GPT-5 Released: But Did Not Meet Expectations

On August 7th, OpenAI held its much-anticipated event to unveil GPT-5. After two and a half years since the release of GPT-4, the world was eager to see what the next generation of AI would bring. However, unlike the jaw-dropping debut of ChatGPT, the leap of GPT-4, or the shockwaves sent by the o1 model, this launch felt surprisingly subdued.

A Flat Launch, Modest Benchmarks

The event, lasting an hour and twenty minutes, was marked by underwhelming benchmarks, a lack of paradigm-shifting features, and even some embarrassing errors in the presentation slides. The use cases demonstrated failed to distinguish GPT-5 from its competitors, and the overall atmosphere was far from electrifying.

But does this mean GPT-5 is a disappointment? Not entirely.

Key Highlights: Lower Hallucination, Better Context, and Unbeatable Price

GPT-5 does bring some notable improvements: a dramatically reduced hallucination rate, enhanced frontend capabilities, a significant leap in context window, and—perhaps most importantly—a highly competitive price. Especially in programming, GPT-5’s API is priced at just 1/15th of Anthropic’s newly released Claude Opus 4.1, and even undercuts Google’s Gemini 2.5 Pro. This aggressive pricing could be a game-changer in the AI market.

Four Versions, Smarter Routing

GPT-5 comes in four versions: GPT-5, GPT-5 mini, GPT-5 nano, and the enterprise/pro-only GPT-5 Pro. For most users, the default is a unified GPT-5 model, which intelligently routes queries to either the “main” model for general tasks or the “thinking” model for deeper reasoning. The Pro mode, similar to Grok 4’s “Hard” mode, leverages parallel computation for the most challenging scientific problems, even setting new world records in some benchmarks.

Incremental Improvements, Not a Giant Leap

Across almost every metric, GPT-5 edges out its predecessors and competitors—but only by a small margin. In intelligence benchmarks, it’s only slightly ahead of o3 and Grok 4. In the much-hyped Arc Prize test (touted as the ultimate AGI benchmark), GPT-5 actually lags far behind Grok 4, prompting even Elon Musk to comment.

However, GPT-5 does show improved computational efficiency, achieving better results with fewer tokens—up to 50-80% less in complex problem-solving compared to o3. In user experience, GPT-5 shines, topping the LMArena leaderboard in blind user tests.

Programming: The Real Star

OpenAI put a spotlight on programming improvements. Thanks to the matured Agentic Coding system, GPT-5 can handle complex, multi-step coding tasks, autonomously use tools, and even communicate its plans and findings like a collaborative team member. It excels at understanding requirements, correcting errors, and using external tools—especially in real-world scenarios like the Tau benchmark for dynamic tool use.

Bug-fixing capabilities have also seen a major upgrade. In live demos, GPT-5 navigated real codebases, identified root causes, and even understood the rationale behind human engineering decisions. It can now build, test, and iteratively improve its own code, marking a significant step toward self-improving AI.

Front-End and Real-World Coding

GPT-5’s front-end coding skills were also on display, generating complex visualizations and interactive games in minutes. However, the much-rumored multimodal leap didn’t materialize—GPT-5 still focuses on text and image understanding, with no audio or video generation yet.

Hallucination and Safety: Quiet but Crucial Progress

One of GPT-5’s most impressive achievements is its drastically reduced hallucination rate—down by 45% compared to GPT-4o and 80% compared to o3. This is a huge win for real-world applications, especially in industrial and professional settings where factual errors can be fatal.

OpenAI attributes this to new training techniques, including better use of browsing tools and reinforcement learning that teaches the model to recognize and correct its own mistakes. Deceptive behaviors have also dropped by up to 90% in some dimensions.

Context Window: A New Frontier

All GPT-5 versions now support a 400k token context window, far surpassing previous models and most competitors (except Gemini’s 1M). In “needle-in-a-haystack” tests, GPT-5’s accuracy nearly doubled over o3, making it much more capable in handling long documents and complex tasks.

New Features: Underwhelming but Useful

Other new features include improved writing quality, more natural and emotionally resonant responses, and better understanding of subtle context. Voice capabilities are more natural, and there’s now video input for the voice assistant—though these are becoming standard across the industry.

Memory upgrades were touted, but in practice, this mostly means integration with Gmail and Google Calendar for scheduling—hardly groundbreaking. Customization now allows users to change chat interface colors, a feature that feels more cosmetic than revolutionary.

Data Bottlenecks and Training Innovations

OpenAI acknowledged the data bottleneck in training GPT-5, revealing that they used new techniques to generate high-quality synthetic data with previous models. This recursive improvement loop is promising, but so far, the scaling benefits seem limited.

The Real Disruption: Price Wars

Perhaps the most disruptive aspect of GPT-5 is its pricing. Free users can now access GPT-5 (with generous usage limits), and API pricing is lower than ever—$1.25 per million input tokens and $10 per million output tokens, undercutting even the “budget” Gemini 2.5 Pro. This could be devastating for competitors like Anthropic, whose models are now 15 times more expensive.

It’s ironic that a company once defined by technical leadership is now leading with price. This shift may signal the end of the “magic” era and the start of a more pragmatic, competitive phase in AI.

A Disastrous Event?

The launch event itself was widely criticized. Errors in benchmark charts, lackluster demos, and awkward jokes made for a dull experience. Compared to Anthropic’s vending machine experiment or Gemini’s Pokémon agent showcase, OpenAI’s event felt uninspired.

Public sentiment, as measured by Polymarkt, dropped sharply after the event. The AI industry is now facing the reality that scaling laws are slowing, and the era of exponential leaps may be over—at least for now.

Conclusion: The End of the Beginning

GPT-5 is not AGI. It’s a solid, incremental upgrade with some important improvements, especially in programming, hallucination reduction, and context handling. But the days of magical, paradigm-shifting AI launches may be behind us. The industry now faces a new phase—one defined by fierce competition, pragmatic innovation, and perhaps, the need for a true breakthrough to reignite the dream of AGI.