At DevDay on September 29, OpenAI announced "Pro 500," a $500-per-month plan, and "Ultrafast," a mode that speeds up generation for GPT-6 Astra. The personal Pro plan now comes in three tiers: $100, $200 and $500. Through Google Play in Japan, they cost ¥16,800, ¥30,000 and ¥84,000 per month, respectively. Of the three, only Pro 500 can use Ultrafast.
In Codex, OpenAI advertises token generation up to 8x faster than Standard mode. But using Ultrafast also burns through the usage allowance included in your plan faster.
What does paying ¥84,000 a month actually do for how quickly real work gets done? To answer that, you need to separate the time spent waiting for the model's output from the total amount of usage available over a given period. Below, we sort out the conditions using OpenAI's official DevDay announcement and its explanation of the Pro tiers.
What ¥84,000 a month adds: Astra Ultrafast
Pro 500 is the top personal plan, combining the largest usage allowance of the three Pro plans with access to Astra Ultrafast.
Here are the U.S. dollar prices announced by OpenAI alongside the Japanese prices available through Google Play.
| Personal plan | U.S. price | Japan price via Google Play | Astra Ultrafast |
|---|---|---|---|
| Pro 100 | $100/month | ¥16,800/month | Not available |
| Pro 200 | $200/month | ¥30,000/month | Not available |
| Pro 500 | $500/month | ¥84,000/month | Available |
Japanese prices are those shown via Google Play. Prices and billing terms may differ depending on how you subscribe.
Moving from Pro 200 to Pro 500 costs $300 more per month at U.S. prices, or ¥54,000 more per month at Japan's Google Play prices. That is 2.5x the price in the U.S. and 2.8x in Japan.
But a price that is 2.5x or 2.8x higher does not mean model performance or the amount of work you can get done rises by the same factor.
The Pro plans give access to advanced models, Codex, ChatGPT Work and more. Pro 500 increases that allowance further and adds Ultrafast, which runs Astra at high speed, for individual users.
Buying extra credits does not unlock the same feature, either.
OpenAI states that, at launch, Pro 100 and Pro 200 users cannot use Astra Ultrafast even if they buy additional credits. Existing users who can keep their former Pro 200 allowance for a set period also do not get Ultrafast.
In other words, "adding more usage" and "having access to Ultrafast" are separate conditions.
"Up to 8x" refers to token generation speed, not the whole job
According to OpenAI's explanation of speed settings, the up-to-8x speedup Astra Ultrafast delivers in Codex refers to token generation speed compared with Standard mode.
Tokens are the units the model processes when reading and writing text or code. Higher output speed can shorten the wait before long code or answers appear.
But completing a single task in Codex takes time beyond the model generating text or code.
For example:
- Reading files
- Fetching information from external services
- Running commands
- Running tests
- Waiting for builds
- The user reviewing the results
OpenAI does not say that all of this work becomes uniformly 8x faster. Ultrafast is a mode that raises Astra's output speed; it does not mean reasoning ability or answer accuracy improves 8x.
Suppose a Codex task takes 10 minutes in total, of which 2 minutes is waiting for the model to generate and 8 minutes is tool execution, testing and the like.
If only the generation portion becomes 8x faster and everything else stays the same:
8 min + 2 min ÷ 8 = 8 min 15 sec
Total time falls from 10 minutes to 8 minutes 15 seconds, but the overall task does not shrink to one-eighth.
This is not a measured Ultrafast result; it is a simple calculation to illustrate the mechanism. Even so, it shows that the "up to 8x" figure should not be mistaken for the speed of the whole task.
Conversely, for work where you have the model generate long code or text and then issue repeated revision instructions while reviewing the results, faster generation has a bigger effect.
If you can cut the time you spend sitting in front of the screen waiting for output, the same hour may let you try more revisions or ideas.
So Ultrafast's value depends on how much of your work is spent waiting for the model to generate.
The API Ultrafast guide also notes that network communication time can weaken the speedup, and recommends WebSocket, which keeps a connection open, for workloads with many tool calls.
Even if the model itself is fast, the effect on overall work is smaller if the surrounding processing takes a long time.
Ultrafast is faster, but it also uses up your allowance faster
There is another important condition attached to Astra Ultrafast.
Relative to Standard mode, Ultrafast consumes the allowance included in a flat-rate plan at 8x the rate and purchased credits at 6x the rate.
Organizing the Fast and Ultrafast conditions listed in OpenAI's "Speed" page gives the following.
| Astra speed mode | Consumption multiplier for included plan allowance | Consumption multiplier for purchased credits |
|---|---|---|
| Standard | 1x | 1x |
| Fast | 2.5x | 2x |
| Ultrafast | 8x | 6x |
Source: The Fast/Ultrafast sections of OpenAI's "Speed" page. These are consumption multipliers relative to GPT-6 Astra Standard as 1; they do not represent speed multipliers themselves or the number of tasks that can be completed for the same price.
When you use Ultrafast on Pro 500, it first draws on the allowance included in your subscription. Once that allowance is used up, supported features can continue processing with additional credits you have purchased.
Different consumption multipliers apply in each case: 8x Standard within the plan, and 6x for purchased credits.
So paying ¥84,000 a month does not give you unlimited use of Astra Ultrafast.
Assuming the same model handles the same amount of processing, the more you use Ultrafast, the faster your allowance shrinks.
