GPT-5: Unified Intelligence¶

Authors: OpenAI Published: August 7, 2025 System Card: cdn.openai.com/gpt-5-system-card.pdf Announcement: openai.com/index/introducing-gpt-5/

Why This Matters¶

GPT-5 is OpenAI's unified model that merged the GPT and o-series lines:

🧠 Unified architecture - One model that routes between fast and deep reasoning
🎯 94.6% on AIME 2025 - State-of-the-art math without tools
💻 74.9% SWE-bench Verified - Near-human software engineering
🔍 80% fewer hallucinations - Dramatic factual accuracy improvement
⚡ 50-80% fewer tokens - More efficient than o3 at same quality

Real-world impact: - Ended the confusing GPT-4/o1/o3 product split - Set new benchmarks across math, coding, and multimodal - Reduced hallucination to lowest levels ever measured - Became the model everything else is benchmarked against in 2025

The insight: Don't make users choose between fast and smart. Build one model with a router that automatically decides when to think deeply and when to respond quickly.

The Breakthrough¶

Unifying GPT and o-Series¶

The problem before GPT-5:

GPT-4o: Fast, good at general tasks, bad at hard reasoning
o1/o3:  Slow, excellent at reasoning, expensive, overkill for simple tasks

Users had to choose: "Is this question hard enough for o3?"
Wrong choice = wasted money or bad answers

GPT-5's solution:

Single model with internal router:
  Simple question → Fast path (GPT-level speed)
  Complex question → Deep thinking (o-series reasoning)
  Ambiguous       → Router decides based on complexity

User experience: Just ask. The model figures out how hard to think.

Architecture¶

GPT-5 internal structure:

User query → Router
              ├→ Fast model (most queries, instant response)
              └→ Thinking model (hard problems, extended reasoning)

The router considers:
- Conversation type
- Problem complexity
- Tool requirements
- User's explicit intent ("think carefully about this")

Performance¶

Mathematics¶

AIME 2025 (Competition Math):

Model	Score
GPT-4o	~20%
Claude 3.5 Sonnet	~25%
OpenAI o3	88.9%
GPT-5	94.6%
GPT-5 (with tools)	96.4%

State-of-the-art math reasoning.

Coding¶

SWE-bench Verified (Real-world software engineering):

Model	Score
GPT-4o	38.4%
Claude 3.5 Sonnet	49.0%
OpenAI o3	69.1%
GPT-5	74.9%

Aider Polyglot (Multi-language coding):

Model	Score
GPT-4o	~65%
Claude 3.5 Sonnet	~70%
GPT-5	88.0%

Multimodal¶

MMMU (Multimodal Understanding):

Model	Score
GPT-4o	69.1%
Gemini 2.5 Pro	75.2%
GPT-5	84.2%

Hallucination Reduction¶

The most impactful improvement:

With web search enabled:
  GPT-5 vs GPT-4o: 45% fewer factual errors

With thinking enabled:
  GPT-5 vs o3: 80% fewer factual errors

This is the biggest single-generation reduction
in hallucination rate ever measured.

Efficiency¶

GPT-5 thinking vs o3:

Same or better quality
50-80% fewer output tokens
= Much cheaper to run
= Much faster responses

GPT-5 learned when NOT to think deeply

Key Features¶

Adaptive Thinking¶

User: "What's the capital of France?"
GPT-5: [Fast path] "Paris."
(No thinking needed, instant response)

User: "Prove that the square root of 2 is irrational."
GPT-5: [Thinking path]
<thinking>
I'll use proof by contradiction...
Assume sqrt(2) = p/q where p,q are integers with no common factors...
Then 2 = p²/q², so p² = 2q²...
This means p² is even, so p is even...
Let p = 2k, then 4k² = 2q², so q² = 2k²...
This means q is also even, contradicting our assumption.
</thinking>
"Here's a proof by contradiction..."

Tool Use¶

GPT-5 can use tools when helpful:

- Web search (for current information)
- Code execution (for computation)
- File analysis (for documents)
- Image generation (via DALL-E)

The model decides WHEN to use tools
No explicit instruction needed

Health and Safety¶

HealthBench Hard: 46.2%
(Evaluates medical question answering)

GPT-5 shows improved performance on sensitive domains
Better at acknowledging uncertainty
More careful with medical/legal/financial advice

GPT-5.2 (Late 2025)¶

Incremental improvement released months later:

Key improvements:
- Better reasoning efficiency
- Improved code generation
- Enhanced multilingual capabilities
- Faster inference
- Lower cost

GPT-5.2 became the default ChatGPT model

Practical Usage¶

API Access¶

from openai import OpenAI

client = OpenAI()

# GPT-5 automatically routes between fast and thinking modes
response = client.chat.completions.create(
    model="gpt-5",
    messages=[
        {"role": "user", "content": "Solve this optimization problem..."}
    ]
)

print(response.choices[0].message.content)

