Version of GPT-5 that produces smarter and more precise responses with deeper reasoning.
GPT-5 Pro is OpenAI's most advanced reasoning model, first announced at DevDay 2025 on October 6. It extends GPT-5's intelligence ceiling with deeper reasoning chains, a 272K maximum output limit (vs. 128K on standard GPT-5), and a mandatory reasoning_effort: high setting — there is no low-power mode. On TheRouter, openai/gpt-5-pro is accessible via the standard chat completions endpoint for users who need near-SOTA inference on the hardest math, code, and scientific reasoning problems.
GPT-5 Pro sits at the top of OpenAI's model hierarchy, above GPT-5 (general-purpose), GPT-5 Mini (balanced), and GPT-5 Nano (cost-efficient). It outputs exclusively as text from text and image inputs. Pricing on TheRouter is $18.75 per 1M input tokens and $150 per 1M output tokens (1.25x upstream). Unlike lower-tier GPT-5 variants, GPT-5 Pro supports only reasoning_effort: high — every response is a deep, multi-step reasoning pass. This makes it unsuitable for latency-sensitive or simple classification tasks but unmatched for difficult analytical work.
| Type | Rate |
|---|---|
| Input | $16.20 / 1M tokens |
| Output | $129.60 / 1M tokens |
| Release date | 2025-10-06developers.openai.com ↗ | verified |
| Knowledge cutoff | September 30, 2024platform.openai.com ↗ | verified |
| Max output tokens | 272,000 (API); YAML listing: 128,000 (TheRouter configured limit)simonwillison.net ↗ | verified |
| Reasoning mode | Mandatory reasoning_effort: high onlydevelopers.openai.com ↗ | verified |
| API availability | Responses API only (native); TheRouter exposes via chat completions endpointdevelopers.openai.com ↗ | verified |
| License | OpenAI Terms of Use (proprietary, API-only) | verified |
| Benchmark | Distribution | Score | Source |
|---|---|---|---|
AIME 2025 Perfect score on the 2025 American Invitational Mathematics Examination when GPT-5 Pro uses Python tools. Without tools, GPT-5 itself scores 94.6%. | 100%% | openai.com ↗ | |
GPQA Diamond SOTA on Google's PhD-level Q&A dataset without external tools. GPT-5 Pro's extended reasoning surpasses all prior models on this benchmark. | 88.4%% | openai.com ↗ | |
Humanity's Last Exam GPT-5 Pro with tools and reasoning achieves 42% accuracy on expert-crafted questions spanning multiple disciplines. Nearly double the prior SOTA (o3: ~27%). | 42%% | openai.com ↗ | |
SWE-bench Verified GPT-5 (base) with thinking scores 74.9%. GPT-5 Pro (extended reasoning) expected to score higher, though OpenAI has not published a separate Pro-variant breakdown for SWE-bench. | 74.9%% | openai.com ↗ | |
MMMU GPT-5 family-wide MMMU score. Improves over GPT-4o on multimodal understanding across charts, diagrams, and visual reasoning tasks. | 84.2%% | openai.com ↗ | |
HealthBench Hard SOTA on physician-defined medical diagnosis and reasoning benchmark. GPT-5 Pro is the first model designed specifically as 'an active health thought partner' per OpenAI. | 46.2%% | openai.com ↗ |
Use the global api.therouter.ai endpoint shown below for new integrations; the legacy China accelerated endpoint is retired.
curl https://api.therouter.ai/v1/chat/completions -H "Content-Type: application/json" -H "Authorization: Bearer $THE_ROUTER_API_KEY" -d '{
"model": "openai/gpt-5-pro",
"messages": [
{"role": "user", "content": "Summarize the key points from this input."}
]
}'GPT-5 Pro is called via TheRouter's standard chat completions endpoint. Unlike other GPT-5 variants, reasoning_effort is locked to 'high' — the model always performs deep reasoning. The 272K max output makes this ideal for long-form generation.
