OpenAI 于 2026 年 9 月 3 日释放了 GPT-6 Astra 的限量预览版,并从 9 月 5 日开始分批次向大众开放。
在随后的一个星期里,你将随处可见“已兼容 GPT-6”的标识。
本文不讨论其性能有多强大,仅聚焦于三个可自行验证的要点:**它的确切名称、定价有多高、
以及如何通过单条指令来测试特定端点是否真正接入了该模型**。本文内涉及的所有返回结果均取自
9 月 5 日当天的实际运行数据,读者现今即可自行重现。
一、首先理清版本号问题,这也是大家最容易产生误解的地方
在过去的五个月中,“GPT-6”这一称呼已被提及不下三次,且每次所指代的对象都不尽相同:
|
2026 年 4 月 |
内部研发代号 Spud,传闻称“GPT-6 已完成预训练阶段”
|
GPT-5.5(于 4 月 23 日正式推出)
|
|
2026 年 6 月 |
— |
GPT-5.6 家族:Sol(旗舰版)/ Terra(均衡版)/ Luna(低成本版)
|
|
2026年9月3日
|
— |
GPT-6 Astra,这回确有其事了。
|
因此,目前网络上依然流传着许多四月份发布的“GPT-6实测”文章,但它们实际评测的对象其实是GPT-5.5。
要想辨别某篇文章探讨的是否为该版本,只需留意其中是否出现了gpt-6-astra——
此乃OpenAI官方提供的API模型标识符,同时也是唯一不会产生混淆的名称。
除此之外还存在一个GPT-6 Astra Pro版本,该版本仅向ChatGPT Pro及企业版订阅用户开放,并不包含在API之中。
二、核心参数:四组数值与一个迁移陷阱
|
API 模型 ID |
|
|
上下文窗口 |
1,050,000 token |
|
最大输出 |
128,000个token
|
|
知识截止 |
2026年4月30日
|
|
输入模态 |
支持文本与图片(原生模态不包含音频和视频)
|
🔴迁移过程中的一个坑在于推理强度。GPT-6 Astra 支持 low、medium、high、xhigh 以及 max 这五个档位,
却不再兼容none。若你的代码是从 GPT-5.6 迁移而来,且请求中包含none或minimal参数,
在切换模型后便会直接触发报错。官方推荐从low档位开始进行测试。
此处值得特别记录一下:这并非模型表现层面的问题,而是参数可选值范围发生了变更,
且报错提示一般仅显示参数无效,并不会明确指出“该档位在新版模型中已被移除”。
三、定价:成本较高,且输入若超过 272K 将触发跳档
计价单位为每百万 token,以美元结算:
|
标准计费(输入 ≤272K)
|
10 |
1 |
50 |
|
输入大于 272K
|
20 |
2 |
75 |
以下三个参照标准,有助于你将这一数值置于具体语境中理解:
- 与 Anthropic 针对 Claude Fable 5.1 制定的定价保持一致
- 相当于当前促销活动中 GPT-5.6 Sol 价格的两倍半
- Batch 与 Flex 模式的费用为标准定价的一半;Fast mode 的收费则是适用费率的二倍
⚠️务必留意跳档规则:并非仅超出部分按高价位计费,而是整次请求都会按高价重新计算。
当输入为 273K token 时,整体将按照每百万 20 美元来计费,而非前 272K 部分按 10 美元、超出部分才按 20 美元计算。
对于从事长文档处理的用户来说,这一规则将直接决定是否需要在 272K 处进行切分操作。
四、上线首日的真正难题:谁真正实现了兼容
新模型发布后的最初几天里,"已兼容"这个说法最不可靠——**在页面上添加一行文字的成本,
远低于实际对接一个模型**。好在这一点完全可以自行验证,不必轻信任何人的说法。
第一步:无需携带 key,直接向端点请求获取模型列表
curl -s https://端点地址/v1/models | head -c 300
9 月 5 日对七个地址进行了实际测试,原始返回结果如下:
|
|
401 |
提供的API密钥不正确 |
|
访问地址为 api.deepseek.com/v1/models
|
401 |
认证失败 (governor)
|
|
调用地址是 open.bigmodel.cn/api/paas/v4/models
|
401 |
请求头部缺失了Authorization参数
|
|
