What’s new in GPT-5, comparing with GPT-4
code
| Area | GPT-5 | GPT-4o | GPT-4 |
|---|---|---|---|
| Overall positioning | Current flagship; system with routing between fast and thinking paths | Fast multimodal generalist; UX speed focus | Prior flagship; older generation |
| Reasoning options | Auto-router; direct access to thinking / mini / nano / Pro modes | Single family; no router tiers | Single family; no router |
| Context window | Larger contexts; better long-task retention | Strong; optimised for responsiveness | Good for its time; smaller than newer models |
| Coding | Best-in-class; GPT-5-Codex for refactor, review, agentic steps | Capable but not specialised for agentic coding | Strong legacy; behind newer stacks |
| Multimodal reasoning | Improved text+image reasoning; stronger on image-grounded tasks | Very capable multimodal chat; speed/UX focus | Limited vs 4o/5 |
| Safety & reliability | Reduced hallucinations; tighter controls | Mature guardrails from 4-series lineage | Older safety stack |
| Latency / cost trade-offs | Router picks fast vs thinking; thinking paths are heavier | Low-latency defaults; fewer deep chains | Heavier than minis; simpler lineup |
| Availability | ChatGPT & API; multiple SKUs + router | ChatGPT & API | ChatGPT & API (legacy in many orgs) |
What’s new / better in GPT-5
1) Unified model + real-time router
– Routes simple tasks to a fast path and harder ones to deeper “thinking” paths.
– In ChatGPT you can still force modes when needed.
2) Direct access to “thinking” tiers
– API access to gpt-5-thinking (plus mini/nano); deeper mode exposed in ChatGPT.
3) Coding jump (GPT-5-Codex)
– Multi-file refactors, larger diffs, agentic execution, stricter code review.
4) Higher utility across domains
– More consistent results across coding, maths, writing, health, visual tasks.
5) Packaging & pricing variety
– Router + multiple SKUs lets you trade latency/cost for depth without leaving GPT-5.
Remaining limitations
– Hallucinations not eliminated
Reduced but not gone, especially in scarce/ambiguous domains.
– Latency / cost trade-offs
Thinking paths are heavier; routing helps but physics still applies.
– Tone / creative feel varies
Some prefer 4o for creative copy; subjective.
– Complex lineup
Multiple variants can confuse selection unless standardised internally.
When to choose which
– Choose GPT-5
For correctness-heavy work, deeper reasoning, serious coding/analysis, and router-driven “think when needed”.
– Choose GPT-4o
For fast multimodal chat and snappy drafting where responsiveness matters more than deep reasoning.
– Choose GPT-4
Only if your stack is locked to it; otherwise upgrade.
