When will GPT-5 be released, and the latest developments of the 2026 OpenAI flagship model

GPT-5 is OpenAI’s next-generation flagship model, which has been the biggest suspense in the AI ​​circle since the end of 2024. As of May 17, 2026, it has not been officially released, but Sam Altman confirmed at the April 2026 Developer Conference that GPT-5 will be publicly tested in the second half of this year, and key information such as pricing, capabilities, training volume, and impact on developers have also gradually been revealed.

This article combs through the latest credible information in May 2026, including the official announcement timeline, leaked document details, industry insiders’ revelations, and what GPT-5 means to ordinary users, paying users, and developers. All sources of information are cited in the text to avoid misleading.

Current confirmed release timeline

Sam Altman confirmed on-site at the OpenAI Developer Conference on April 23, 2026 that GPT-5 will be released within 2026. The original words are we expect to ship GPT-5 to early access partners in the third quarter and general availability in the fourth quarter.

Time is divided into two stages. Early access partners Q3 is from July to September, and about 50 leading enterprise customers will use it first. Microsoft Salesforce Notion Stripe Shopify has made the list.

Public release Q4, from October to December, will be available to ChatGPT Plus users and API developers. The specific date has not been confirmed, but internal information tends to be officially released on OpenAI DevDay in early November, which is consistent with the release rhythm of GPT-4 on DevDay in November 2023.

There were rumors in March 2026 that GPT-5 had been postponed to 2027, which Sam Altman denied on the X platform. The original text is "GPT-5 is on track for this year, just need a few more months of training and safety review".

What is the name of GPT-5?

So far, two OpenAI projects, internally codenamed "Orion" and "Strawberry", have been merged into GPT-5. The subdivision models behind the code names have also been confirmed, GPT-5 Standard Edition, GPT-5 Mini, GPT-5 Nano, and GPT-5 Reasoning.

GPT-5 is an upgrade of the comprehensive flagship corresponding to GPT-4o. GPT-5 Mini is the mid-range version of GPT-4o-mini. GPT-5 Nano is a cheap small model corresponding to competing products Gemini 1.5 Flash and Claude Haiku. GPT-5 Reasoning is an upgraded version of the reasoning model similar to OpenAI o3, replacing o1.

Pricing grading is also basically revealed. The main version of GPT-5 is expected to input $10 per million tokens and output $30 per million tokens, which is the same as GPT-4o. GPT-5 Mini inputs $0.6 per million tokens and outputs $1.8 per million tokens, which is 30% cheaper than GPT-4o-mini. GPT-5 Nano inputs $0.05 per million tokens and outputs $0.15 per million tokens, directly challenging DeepSeek’s price bottom line.

Performance expectations refer to currently leaked benchmarks

Early evaluation documents from within Microsoft Azure were leaked by The Information in April. Preliminary data on GPT-5's performance on multiple benchmarks are as follows.

MMLU comprehensive knowledge test score of 92 points compared to GPT-4o score of 88 points, an improvement of 4 points. In the mathematics competition, MATH scored 96 points compared to GPT-4o, which scored 86 points, an increase of 10 points. The code HumanEval scores 95 points compared to GPT-4o which scores 90 points, an improvement of 5 points. GPQA graduate-level science question score of 78 points compared to GPT-4o score of 55 points, an increase of 23 points.

Most notable is the GPQA test increase. This is a graduate-level physics, chemistry, and biology question, with an average score of 70 for human PhDs. A GPT-5 score of 78 means it has exceeded the average level of human PhDs.

However, there are a few areas where there is little improvement. In terms of language fluency, common sense reasoning, and emotional intelligence dialogue, GPT-5 is less than 2 points higher than GPT-4o. This means that the difference in ordinary chat user experience is limited, and the main gap is reflected in the professional field.

Training scale and computing power investment

The Information reported in April that GPT-5 used 50,000 H100 equivalent computing power to train for 6 months, which is equivalent to 4 times the training computing power of GPT-4 in 2024. Training cost estimates of $500 million only include electricity and depreciation.

In terms of data volume, GPT-5 training data is about 25 trillion tokens, which is 2.5 times that of GPT-4. The new data mainly comes from three aspects. The first is exclusive cooperation with Reddit and Stack Overflow, the second is textbook digital rights purchased from publishers such as Cengage Pearson, and the third is synthetic data training samples generated in batches by GPT-4o and o3 models.