请求发送至 api.siliconflow.cn/v1/models
|
401 |
提供的Token无效
|
|
接口URL为 ark.cn-beijing.volces.com/api/v3/models
|
401 |
|
|
|
401 |
You did not provide an API key
|
|
|
200 |
完整的 JSON 数据,共计 17 个模型 ID
|
结论非常清晰:绝大多数端点返回的模型列表都受权限保护,如果没有提供 key 是无法获取的。
由此可见,首轮验证的核心目的并非为了“查看列表”,而是用来过滤掉那些连接口都不具备的地址——
收到 401 状态码代表你找对了入口,而收到 404 则意味着你连门都没找对。这两个返回码的差异,其实远超许多人的预期,至关重要。
⚠️ 另外还有一种不易察觉的情形:状态码明明是200,返回的却是一个HTML页面。
这意味着/v1/models在该域名下并非API接口路径,你的请求实际上被导向了网站的页面路由。
仅凭状态码很容易被蒙蔽,因此前面那条命令务必加上head -c 300来检查实际返回内容。
第二步:携带你自己的key再查询一次
curl -s https://端点地址/v1/models \
-H "Authorization: Bearer 你的key" | grep -io "gpt-6[a-z0-9.-]*"
只要有输出,就说明该端点的模型清单中确实包含了它。
第三步:实际发送一个请求过去(这一步最容易被忽略,但也最为关键)
清单中列出不等于真的能调用。有些服务会将不存在的模型名悄悄映射到其他模型上——
你以为自己在用GPT-6,实际跑的却是5.6甚至更早的版本,而返回结果一切正常,不会报错。
验证方法是发送一个最小化的请求,然后检查响应体中model字段返回的是什么:
curl -s https://端点地址/v1/chat/completions \
-H "Authorization: Bearer 你的key" \
-H "Content-Type: application/json" \
-d '{"model":"gpt-6-astra","messages":[{"role":"user","content":"hi"}],"max_tokens":8}' \
| grep -o '"model":"[^"]*"'
如果请求中写的模型名和响应中返回的对不上,就说明被映射了。这一点对于中转服务尤其需要验证,
这比任何宣传页面都靠谱——毕竟你检验的是它自身返回的数据。
五、一条建议
新模型上线后的两周内,不要直接将生产流量切换过去。原因并非模型本身有问题,而是周边生态尚未跟上:
- 迁移参数方面的坑(比如前面提到的none被取消)通常需要一周时间才会有人总结出来
- 各家中转服务的接入进度参差不齐,同一个模型名在不同端点上的实际表现可能存在差异
- 价格跳档规则会导致长文档场景下的成本估算严重失准
先用第三步验证确认你手上的端点是真实的,再以小流量运行几天,看清楚账单情况后再做决定。
附录:本文所使用的可复现命令
这三条命令均无需注册任何服务即可运行第一条,而后续两条使用你现有的 key 便足够了。
自行验证不花一分钱,但若被一句“已兼容”误导,代价可就不止这些了。
*笔者自身也从事 AI 接入相关的服务。上文表格中 flex-api.code2ai.codes 那一行属于
自家服务的端点,将其列入表中的原因在于,它是这七家当中唯一一个无需 key 即可查阅完整模型列表的——
关于这一点,你只需自行 curl 测试一番便可证实,无需盲信我。*
本文来源:https://cloud.tencent.com/developer/article/2737783
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
On September 3, 2026, OpenAI released a limited preview of GPT-6 Astra, and starting September 5, it began rolling out to the general public in batches.
In the following week, you will see "GPT-6 compatible" badges everywhere.
