一个模型拿了总榜第一，落到你的账单上却未必便宜。2026 年 10 月 29 日，Artificial Analysis 公布 Gemini 4 Carbon 登顶 Intelligence Index v5，得分 59。同一份报告里还写着另一个数字：它每个任务的成本是 7.76 美元，比 Claude Opus 5.5 便宜约 23%，却比 GPT-6 Astra 贵出约两成。第一名、最便宜、最适合你，是三件不同的事。
真正值得抄的不是结论，而是这份报告拆指标的思路。它把一次模型选型拆成三条线：能力分够不够、每任务烧多少钱、可用性到不到位。下面按这三条线走一遍，顺手把可以落进代码的部分写成脚本。
一、榜单换了尺子，59 分和旧榜的 59 分不是一回事
Intelligence Index v5 是一次换代，不是小版本更新。Carbon 的 59 分是在九项评测上加权出来的，它在其中四项拿到第一：HLE-Diamond 考复杂推理，Terminal-Bench-Science 考科学场景下的终端操作，AA-Omniscience 考知识边界与幻觉控制，AutomationBench-AA 考自动化任务。
这份报告最诚实的地方在于它同时说了 Carbon 没赢的地方。报告原文写得很直白，Carbon 在编程和知识工作上都有明显进步，但并没有领先。这意味着一件事：如果你的场景是写代码，59 分这个总分的说服力要打折，得去看分项。
还有个更隐蔽的坑。同一批模型的旧分数不能直接跨版本比较。指数换了评测集和权重，上一代 58 分和这一代 58 分代表的水平并不相同。团队内部如果还留着半年前那张对比截图，结论很可能已经过期。榜单的价值在于当下的相对位置，不在于历史曲线。
二、九项评测里，只有一两项对得上你手里的活
把评测项和你手上的任务对齐，比记总分有用得多。Carbon 领先的四项，翻译成工程语言大致是这样。
Coding Agent Index 这一栏值得单独说。Carbon 配合 Antigravity CLI 拿到 59 分，比 Claude Code 配 Claude Opus 5.5 的 62 分只差 3 分。分项上它排第二：Terminal-Bench 4.0 拿到 64%，FrontierCode v1.1 拿到 51%，两项都落后于 Opus 5.5 的 56%。Google 在这个方向历史上一直偏弱，这次的追赶幅度确实明显，但差距没有消失。
知识工作上也是同一套读法。Carbon 在 AA-Briefcase 拿到 1609 Elo，比上一代 Argon 的 1483 高出 126 分，进步集中在演示环节，单项涨了 293 Elo 到 1598，分析质量涨了 120 分到 1671。但它仍然排在 Grok 4.7 的 1638、Muse Wonder 的 1659、Kimi K3.1 Pro 的 1678 之后。如果你的活是给客户出方案，这一栏比总分重要。
分项对齐这件事没有捷径，只能拿你自己的历史任务去套。把过去一个月的真实请求按类型打标，再看每类任务落在哪一项评测上，通常二十条样本就够看出方向。没有样本就凭总分投票，等于把选型的责任交给了榜单编辑。
三、准确率、幻觉率、每任务成本必须一起看
Carbon 在 AA-Omniscience v1.1 拿到 54 分，比 Claude Opus 5.5 的 48 分高 6 分，比 Argon 的 46 分高 8 分。准确率从 Argon 的 48% 涨到 58%，涨了 10 个百分点。
与此同时它的幻觉率从 4% 涨到 11%。单看这条会觉得退步了，放到同类里看才清楚：报告明确指出这仍是所有模型里第二低的水平，只比自家 Argon 高，而 Claude Opus 5.5 和 GPT-6 Astra 两家的幻觉率都是 50%。
这个对比暴露了一个常见误读。很多团队把幻觉率当绝对值看，看到从 4% 变 11% 就判定模型变差。真正该算的是错误成本。50% 的幻觉率意味着每两次回答就有一次在编，这个量级在客服、配置修改、法务摘要这类场景里根本不能用，哪怕它的语气更自信。11% 则需要配上校验环节，比如让检索结果回填引用、让关键数字来自可核验来源。
成本那一栏更容易被忽略。Carbon 每个任务的标准价是 7.76 美元，Opus 5.5 是 10.04 美元，GPT-6 Astra 是 6.25 美元。看起来 Carbon 比 Astra 贵，但报告给了一个解释：Carbon 每个任务只用 123k 输出 token，比 Opus 5.5 的 173k 少 29%。贵和便宜不能只看单价，得看完成同一个任务要吐多少字。
这三条放在一起看，评测分数的含义就清晰了：总分说明能力上限，幻觉率说明风险下限，每任务成本说明可行性。任何一项不达标，另外两项再漂亮也不该进生产。采购讨论里最常见的错误，是拿上限去论证可行性。
四、把官方价目表换算成你自己的单位成本
官方标价是 4 美元每百万输入 token、20 美元每百万输出 token，和上一代 Argon 的标准价一致，另外还有一个 50% 的上线折扣，报告明确说这个折扣没有确定的结束日期。缓存命中的输入打 5% 折扣，折算下来是每百万 0.20 美元。
这三个数足够反推一个任务到底花多少。下面这段脚本按任务画像估算单次成本，输入侧拆出缓存命中比例，输出侧按你自己实测的 token 数填。跑通之后你会得到一张自己的价目表，而不是抄报告里的数字。
PRICE_IN = 4.0 / 1_000_000
PRICE_OUT = 20.0 / 1_000_000
PRICE_CACHE = 0.20 / 1_000_000
def task_cost(prompt_tokens, output_tokens, cache_ratio=0.0, retries=1.0):
cached = prompt_tokens * cache_ratio
fresh = prompt_tokens - cached