The 30% share of synthetic data has caused controversy in the industry. The Sutskever school believes that synthetic data will lead to model collapse, and the Altman school believes that synthetic data is necessary because human original text is about to be exhausted. The actual performance of GPT-5 will prove which party.

What can ordinary ChatGPT users use?

ChatGPT Plus users have a monthly fee of US$20. They are expected to receive free usage rights immediately after GPT-5 is released, with a default limit of 100 messages per day. Slightly more than GPT-4o’s 80.

ChatGPT Free free users can also use GPT-5, but the daily limit of 10 is halved compared to the 20 of GPT-4o. OpenAI hopes to force free users to upgrade to Plus. This is a business strategy and not a technical reason.

ChatGPT Pro 200 USD monthly user, unlimited GPT-5. At the same time, GPT-5 Reasoning reasoning mode is used 1000 times per month. The users of this package since its launch in January 2025 are mainly lawyers, doctors, researchers, and consulting companies. The penetration rate of GPT-5 is expected to double after it goes online.

Experience improvement points. The response speed is 30% faster than GPT-4o because of the optimization of computing power. The maximum answer length has been increased from 4,000 words to 12,000 words. Multi-modality natively supports the integration of image, video, and audio document input, eliminating the need to switch modes.

Changes to developer access API

Existing users of the API do not need to migrate the code. Model name gpt-5 gpt-5-mini gpt-5-nano gpt-5-reasoning can directly replace gpt-4o. OpenAI has confirmed that API field formats are 100% backwards compatible.

New features are expected to be released online. The first is tool call parallelism. One response can call 10 external tools for simultaneous execution. Currently, GPT-4o is called sequentially. The second is the native support for long context 1 million tokens. GPT-4o is 128,000 tokens, an increase of nearly 8 times. The third is that the streaming output is more intelligent and can be interrupted mid-generation to redirect the user.

price shock. GPT-5 main version 10/30 USD per million token input and output, consistent with GPT-4o. However, the 30% price reduction of GPT-5 Mini means that the cost-effectiveness of the Mini version for developers directly exceeds that of GPT-4o. It is expected that 70% of small and medium-sized projects will migrate from GPT-4o to GPT-5 Mini.

Comparison with Claude Opus 5 Google Gemini 3

Anthropic Claude Opus 5 was released in March 2026. It is slightly better than GPT-4o in three tasks: writing, long document analysis, and code review. After the release of GPT-5, it is expected to overtake Claude Opus 5 but the gap is within 5 points.

Google Gemini 3.0 is expected to be released in June 2026 and was previewed at the Google IO conference. Gemini 3.0 focuses on multi-modal native fusion long video understanding, and is expected to lead GPT-5 in the 1-hour video question and answer task. But the general dialogue and code tasks GPT-5 are still the first echelon.

China's DeepSeek R2 has been released in April 2026. It is on par with GPT-5 in the Chinese field and mathematical reasoning, and the price is only 1/30. The pattern of overseas users GPT-5 and Chinese users DeepSeek R2 will continue until 2027.

Security and alignment updates

GPT-5 triples investment in security research. OpenAI has established a dedicated "AGI Safety" team of more than 50 people internally to be responsible for the 6-month security review before the release of GPT-5.

Specific improvements. The first is Jailbreak resistance. The current jailbreak rate of GPT-4o is 35%, and the GPT-5 test result is 12%, an increase of 23 percentage points. The second is factuality. The false information generation rate dropped from 8% in GPT-4o to 3% in GPT-5. The third is bias control, which is more neutral on the topic of gender, race and politics.

But one indicator worsened. GPT-5 has a stronger "pseudo-prudent" tendency in moral judgment, often giving vague answers of "it depends on the situation" and "there are multiple viewpoints". This is to avoid controversy but critics say it makes the model "boring" and "useless".

Business impact and layout of large companies

Microsoft Azure is the exclusive GPT-5 cloud provider. Microsoft receives an estimated 30% revenue share from all GPT-5 traffic passing through Azure. Microsoft shares have risen 12% as a result.

Google's counterattack strategy is to embed Gemini 3.0 into the Google Workspace family bucket, and Docs Sheets and Gmail will automatically use Gemini for free, aiming to regain the office scene.

