OpenAI officially released GPT-5 last week, and the AI research community has been dissecting it ever since. The company claims GPT-5 represents a 'significant step' beyond GPT-4, with benchmark performance that matches or exceeds PhD-level experts in a range of scientific disciplines. Here's what we know about what GPT-5 can do, what it can't, and what it means for the industry.
The Benchmarks Tell a Compelling Story
GPT-5 scores 92.3% on MMLU (a 57-subject academic knowledge benchmark), up from GPT-4's 86.4%. On the MATH dataset — a notoriously difficult set of competition math problems — it jumps from 67% to 88%. On HumanEval (a coding benchmark), it reaches 91%, compared to GPT-4 Turbo's 81%.
These aren't marginal improvements. An 8-point jump on MMLU puts GPT-5 firmly in the range of expert human performance on most subjects. The coding improvement is particularly significant for the large population of developers who rely on AI coding assistance.
What Makes GPT-5 Different
According to OpenAI's technical blog, GPT-5 uses a new training approach that emphasizes 'deliberative reasoning' — essentially, the model is trained to think through problems more carefully before responding, rather than pattern-matching to the most likely-looking answer.
This shows up in practice as significantly better instruction following, more consistent output formatting, and fewer confident wrong answers. Early users report that GPT-5 is notably better at maintaining context and consistency across long, multi-turn conversations.
Access and Pricing
GPT-5 is available to ChatGPT Plus subscribers ($20/month) with standard access limits. Heavier usage and access to the most capable version requires the ChatGPT Pro plan ($200/month). API access is available through OpenAI's platform with per-token pricing. The model is rolling out gradually, so not all users will have immediate access.
Industry Reaction
The response from the AI community has been cautiously positive. Independent researchers who were granted early access confirm the benchmark improvements but also note that hallucinations haven't been eliminated — just reduced. 'It's meaningfully better,' wrote AI researcher Andrej Karpathy on X. 'But it's not qualitatively different in the ways that would indicate a path to AGI.'
Anthropic and Google have not yet responded publicly, but sources suggest both companies are accelerating release timelines for their next major models.
The Takeaway
GPT-5 is a real, meaningful upgrade over GPT-4 — particularly for technical and analytical tasks. The benchmark numbers back up the marketing, and early user reports are broadly positive. However, the core limitations of large language models — hallucinations, limited real-world knowledge, context window constraints — remain. This is iteration, not revolution.