per_call = fresh * PRICE_IN + cached * PRICE_CACHE + output_tokens * PRICE_OUT
return per_call * retries
def compare(profiles):
rows = []
for name, p in profiles.items():
cost = task_cost(p["prompt"], p["output"], p.get("cache", 0.0), p.get("retries", 1.0))
rows.append((name, round(cost, 4)))
rows.sort(key=lambda r: r[1])
for name, cost in rows:
print(f"{name:24s} {cost:>8.4f} USD/task")
profiles = {
"gemini-4-carbon": {"prompt": 90_000, "output": 123_000, "cache": 0.7, "retries": 1.1},
"claude-opus-5.5": {"prompt": 90_000, "output": 173_000, "cache": 0.7, "retries": 1.2},
"gpt-6-astra": {"prompt": 90_000, "output": 95_000, "cache": 0.5, "retries": 1.6},
}
compare(profiles)
脚本里 retries 那一栏是很多人漏掉的一环。模型失败一次就得重试，重试的输入通常还得带上前一轮的上下文，真实成本会偏离标价。上面 Astra 的 retries 给到 1.6，是因为低幻觉率缺失时，人工复核和重跑都会被算进单位成本。
跑出来的结果通常会颠覆直觉：单价最贵的那家，可能不是每任务最贵的。真正决定账单的是输出长度乘以失败率，而不是输入单价那几毛钱。
缓存那一项也值得专门抠一次。报告里缓存输入折扣到每百万 0.20 美元，相对 4 美元的标准价是 95% 的折扣。多轮对话和长系统提示的场景里，命中率从七成提到九成，省下来的钱往往比换一家模型更多。这类优化不改变任何能力指标，却是性价比最直接的一条路。
五、四个模型横向摆开
把报告里的数字放进一张表，方便直接对照。
看这张表的方式是先把两处空白放一边，剩下的行里只有幻觉率这一项差距最大。一个 11% 对一个 50%，这不是优化空间，是场景准入线。
第二层看碳与氩的对比。同价，分数高 9 分，幻觉率高 7 个百分点。如果任务是给内部工程师查资料，多出来的幻觉可以用代码和日志当场证伪，选高分那版更划算。如果任务是给外部客户输出结论，低温的那版更稳。
第三层才是看价格标签。7.76 和 7.68 之间几乎没有区别，真正的差距发生在 Opus 的 10.04 上，而它换回来的是编程方向多出的那 3 分和 56% 对 51% 的分项优势。如果你的场景就是改仓库，这 30% 的溢价可能值得付。
六、能不能真的调到，是榜单不会写的指标
报告里有一句容易被忽略的话：Carbon 正在向部分用户灰度，尚未公开可用。前面那些分数，包括 59 分总分和 51% 的 FrontierCode 成绩，都是用高推理档位、配合特定 CLI 跑出来的。
这一条在 Hacker News 的讨论里被反复吐槽。有用户提到自己走 Pro 订阅用上一代 Argon，中等档位发五条消息就触发四小时冷却。还有人贴了自己的复现经验，说 Argon 在读取文件名后凭空编出一整段根本不存在的对话，内容离谱到像小说。同一天有人补了一句股价，说这轮发布之后股价跌了 13%。这些是用户视角和资本市场视角，不一定代表技术判断，但它们指向同一个现实：发布会和可用性是两件事。
这些讨论里还有一条更实用的信息。有人提到自己通过非官方客户端用上了上一代模型，并把它的修复效果和别家做了对比，结论是差别真实存在，但需要自己搭一层客户端才能拿到。这类折腾对小团队来说成本不低，因为它把维护负担从供应商转移到了自己身上。
对工程团队来说，这条应该变成一个明确的决策规则：不要基于一个你还调不到的模型做架构迁移。灰度模型适合放进评估池，按月重测一次；生产流量必须跑在已经稳定可用、有明确 SLA 和配额承诺的版本上。把评测分数当成采购依据没问题，把评测分数当成今天就能上线的依据就是事故。
七、把选型做成路由，而不是做成一次投票
选型投票的典型结局是团队吵一个月，收尾时统一选一家，然后在遇到瓶颈时无路可退。更耐用的做法是做路由，让请求按任务类型分派。
2026 年 7 月的 Copilot 更新里有一个可以参考的做法。Copilot CLI 上线了自动模型选择，简单查询不会占用复杂多步任务那份算力预算。同一批更新还给了会话级的 AI 额度上限，可以从 CLI 和 SDK 两侧配置。这两件事合起来就是一套最小可用的成本治理：路由负责把便宜的活交给便宜的模型，额度闸门负责兜住失控的长任务。
同月还有个信号值得记一笔，Kimi K2.7 Code 进了 Copilot 的模型选择器，是第一个可以在里面选的开源权重模型。开源权重意味着推理侧可以自建，单位成本和数据边界都能自己控。对预算敏感或者有合规要求的团队，这条路比再等一个灰度模型现实得多。
顺带一个必须提前排期的动作。GitHub Models 在 2026 年 7 月 30 日退役，六月份就已经停止接纳新客户。任何把业务流程挂在这个端点上、又没有做抽象层的系统，都会在退役当天被迫临时改代码。这也正好说明为什么路由层要抽象：换模型应该是改配置，不是改业务逻辑。
八、今天能做的四件事
第一件，把你能调到的模型列成一张表，只填三列：能不能调、每任务实测成本、失败率。三列齐了，选型会自己浮出来。
第二件，用上面的脚本把官方价目表换成自己的报价。输入 token 用真实日志统计，别用估算。缓存命中比例从账单里取，这一项经常被白送掉三成成本。