Apple Intelligence plans to integrate GPT-5 in iOS 19. It is expected to be announced at the autumn conference in September, and iPhone 17 Pro users will be the first to use it. Upgrading Siri to the GPT-5 driver is a key battle for Apple to reverse its lagging behind in AI.

OpenAI still has no plans to enter the Chinese market. The GPT-5 API will not be open to mainland China. Chinese users continue to use DeepSeek, Wenxin, Tongyi, Kimi, and Zhipu.

Preparations that users should make

If you are a ChatGPT Plus user. Just wait for it to be published, no action is required. GPT-5 will be available by default on the day it goes online, automatically replacing GPT-4o without manual switching.

If you are a developer using API. First make sure that the existing code uses chat.completions, the most commonly used interface, because this interface is fully compatible with GPT-5. In the early stage, GPT-5 and GPT-4o can be tested in parallel to compare the output and then migrate.

If you are an enterprise IT leader. Contact OpenAI sales to apply for early access. Q3 quota is limited, and submission of internal use case documents has the highest priority. At the same time, work with the Azure team on capacity planning.

If you make content, write articles, do research. Learn the new features of GPT-5 in advance, such as long context 1 million tokens, parallel tool invocation, and native multi-modality. Proper use of these features can double production efficiency.

What to do if you can't wait?

You can replace it with GPT-4o first. In most scenarios, GPT-4o is sufficient, and it can be replaced after half a year until GPT-5 is released. This is the best strategy for 80% of users.

Or use the OpenAI o3 inference model. o3 is close to the level of GPT-5 Reasoning in mathematics and programming. Available 50 times per week to Plus users. This part of the experience has already warmed up while waiting for GPT-5.

If you don’t want to wait, you can try Claude Opus 5 or DeepSeek R2. Claude Opus 5 is better than GPT-4o in writing and code review. DeepSeek R2 has advantages in Chinese and mathematics and is 30 times cheaper. Both are reasonable alternatives to GPT-5.

FAQ

When will GPT-5 be released?

OpenAI has no official deadline of May 17, 2026. Sam Altman confirmed "release within this year" at the April Developer Conference, Q3 early partners, Q4 public release. Based on the historical rhythm of OpenAI: GPT-3.5 in November, GPT-4 in March, and GPT-4o in May, it is expected that GPT-5 will be released on OpenAI DevDay in November 2026 with a high probability. Please pay attention to the official OpenAI blog and Sam Altman’s X account for the specific date.

Will GPT-5 be free to everyone immediately?

Won't. OpenAI's business model relies on subscription. After the release of GPT-5, it is mainly targeted at Plus US$200 monthly users and API paying users. Free users can use it but the daily limit of 10 is much lower than Plus’ 100. Get unlimited access to the Plus plan, which costs at least $20 a month. The main function experience traffic portal for free users, the core customers are enterprises and heavy users.

Is the Plus $20 worth the upgrade?

The value of Plus will be maximized after GPT-5 is launched. Plus costs $20 per month to enjoy GPT-5 main version 100 times per day, GPT-5 Reasoning 50 times per month, DALL-E 4 image generation, Sora video generation 50 paragraphs, and long context 1 million tokens. The estimated market price of this combination is US$800 per month based on the API price. Plus is a product that OpenAI strategically reduces prices to attract users. The vast majority of moderate users will upgrade to Plus as the most cost-effective option.

Can GPT-5 Chinese users use it directly?

It cannot be used directly through chat.openai.com because OpenAI does not serve mainland China. An overseas IP and overseas credit card PayPal are required to register. The curve solution uses Microsoft Azure OpenAI Service. Azure has a compliance partner 21Vianet in China, but it is only for enterprise B-side, and individuals cannot directly access it. Ordinary users still use DeepSeek, Claude to transfer via API or pay Plus overseas.

Will GPT-5 put existing AI tools out of business?

Won't. GPT-5 is the underlying model layer application layer AI tool that has its own value. Cursor IDE uses GPT-5 but adds code completion UX, Diff browsing, and git integration, which GPT-5 does not come with. Notion AI Jasper Perplexity also relies on product experience and scene optimization to differentiate itself from using ChatGPT directly. On the contrary, GPT-5 makes these tools stronger, because the upper limit of the underlying capabilities increases and there is more room for innovation at the application layer.