Without external aids, GPT-5 won't beat Meta's Cicero at Diplomacy or drive a car reliably.

The claim has two parts, both about GPT-5 without external aids (tools or specialised scaffolding): (1) it will not beat Meta's Cicero at Diplomacy; (2) it will not drive a car reliably. True: there is no credible demonstration of GPT-5 alone doing either. Mostly true: neither has been done, apart from borderline demonstrations that use limited aids. Mixed: one part is credibly contradicted and the other holds. Mostly false: one part is clearly contradicted and the other is near-contradicted. False: both are credibly demonstrated by GPT-5 alone.

Set before the research began.

1. E1 DrivingBench: Can Vision-Language Models Drive a Toyota Corolla? arXiv (preprint; authors Aditya Ramabadran, Simon Mahns, Tobias Gessler, et al.) · 1 October 2026

   “and GPT-5.6 Sol achieved 6% in all three attempts.”

2. E2 DrivingBench: Can Vision-Language Models Drive a Toyota Corolla? arXiv (preprint; authors Aditya Ramabadran, Simon Mahns, Tobias Gessler, et al.) · 1 October 2026

   “Astra is the only model to finish the course, on its second attempt, with no other attempt passing 50% of the course.”

3. E3 DrivingBench: Can Vision-Language Models Drive a Toyota Corolla? arXiv (preprint; authors Aditya Ramabadran, Simon Mahns, Tobias Gessler, et al.) · 1 October 2026

   “with nothing between it and the car except for openpilot’s low-level actuation (comma.ai, 2026).”

4. E4 DRIVINGBENCH: CAN VISION-LANGUAGE MODELS DRIVE A TOYOTA COROLLA? arXiv (preprint; authors Aditya Ramabadran, Simon Mahns, Tobias Gessler and others)

   “Grok 4.6 and GPT-5.6 Sol failed to pass the first corner.”

5. E5 DRIVINGBENCH: CAN VISION-LANGUAGE MODELS DRIVE A TOYOTA COROLLA? arXiv (preprint; authors Aditya Ramabadran, Simon Mahns, Tobias Gessler and others)

   “The model drives it via three Model Context Protocol tools (Anthropic, 2024): observe, set_motion and stop_now.”

6. E6 DRIVINGBENCH: CAN VISION-LANGUAGE MODELS DRIVE A TOYOTA COROLLA? arXiv (preprint; authors Aditya Ramabadran, Simon Mahns, Tobias Gessler and others)

   “we retrofitted a comma four running a modified version of the openpilot software”

7. E7 DRIVINGBENCH: CAN VISION-LANGUAGE MODELS DRIVE A TOYOTA COROLLA? arXiv (preprint; authors Aditya Ramabadran, Simon Mahns, Tobias Gessler and others)

   “Its second attempt, however, finished the entire course and parked in the blue finish zone.”

8. E8 DrivingBench: Can Vision-Language Models Drive a Toyota Corolla? arXiv (DrivingBench authors) · 24 September 2026

   “Grok 4.6 in its three attempts achieved 8%, 11%, and 10%; and GPT-5.6 Sol achieved 6% in all three attempts.”

9. E9 DrivingBench: Can Vision-Language Models Drive a Toyota Corolla? arXiv (DrivingBench authors) · 24 September 2026

   “Astra is the only model to finish the course, on its second attempt, with no other attempt passing 50% of the course.”

10. E10 Report · DrivingBench DrivingBench · 1 January 2026

    “Below you can see the 3 MCP tools we give our agents (in their native harness - Codex/Claude Code/Cursor).”

11. E11 Diplomacy and CICERO Meta AI · 1 January 2026

    “CICERO achieved more than 2x the average score of its opponents.”

The only driving evidence is a DrivingBench preprint dated September–October 2026 (E1, E4, E8). It tested GPT-5.6 Sol rather than GPT-5, and that model reached 6% of the course and failed to pass the first corner. The test also gave every model MCP tools and openpilot actuation (E3, E5, E6, E10), so it does not measure GPT-5 driving without external aids by the end of 2025. For Diplomacy, the evidence describes only Cicero's own results (E11) and nothing on how GPT-5 did against it. The evidence does not show either part for GPT-5 without external aids, so the claim is unresolvable.