What's New in GPT-5: Features, Pricing, Context Window, and How It Compares

GPT-5 at a glance

• Unified system: Routes between a fast general model and a deeper "GPT-5 thinking" mode; GPT-5 Pro offers even longer test-time reasoning for the hardest problems.
• Standout gains: SOTA-level results in coding (e.g., SWE-bench Verified), math, multimodal understanding, and health (HealthBench).
• Developer knobs: New verbosity and reasoning_effort (now includes minimal), plus custom tools and visible preamble messages between tool calls.
• Sizes & names: gpt-5, gpt-5-mini, gpt-5-nano; ChatGPT's non-reasoning chat model appears as gpt-5-chat-latest.
• Rollout: Default in ChatGPT from Aug 7, 2025; Enterprise/Edu access one week later.

What's actually new (and why it matters)

1) Unified routing = fewer model picks, better answers

GPT-5 decides when to answer fast and when to think longer, and it can respond to cues like "think hard about this." This reduces the friction of choosing models and boosts reliability on real-world tasks.

2) Coding that feels collaborative

On real coding work, GPT-5 is stronger than prior OpenAI reasoning models, with improvements in front-end generation, code editing, long-running agentic workflows, and fewer tool-calling errors. Devs also get verbosity control, minimal reasoning for latency, and custom tools (plaintext tool calls).

3) Health answers with better signal

OpenAI reports higher scores on HealthBench and reduced "deception" (over-confident claims) vs. earlier models—paired with "safe-completions" training and a published system card that details risk safeguards. (Reminder: it's not a doctor.)

4) Fewer hallucinations, less sycophancy

Internal evals show GPT-5 reduces factual errors vs. GPT-4o and o3 when reasoning, and it's trained to be more honest about what it can't do—useful for high-stakes workflows.

Pricing and plans (ChatGPT vs API)

API (standard processing rates):

• GPT-5: $1.25 / 1M input; $10 / 1M output
• GPT-5 mini: $0.25 / 1M input; $2 / 1M output
• GPT-5 nano: $0.05 / 1M input; $0.40 / 1M output
• Non-reasoning chat variant gpt-5-chat-latest is priced like GPT-5.

ChatGPT plans (user-facing):

Free (limited), Plus ($20), Pro ($200, includes GPT-5 Pro), Team, Enterprise; all include GPT-5 with varying usage limits and features.

Context window & output limits

API total context: ~400K tokens (~272K input + up to 128K reasoning/output). That enables very long documents, multi-file repos, and broad retrieval.

GPT-5 vs GPT-4 (quick reality check)

• Quality: GPT-5 sets new highs on math/coding/multimodal/health benchmarks relative to GPT-4-family models (e.g., 4.1/4o).
• UX: You'll see fewer hallucinations, better instruction-following, clearer honesty about limits, and smarter tool use.
• Dev controls: verbosity, minimal reasoning, custom tools, and improved long-context retrieval go beyond prior GPT-4-series APIs.

GPT-5 vs Claude / Gemini / Grok (state of play)

• Claude 3.7 Sonnet (Anthropic): hybrid "visible thinking" mode and fine-grained control; strong competitor in structured reasoning.
• Gemini 2.x (Google): "agentic era" push with 2.0/2.5 updates; deeper app integrations and research prototypes.
• Grok 4 (xAI): strong reasoning claims and noisy debate—Musk is publicly saying Grok 4 Heavy outperforms GPT-5; treat as claims, keep testing.

Bottom line: in day-to-day coding + agentic tool use, GPT-5's ecosystem (ChatGPT, API, and Microsoft surfaces) is now the most turnkey—verify with your own evals.

How to use GPT-5 (fast start)

In ChatGPT:

GPT-5 is now the default for signed-in users. Type your prompt; it will auto-route, and you can say "think hard about this" to invoke deeper reasoning. Pro users can select GPT-5 Pro from the picker. Rollout to Enterprise/Edu is one week after Aug 7, 2025.

In the API:

Pick a size (gpt-5, gpt-5-mini, gpt-5-nano), then tune reasoning_effort (try minimal for speed) and verbosity. Use custom tools when JSON is brittle, and lean on prompt caching/batch for cost. Context up to 400K helps for long repos & briefs.

Using connectors/agents:

ChatGPT's Agent and Connectors (e.g., Gmail, Google Calendar, GitHub) let GPT-5 pull context and take actions—with guardrails. Great for research, planning, and ops.

Comparison table (skim-friendly)

Practical prompts to try (copy/paste)

FAQs (rich-snippet length)

Sources

• Introducing GPT-5 (research/product overview, rollout, performance, safety): OpenAI
• Introducing GPT-5 for developers (API sizes, context 400K, dev knobs, benchmarks): OpenAI
• API Pricing (token rates across GPT-5 family): OpenAI
• ChatGPT Pricing (Free/Plus/Pro/Team/Enterprise features): OpenAI
• Competitor references and press coverage from major outlets