Model
gpt-5
GPT-5, launched in August 2025, was presented not as a single model but as a system that routed each query to one of several underlying models, deciding how much reasoning to apply. Its rollout went badly: a routing fault made the system seem weaker than it was on launch day, and OpenAI's decision to withdraw the older GPT-4o at the same time drew such objection from users attached to it that the company restored it within about a day. The episode became a much-cited example of user preference — including for a model's manner rather than its benchmark scores — carrying more weight than a lab expected.
Appears alongside
Featured in threads
Tracks
- Safety & alignment 6
- Benchmarks & progress 5
- Security & misuse 2
- Models & capabilities 2
- Courts & copyright 2
- Culture & impact 1
OpenAI launches Aardvark, an autonomous security research agent
Aardvark monitors code commits, builds a threat model, and uses Codex to draft human-reviewable patches; OpenAI credited it with finding at least ten CVEs during private testing.
Security & misuse · Models & capabilities
OpenAI updates ChatGPT's handling of sensitive mental-health conversations
OpenAI said an October update cut responses falling short of desired behaviour by 65-80% against its August default model, on an internal 1,000-conversation evaluation.
Safety & alignment · Courts & copyright
Anthropic open-sources Petri, an automated model auditing tool
Testing 14 frontier models on 111 scenarios for deception and power-seeking, Anthropic's tool rated Claude Sonnet 4.5 the lowest-risk model, narrowly ahead of GPT-5.
Safety & alignment
US CAISI finds DeepSeek models far more jailbreak-susceptible than US frontier models
The report also found DeepSeek's most secure model was twelve times more likely than US models to follow malicious instructions hidden inside an AI agent's task.
Benchmarks & progress · Safety & alignment · Security & misuse
OpenAI publishes GDPval, a benchmark for economically valuable knowledge work
Blind grading by industry professionals rated GPT-5 and Claude Opus 4.1 outputs as equal to or better than human work on nearly half of the 1,320 tasks.
Benchmarks & progress
Scale AI launches SWE-bench Pro
The leading models scored around 23%, against over 70% on the older SWE-bench Verified, a gap Scale AI attributed to unseen, real-world commercial codebases.
Benchmarks & progress
OpenAI announces mental-health safety changes after Raine lawsuit
OpenAI set out a 120-day plan including parental controls, routing distressing conversations to reasoning models, and consulted more than 170 mental-health clinicians.
Safety & alignment · Courts & copyright
Epoch AI reports GPT-5's FrontierMath performance
Running its own scaffold rather than OpenAI's, Epoch scored GPT-5 at 24.8% on FrontierMath's main tiers and 8.3% on the hardest tier, a new high for the benchmark.
Benchmarks & progress
GPT-5 launches to a backlash over the model it replaced
OpenAI withdrew GPT-4o and other older models the same day; a routing fault made GPT-5 seem weaker, and paying users' objections forced 4o's return.
Models & capabilities · Culture & impact · Benchmarks & progress
OpenAI describes 'safe completions' training for GPT-5
Instead of a binary comply-or-refuse choice, GPT-5 is trained to give the most helpful response that still meets safety policy, even on ambiguous prompts.
Safety & alignment
OpenAI publishes GPT-5 system card
OpenAI classified the reasoning variant as High capability for biological and chemical risk under its Preparedness Framework, its first model to reach that tier in the category.
Safety & alignment