curl https://api.therouter.ai/v1/chat/completions \
-H "Authorization: Bearer $THEROUTER_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "openai/gpt-5-pro",
"messages": [
{"role": "user", "content": "Prove the Riemann hypothesis step by step, explaining each lemma."}
]
}'OpenAI unveiled GPT-5 Pro as its most advanced reasoning model at DevDay 2025. Available exclusively through the Responses API with mandatory reasoning_effort:high, a 272K max output, and $15/$120 upfront pricing (TheRouter: $18.75/$150). The model achieves 100% on AIME 2025 with tools and 88.4% on GPQA Diamond without tools.
OpenAI launched the GPT-5 family in August 2025, with GPT-5 Pro initially available only to ChatGPT Pro subscribers. The API version followed at DevDay in October 2025 with a dedicated Responses API implementation.
Natively, no — GPT-5 Pro is only available through OpenAI's Responses API. However, TheRouter maps it to the standard chat completions endpoint, so you can call it via the familiar OpenAI SDK. The reasoning_effort parameter is locked to 'high'.
272,000 tokens in OpenAI's own API, compared to 128,000 for standard GPT-5. On TheRouter, the YAML-configured limit is currently 128,000 — if you need the full 272K, contact support to discuss custom configuration.
Yes — GPT-5 Pro accepts image inputs (text and image modality) but outputs text only. Images are processed through the deep reasoning chain for visual understanding tasks.
$18.75 per 1M input tokens and $150 per 1M output tokens (1.25x upstream pricing from OpenAI's $15/$120). Cache-read discount is not available on GPT-5 Pro. Given the mandatory deep reasoning, a single complex query can consume hundreds to thousands of output tokens.
Every request to GPT-5 Pro goes through full deep reasoning, similar to telling GPT-5 or GPT-5 Mini to 'think long and hard'. Response times are measured in minutes rather than seconds, and output token costs accumulate quickly. This is deliberate — GPT-5 Pro is designed for problems where you'd wait longer for a higher-quality answer.
| source | URL | retrieved | |
|---|---|---|---|
| Release date | developers.openai.com ↗ | 2026-05-29 | verified |
| Knowledge cutoff | platform.openai.com ↗ | 2026-05-29 | verified |
| Max output tokens | simonwillison.net ↗ | 2026-05-29 | verified |
| Reasoning mode | developers.openai.com ↗ | 2026-05-29 | verified |
| API availability | developers.openai.com ↗ | 2026-05-29 | verified |
| License | — | — | verified |
| AIME 2025 | openai.com ↗ | 2026-05-29 | verified |
| GPQA Diamond | openai.com ↗ | 2026-05-29 | verified |
| Humanity's Last Exam | openai.com ↗ | 2026-05-29 | verified |
| SWE-bench Verified | openai.com ↗ | 2026-05-29 | to verify |
| MMMU | openai.com ↗ | 2026-05-29 | to verify |
| HealthBench Hard | openai.com ↗ | 2026-05-29 | to verify |
| OpenAI launches GPT-5 Pro at DevDay 2025 | developers.openai.com ↗ | 2026-05-29 | verified |
| GPT-5 announced — Pro variant teased for Pro subscribers | openai.com ↗ | 2026-05-29 | verified |
| Is GPT-5 Pro available via the standard OpenAI Chat Completions API? | developers.openai.com ↗ | 2026-05-29 | to verify |
| What is the maximum output length of GPT-5 Pro? | simonwillison.net ↗ | 2026-05-29 | to verify |
| Can GPT-5 Pro see images? | platform.openai.com ↗ | 2026-05-29 | to verify |
| How much does GPT-5 Pro cost on TheRouter? | platform.openai.com ↗ | 2026-05-29 | to verify |
| What does 'reasoning_effort: high only' mean in practice? | developers.openai.com ↗ | 2026-05-29 | to verify |