use, our take is nuanced: GPT-5 is genuinely more capable than GPT-4o in measurable ways, but the experience is uneven in ways previous models were not.

Sam Altman described GPT-5 as having "PhD-level abilities across tasks," which is the kind of statement that sounds impressive in a keynote and then gets stress-tested by millions of real users. The truth is somewhere in between. For coding, interactive app creation, and structured reasoning, GPT-5 regularly produces output that would genuinely impress a domain expert. For creative writing, casual conversation, and personality — areas where GPT-4o had a surprising amount of charm — GPT-5 feels like it traded warmth for precision.

GPT-5 is available through ChatGPT (free with limits, higher limits on Plus and Pro), the OpenAI API, Microsoft Copilot, and is planned for Apple Intelligence integration. That distribution reach is hard to overstate — no other frontier model ships embedded in this many products out of the box.

Key Features

Here are the six capabilities that define GPT-5 and set it apart from everything that came before it:

How GPT-5 Works: The Router Architecture

The single biggest architectural change in GPT-5 is the router system, and it deserves its own section because it explains both the model's strengths and its most frustrating inconsistencies.

Here is how it works in practice. When you send a prompt to GPT-5, it does not go directly to a single model. Instead, a routing layer evaluates the complexity of your request and directs it to one of three sub-models:

The idea is sound: why burn expensive deep-reasoning compute on "what's the capital of France?" The router saves resources and improves speed for simple tasks while reserving heavyweight processing for queries that need it.

The problem — and this is the biggest complaint we have after eight months — is that the router gets it wrong more often than it should. We have had instances where a moderately complex coding question got routed to gpt-5-main-mini and produced garbage output, while a simple "rewrite this paragraph" request got sent to gpt-5-thinking with a 15-second delay. In the first few weeks after launch, the router malfunctions were bad enough that many users thought the model itself was worse than GPT-4o. OpenAI has improved routing accuracy since then, but it is still not where it needs to be.

The adjustable reasoning effort feature partially addresses this. Through the API, you can override the router and set reasoning to minimal, low, medium, or high. We strongly recommend using this for any production application — it gives you predictable quality instead of leaving it up to the router's judgment.

Pricing Plans

OpenAI made an interesting move with GPT-5 pricing: everyone gets access, but the usage limits create a clear tiering system. Here is the breakdown through ChatGPT:

Our recommendation: The free tier is worth testing to see if GPT-5 fits your workflow. If you use it more than a few times a day, Plus at $20/month is a no-brainer — the higher message limits alone justify it. Pro at $200/month only makes sense if you are doing heavy research or development work where you need unlimited GPT-5 thinking mode and cannot afford to wait in queue. We ran on Plus for the first three months and only upgraded to Pro when we started running agentic workflows that burned through the Plus limits in a single afternoon.

Enterprise pricing is custom and negotiated directly with OpenAI. If you are evaluating GPT-5 for an organization, the Team plan at $25/user/month is the easiest starting point — it includes the guarantee that your data will not be used for model training, which is the first question every legal team asks.

Pros and Cons

After eight months of daily use, here is where GPT-5 genuinely impresses and where it falls short:

GPT-5 vs Claude Opus 4.6 vs Gemini 3 Pro

The frontier model landscape in 2026 is a genuine three-way race. Here is how GPT-5 stacks up against its two closest competitors based on our testing across coding, writing, reasoning, and multimodal tasks:

The honest breakdown: If you primarily code and want the fastest interactive experience, GPT-5 wins. If you need deep reasoning, nuanced writing, or work with massive codebases, Claude Opus 4.6 is the better choice. If you live in the Google ecosystem and need long context or video understanding, Gemini 3 Pro makes the most sense. There is no single "best model" anymore — the right answer depends entirely on your use case.

We personally use all three. GPT-5 for quick prototyping and browser agents, Claude Opus 4.6 for long-form analysis and complex code refactoring, and Gemini 3 Pro for anything involving video or extremely long documents. The models are converging in capability but diverging in personality and strengths, which is actually healthy for the ecosystem.

Final Verdict

GPT-5 is, by measurable benchmarks, the most capable model OpenAI has ever released. The router architecture is genuinely innovative, native multimodal training shows clear improvements over GPT-4o, and the coding and agentic capabilities are the best we have used from any ChatGPT model. For developers who want to prototype fast and ship interactive apps, GPT-5 is hard to beat right now.

But we would be dishonest if we pretended this was the seismic shift that OpenAI's marketing suggested. The router misfires enough to be annoying. The personality regression from GPT-4o is real and noticeable — if you enjoyed GPT-4o as a conversational partner, GPT-5 feels like talking to a more capable but less interesting colleague. The day-one jailbreak was a bad look for a company that positions safety as a core differentiator. And MIT Tech Review's assessment of "a refined product, not a transformative breakthrough" has proven accurate over time.

Here is who should use GPT-5:

• Developers and builders: Yes, absolutely. The coding and app creation capabilities justify the price alone. Start with Plus at $20/month.
• General productivity users: The free tier is worth trying. Upgrade to Plus if you hit the limits more than twice a week.
• Researchers and analysts: Consider it, but compare with Claude Opus 4.6 first. For deep reasoning and long-context analysis, Claude is still ahead.
• Writers and creatives: Proceed with caution. GPT-5's writing is technically proficient but lacks personality. Claude is the better choice for creative work.
• Enterprise teams: The Team plan at $25/user/month is competitive, but evaluate against Anthropic's team plans and Google Workspace AI before committing.

GPT-5 earns a 4.3 out of 5 from us. It is a strong model that does many things well and a few things best-in-class. The router architecture is a smart engineering decision that needs better execution. The free access for all users is a genuinely pro-consumer move. But in a market where Claude Opus 4.6 exists for the same price point and Gemini 3 Pro offers 2M token context, GPT-5 is no longer the obvious default choice it once was. It is one of three excellent options, and which one is "best" depends entirely on what you are building.

Frequently Asked Questions