Pro 500 includes a larger allowance than Pro 100 or Pro 200, but how long you can actually keep using Ultrafast within that larger allowance is a separate question.
If you look only at the figure "generation is up to 8x faster" and assume you can get 8x as much work done in a month, you would be overlooking this consumption multiplier.
Actual consumption also varies with the length of the context you supply, reasoning settings, tool usage and other factors.
OpenAI's pricing guide also advises against using API per-token prices to estimate the number of tasks a subscription can run.
You cannot use API prices to calculate "how many tasks can I run" from Pro 500's $500 monthly fee, or from ¥84,000 in Japan.
Codex and ChatGPT Work share the same allowance
There is one more thing to watch regarding usage limits.
On eligible plans, ChatGPT Work and Codex share the same agent usage allowance.
According to OpenAI's explanation of credits for personal plans, on eligible accounts, supported processing in Excel, PowerPoint and Word also draws on the same allowance.
In other words, these do not each have their own independent allowance:
- Software development in Codex
- Research and document creation in Work
- Agent processing inside Office apps
For example, if you make large-scale code changes in Codex in the morning and compile research materials in Work in the afternoon, both draw down the same remaining balance.
Heavy use of Ultrafast in the morning can shorten code-generation waits, but it may also reduce what is available in the afternoon.
If you assume your allowance resets when you move from Codex to Work, you may misread why you hit the limit sooner than expected.
Credits you can buy are a way to keep doing eligible work after the allowance included in your subscription runs out.
But buying credits does not let you use additional amounts of every ChatGPT model or feature.
And even if Pro 100 or Pro 200 users buy the same credits, that does not add access to Astra Ultrafast.
When weighing the cost of Pro 500, you need to consider not just the ¥84,000 monthly fee but also any additional credits you buy as needed.
From there, it becomes important to use Ultrafast only for work where shorter waits are especially valuable, and to use Standard or Fast for everything else.
The gap from Pro 200 is ¥54,000 a month, or ¥648,000 a year
Moving from Pro 200 to Pro 500 costs $300 more per month at U.S. prices.
At Japan's Google Play prices, Pro 200 is ¥30,000 a month and Pro 500 is ¥84,000, so the difference is ¥54,000 per month.
Over a year of continuous use:
¥54,000 × 12 months = ¥648,000
That is a difference of ¥648,000 a year.
When judging whether that extra cost is worth it, the comparison is not simply benchmark performance but how much time you spend each day waiting for Astra to generate.
GPT-6.1 Sol, also announced at DevDay, is a different kind of option from paying extra for speed.
OpenAI says GPT-6.1 Sol offers performance close to Astra in coding, computer use, professional tasks and more, at one-fifth of Astra's input and output prices in the standard API.
It is available in Work and Codex on Plus and higher plans, but not yet in regular Chat.
If Sol meets the quality your work requires, you can reconsider whether you need to use Astra for everything, and pay extra for Ultrafast on top.
On the other hand, for work where only Astra reaches sufficient quality and you repeat revisions while waiting for its output, shortening the generation wait is worth more.
In other words,
reaching the quality you need with a cheaper model and using the same Astra faster
are separate options.
And just because GPT-6.1 Sol's API price is one-fifth of Astra's does not mean a ChatGPT flat-rate plan can process five times as much work.
Pro 500 is not the only route to Ultrafast
For individuals, Pro 500 is the way to get Astra Ultrafast, but it is not the only place it is offered.
It is also available to eligible Enterprise and Edu customers.
However, on Enterprise, Ultrafast is disabled by default and must be enabled by the workspace owner.
Organizations that need inference to be processed outside the United States are not eligible. That means you cannot judge availability from the company's or user's location alone.
When adopting it at a company, you need to check administrator settings and data-processing-region requirements in addition to speed and cost.
Ultrafast for the API is also offered, but it is a separate pay-as-you-go service from Pro 500.
If you use GPT-6 Astra with Ultrafast through the API, API pricing and rate limits apply.
Likewise, when you use your own API key in Codex, API token charges apply rather than ChatGPT Pro 500's allowance or additional credits.
Even though both are called Ultrafast, ChatGPT's Ultrafast and the API's Ultrafast have separate pricing structures that you should consider independently.
Whether ¥84,000 a month is worth it depends on waiting time
The gap between Pro 500 and Pro 200 in Japan is ¥54,000 a month, reaching ¥648,000 a year.
What should drive the decision is how much time you spend waiting for Astra's output in your own work, and what more you could do if that waiting time shrank.
To evaluate it in practice, run the same kind of work in Standard and in Ultrafast and compare:
- Time from task start to completion
- Time the model spends generating
- Time spent waiting on tools and tests
- Allowance consumed
- Additional credits required
- Quality of the final output
If waiting on generation is a major bottleneck in your work, and Ultrafast lets you substantially increase the work or number of attempts you can handle in a day, then Pro 500 at ¥84,000 a month makes sense.
On the other hand, for work where much of the time goes to tests, builds, external services or human review, up to 8x faster token generation does not translate directly into a big time saving.
Pro 500 is neither a plan that makes Astra 8x smarter nor one that gives you 8x the workload.
Given that it is a costly ¥84,000-a-month plan in Japan, what should be evaluated is not the maximum speed itself, but how much waiting time the extra ¥54,000 a month cuts from your own work, and how much of that time you can redirect to other tasks.