This article does not discuss how powerful it is; it focuses only on three points you can verify yourself: **its exact name, how expensive it is,
and how to test whether a specific endpoint is truly connected to the model with a single command**. All results returned in this article were taken from
actual runs on September 5, and readers can reproduce them right now.
1. First, let's clarify the version number issue, which is also where most misunderstandings occur
Over the past five months, the term "GPT-6" has been mentioned no fewer than three times, and each time it referred to something different:
April 2026
Internal R&D codename Spud, with rumors that "GPT-6 has completed the pre-training phase"
GPT-5.5 (officially launched on April 23)
June 2026
—
GPT-5.6 family: Sol (flagship) / Terra (balanced) / Luna (low-cost)
September 3, 2026
—
GPT-6 Astra, and this time it is real.
Therefore, many "GPT-6 hands-on review" articles from April still circulate online, but what they actually reviewed was GPT-5.5.
To tell whether an article is about this version, just check whether gpt-6-astra appears in it—
this is OpenAI's official API model identifier and the only name that cannot cause confusion.
There is also a GPT-6 Astra Pro version, which is available only to ChatGPT Pro and enterprise subscription users and is not included in the API.
2. Core specifications: four sets of values and one migration pitfall
API model ID
gpt-6-astra
Context window
1,050,000 tokens
Maximum output
128,000 tokens
Knowledge cutoff
April 30, 2026
Input modalities
Supports text and images (native modalities do not include audio or video)
🔴 One pitfall in migration involves reasoning effort. GPT-6 Astra supports five levels: low, medium, high, xhigh, and max,
but it no longer supports none. If your code is migrated from GPT-5.6 and your request contains the none or minimal parameter,
it will directly trigger an error after switching models. OpenAI recommends starting tests with the low level.
This is worth noting: it is not a model performance issue, but a change in the range of allowed parameter values,
and the error message usually only says the parameter is invalid, without explicitly stating that "this level has been removed in the new model".
3. Pricing: costs are high, and input exceeding 272K triggers a pricing tier jump
Pricing is per million tokens, settled in US dollars:
Standard billing (input ≤272K)
10
1
50
Input greater than 272K
20
2
75
The following three reference points help put this number into context:
Matches Anthropic's pricing for Claude Fable 5.1
Equivalent to two and a half times the price of GPT-5.6 Sol during the current promotion
Batch and Flex modes cost half the standard price; Fast mode charges twice the applicable rate
⚠️ Be sure to note the tier-jump rule: it is not that only the excess portion is billed at the higher rate; the entire request is recalculated at the higher rate.
When input is 273K tokens, the whole request is billed at $20 per million, rather than the first 272K at $10 and the excess at $20.
For users handling long documents, this rule directly determines whether you need to split content at the 272K boundary.
4. The real challenge on launch day: who actually achieved compatibility
In the first few days after a new model launches, "compatible" claims are the least reliable—**the cost of adding a line of text to a webpage
is far lower than actually integrating a model**. Fortunately, this can be verified on your own without trusting anyone's word.
Step 1: Without carrying a key, directly request the model list from the endpoint
Code language: bash
Copy
On September 5, seven addresses were actually tested, and the raw results are as follows:
api.openai.com/v1/models
401
The provided API key is incorrect
Request address: api.deepseek.com/v1/models
401
Authentication failed (governor)
Call address: open.bigmodel.cn/api/paas/v4/models
401
The request header is missing the Authorization parameter
Request sent to api.siliconflow.cn/v1/models
401
The provided Token is invalid
API URL: ark.cn-beijing.volces.com/api/v3/models
401
AuthenticationError
dashscope.aliyuncs.com/compatible-mode/v1/models
401
You did not provide an API key
flex-api.code2ai.codes/v1/models
200
Complete JSON data, containing 17 model IDs in total
The conclusion is very clear: the model lists returned by most endpoints are permission-protected and cannot be retrieved without a key.