With Thinking Control¶

# Force deep thinking
response = client.chat.completions.create(
    model="gpt-5",
    messages=[
        {"role": "user", "content": "Think carefully: prove Fermat's Last Theorem for n=3"}
    ],
    reasoning_effort="high"  # Force deep thinking
)

# Quick response
response = client.chat.completions.create(
    model="gpt-5",
    messages=[
        {"role": "user", "content": "What's 2+2?"}
    ],
    reasoning_effort="low"  # Skip deep thinking
)

Pricing¶

GPT-5:
  Input: ~$10/1M tokens
  Output: ~$30/1M tokens
  Thinking tokens: Varies

GPT-5 mini:
  Significantly cheaper
  Good for most tasks

Impact on the Field¶

Ended the Model Confusion¶

Before GPT-5:
  "Should I use GPT-4o or o1 or o3 or o3-mini?"
  Confusing product lineup

After GPT-5:
  "Just use GPT-5"
  One model, automatic routing

New Benchmark Standard¶

GPT-5 became the new bar:
- Every new model compared to GPT-5
- Claude 4.5 Opus, Gemini 3, DeepSeek targeted GPT-5 level
- SWE-bench Verified became the key coding benchmark

Hallucination Progress¶

The 80% reduction in hallucination (vs o3) was
the most significant safety improvement in a generation.

Showed that reasoning + retrieval can dramatically
reduce factual errors.

Limitations¶

1. Still Hallucinates¶

80% fewer errors is impressive but not zero
Still can't be fully trusted for critical decisions
Verification still required

2. Expensive for Thinking¶

When GPT-5 decides to think deeply:
- Uses many more tokens
- Significantly more expensive
- Slower response time
- Not always necessary

3. Closed Source¶

No weights, no architecture details
Can't fine-tune (only through OpenAI's API)
Dependent on OpenAI's infrastructure and pricing

4. Router Imperfect¶

Sometimes thinks too deeply on simple questions
Sometimes doesn't think enough on hard questions
Users can override, but default routing isn't perfect

Comparison with Competitors¶

GPT-5 vs Claude Opus 4.5 vs Gemini 3¶

Aspect	GPT-5	Claude Opus 4.5	Gemini 3 Pro
SWE-bench	74.9%	80.9%	76.2%
AIME 2025	94.6%	~85%	~88%
MMMU	84.2%	~78%	~80%
Hallucination	Lowest	Low	Low
Agentic tasks	Good	Best	Good
Context	1M+	200K	1M+
Open source	No	No	No

Each model leads in different areas.

Key Takeaways¶

Unified model - Merged fast (GPT) and deep (o-series) into one system
Adaptive routing - Automatically decides how hard to think
94.6% AIME - State-of-the-art mathematical reasoning
80% fewer hallucinations - Biggest factual accuracy improvement ever
50-80% more efficient - Better than o3 with far fewer tokens

Bottom line: GPT-5 ended the era of choosing between "fast" and "smart" models. Its unified architecture with adaptive routing set a new standard for what frontier AI should look like - one model that knows when to think quickly and when to think deeply.

GPT-5: Unified Intelligence¶

Why This Matters¶

The Breakthrough¶

Unifying GPT and o-Series¶

Architecture¶

Performance¶

Mathematics¶

Coding¶

Multimodal¶

Hallucination Reduction¶

Efficiency¶

Key Features¶

Adaptive Thinking¶

Tool Use¶

Health and Safety¶

GPT-5.2 (Late 2025)¶

Practical Usage¶

API Access¶

With Thinking Control¶

Pricing¶

Impact on the Field¶

Ended the Model Confusion¶

New Benchmark Standard¶

Hallucination Progress¶

Limitations¶

1. Still Hallucinates¶

2. Expensive for Thinking¶

3. Closed Source¶

4. Router Imperfect¶

Comparison with Competitors¶

GPT-5 vs Claude Opus 4.5 vs Gemini 3¶

Key Takeaways¶

Further Reading¶

Official Resources¶

GPT-5: Unified Intelligence¶

Why This Matters¶

The Breakthrough¶

Unifying GPT and o-Series¶

Architecture¶

Performance¶

Mathematics¶

Coding¶

Multimodal¶

Hallucination Reduction¶

Efficiency¶

Key Features¶

Adaptive Thinking¶

Tool Use¶

Health and Safety¶

GPT-5.2 (Late 2025)¶

Practical Usage¶

API Access¶

With Thinking Control¶

Pricing¶

Impact on the Field¶

Ended the Model Confusion¶

New Benchmark Standard¶

Hallucination Progress¶

Limitations¶

1. Still Hallucinates¶

2. Expensive for Thinking¶

3. Closed Source¶

4. Router Imperfect¶

Comparison with Competitors¶

GPT-5 vs Claude Opus 4.5 vs Gemini 3¶

Key Takeaways¶

Further Reading¶

Official Resources¶

Related Work¶

Related in This Collection¶