GPT-5: Key characteristics, pricing and model card

• GPT-5 costs $1.25/m input vs. $2.50 for GPT-4o and beats Claude 4.1 ($15/m), Gemini 2.5 Pro ($2.50/m), and Grok 4 ($3/m)—with 90% discount on cached tokens
• Replaces six different OpenAI models (GPT-4o, o3, o4-mini, etc.) with unified three-tier system: regular/mini/nano, each with four reasoning levels
• "Safe-completions" training means GPT-5 provides moderated answers instead of refusing requests—focusing on output safety rather than binary intent classification
• Prompt injection attacks still succeed 56.8% of the time (k=10)—best in class but confirms it remains an unsolved problem for production apps
• Hidden API feature lets you access reasoning traces via "reasoning": {"summary": "auto"} in the responses endpoint

Simon Willison's two-week early access review reveals GPT-5 as OpenAI's aggressive market consolidation play. The model replaces their fragmented lineup (GPT-4o, o3, o4-mini, GPT-4.1-nano, o3 Pro) with three tiers that are both better and significantly cheaper. The pricing strategy is striking: GPT-5 at $1.25/m input undercuts GPT-4o by 50% and beats every major competitor—Claude Opus 4.1 ($15/m), Sonnet 4 ($3/m), Grok 4 ($3/m), and Gemini 2.5 Pro ($2.50/m). The 90% discount on cached tokens makes chat interfaces particularly cost-effective.

The system card reveals three key architectural decisions. First, "safe-completions" training replaces binary refusals with content moderation—the model provides answers but moderates output to avoid harmful content, especially for dual-use cases like biology or cybersecurity. Second, post-training specifically targets sycophancy using production conversation data with reward signals. Third, the model was optimized for ChatGPT's three most common use cases: writing, coding, and health. The ChatGPT version uses a real-time router that switches between fast and reasoning models based on complexity and explicit user intent ("think hard about this"), while the API offers simpler control via three models with four reasoning effort levels.

The prompt injection results are sobering: external red teams achieved 56.8% attack success rate (k=10 queries), beating all competitors but still failing more than half the time. This is the best result in the field—Claude 3.7 scores 63.6%, everything else is 70%+—but confirms prompt injection remains fundamentally unsolved. Willison emphasizes this means developers can't assume it won't be a problem. An undocumented API feature lets you access reasoning traces via the responses endpoint with "reasoning": {"summary": "auto"}—critical for avoiding long delays while invisible thinking tokens generate. His verdict: not a dramatic leap, but consistently competent and now his daily driver.