Thus, the core purpose of the first verification is not to "view the list" but to filter out addresses that do not even have the interface—
receiving a 401 means you found the right entrance, while receiving a 404 means you have not even found the door. The difference between these two status codes is far greater than many people expect, and it is critical.
⚠️ There is also an easily overlooked case: the status code is 200, but what is returned is an HTML page.
This means that /v1/models is not an API path on that domain, and your request was actually routed to the website's page route.
Relying solely on the status code can easily mislead you, so be sure to add head -c 300 to the previous command to check the actual returned content.
Step 2: Query again with your own key
Code language: bash
Copy
As long as there is output, it proves that the endpoint's model list truly contains it.
Step 3: Actually send a request (this step is the most easily overlooked, but also the most critical)
Being listed does not mean it can actually be called. Some services silently map nonexistent model names to other models—
you think you are using GPT-6, but you are actually running 5.6 or even an earlier version, and the returned results look normal, with no error.
The verification method is to send a minimal request and then check what the model field in the response body returns:
Code language: bash
Copy
If the model name written in the request does not match the one returned in the response, then mapping has occurred. This is especially important to verify for relay services,
and it is more reliable than any promotional page—after all, you are checking the data returned by the service itself.
5. One suggestion
Within two weeks after a new model launches, do not switch production traffic directly. The reason is not that the model itself has problems, but that the surrounding ecosystem has not caught up yet:
Migration parameter pitfalls (such as the removal of none mentioned earlier) usually take about a week before someone summarizes them
The integration progress of various relay services is uneven, and the actual behavior of the same model name may differ across endpoints
The pricing tier-jump rule can seriously skew cost estimates for long-document scenarios
First use Step 3 to confirm that the endpoint you have is genuine, then run a small amount of traffic for a few days, and check the billing details before making a decision.
Appendix: reproducible commands used in this article
Code language: bashCopy# 1. Check whether the endpoint provides an Anthropic/OpenAI-format model API curl -s https://endpoint-address/v1/models | head -c 300 # 2. Check whether GPT-6 appears in the model list (key required) curl -s https://endpoint-address/v1/models -H "Authorization: Bearer your-key" \ | grep -io "gpt-6[a-z0-9.-]*" # 3. Actually send a request and verify whether the model field in the response matches curl -s https://endpoint-address/v1/chat/completions \ -H "Authorization: Bearer your-key" -H "Content-Type: application/json" \ -d '{"model":"gpt-6-astra","messages":[{"role":"user","content":"hi"}],"max_tokens":8}' \ | grep -o '"model":"[^"]*"'
These three commands require no registration for the first one, and the latter two only need your existing key.
Self-verification costs nothing, but being misled by a single "compatible" claim can cost far more.
*The author also works on AI integration services. In the table above, the flex-api.code2ai.codes row belongs to
our own service's endpoint. It is included in the table because it is the only one among the seven that allows viewing the complete model list without a key—
you can verify this simply by running curl yourself; there is no need to take my word for it.*
gpt large model deployment cloud AI API openai #GPT-6 #gpt-6-astra #model API #endpoint verification
Table of contents 1. First, clarify the version number, where misunderstandings are most concentrated 2. Hardware specifications: four numbers and one migration pitfall 3. Pricing: on the high side, with a tier jump beyond 272K 4. The real test on launch day: who actually achieved compatibility Level 1: Without a key, directly request the model list from the endpoint Level 2: Query again using your own key Level 3: Actually send a request (this step is the most easily overlooked but also the most practical) 5. One suggestion Appendix: reproducible commands used in this article Related models and services
Shenzhen Tencent Computer Systems Co., Ltd. ICP filing/license number: 粤B2-20090059 粤公网安备44030502008569号![]()
Source of this article: https://cloud.tencent.com/developer/article/2737783 If there is any copyright issue, please notify us via email: 13951431913#139.com or submit a comment on the right.
