GPT-5's robustness and reliability have greatly increased compared with previous OpenAI models, which helps agentic workflows.

True: as of 2025-08-14, published benchmarks or evaluations (from OpenAI's GPT-5 system card or independent tests) show GPT-5 performing much better than earlier OpenAI models on reliability measures relevant to agents, such as lower hallucination or error rates, completing tool-use or multi-step tasks, and agentic benchmarks. Mostly true: improvements are clear but modest rather than 'great', or appear on most but not all such measures. Mixed: evidence is about evenly split, with gains on some reliability measures and none or regressions on others. Mostly false: only small or narrow gains that the claim overstates, or gains shown only in OpenAI's own claims and not in its published numbers or independent tests. False: evaluations show no improvement, or worse reliability, than earlier models such as o3 and GPT-4o.

Set before the research began.

1. E1 GPT-5 System Card - OpenAI Deployment Safety Hub OpenAI · 7 August 2025 Own record

   “the GPT-5 models have significantly lower hallucination rates in both "browse-on" and "browse-off" settings. In particular, gpt-5-thinking makes over 5 times fewer factual errors than OpenAI o3 in both browsing settings across the three benchmarks.”

2. E2 Introducing GPT-5 OpenAI · 7 August 2025 Own record

   “With web search enabled on anonymized prompts representative of ChatGPT production traffic, GPT‑5’s responses are ~45% less likely to contain a factual error than GPT‑4o, and when thinking, GPT‑5’s responses are ~80% less likely to contain a factual error than OpenAI o3.”

3. E3 OpenAI says GPT-5 hallucinates less — what does the data say? Mashable · 7 August 2025

   “GPT-5 with thinking was marginally better than o3, while the normal GPT-5 hallucinated one percent higher than o3 and a few percentage points below GPT-4o.”

4. E4 HAL: TAU-bench Tool Calling - Agent Performance Princeton University HAL (Holistic Agent Leaderboard) · 22 September 2026

   “| Taubench Airline | GPT-5 Medium (August 2025) | 48.00% | $23.83 | No |”

OpenAI's own figures show large drops in factual errors (E1, E2). The one independent analysis finds only marginal gains over o3, with standard GPT-5 slightly worse than o3 (E3). The only agentic benchmark entry has no earlier-model baseline and is dated after the judge date (E4), so under the criteria the evidence does not settle how much reliability for agentic tasks improved.