第三件，给每类任务定一条准入线。比如面向外部客户的输出必须低于 15% 幻觉率，编程类任务可以放宽，但必须带测试回填。准入线一旦写下来，模型换代时就不需要重新吵架。
第四件，把灰度模型丢进评估池，按月度重测，不要进生产。评估池不需要复杂的平台，一个能把任务重放、记录成本和失败率的脚本就够了，关键是让每次换代都有可比数据。分数会变，评测集也会变，今天的第一名三个月后可能连前十都进不去，而你的系统不该跟着榜单抖动。
排行榜解决的是信息不对称，不解决决策。Carbon 拿 59 分是真的，它每任务 7.76 美元也是真的，它现在多数人还调不到同样是真的。把这三句同时写进决策文档，选型才算是做完了。
MODEL_POLICY = {
"cheap_qa": {"model": "gemini-4-carbon", "max_output": 4_000, "retry": 1},
"code_edit": {"model": "claude-opus-5.5", "max_output": 32_000, "retry": 2},
"client_report":{"model": "gemini-4-argon", "max_output": 8_000, "retry": 1},
}
def pick(task_type, budget_left):
rule = MODEL_POLICY[task_type]
if budget_left < rule["max_output"] * PRICE_OUT:
raise SystemExit("额度不足，拒绝下发高价任务: " + task_type)
return rule
这段策略表小到不值得叫架构，但它拦住的问题很实际：额度见底时，最贵的那个任务往往还在排队等着跑，而它恰好是最不该跑的那一个。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
A model gets first on the list, but it's not cheap on your bill. On October 29, 2026, Artificial Analysis announced Gemini 4 Carbon top of Intelligence Index v5, score 59. The same report contained another figure: it cost $7.76 per mission, about 23 percent less than Claude Opus 5.5, but about 20 percent more than GPT-6 Astra. First, the cheapest, the best for you, is three different things.
It is not the conclusions that are truly worthy of being copied, but rather the thinking behind the dismantling of the indicators in this report. It splits the model selection into three lines: sufficient capability, how much per mission burned, and availability. Following these three lines, the parts that can be put into the code are written in scripts.
One, the list changed the ruler, 59 points and 59 points from the old list. Yes.
Inteligence Index v5 is a change, not a small update. Carbon ' s 59 scores were weighted on nine assessments, and in four of them it received the first: HLE-Diamund ' s complex reasoning, Terminal-Bench-Science ' s terminal operation in the scientific scene, AA-Omniscence ' s knowledge boundary and hallucination control, Automation Bench-AAA mission.
The most honest part of this report is where it says that carbon didn't win. The original report was very clear, and Carbon had made significant progress in programming and knowledge, but not leading. And that means one thing: if your scene is a code, 59 is a discount for the whole score, and you have to see the sub-point.
There's a more hidden pit. The old scores of the same models cannot be directly compared across the versions. The index changed the rating and weight, with the previous generation 58 points different from the levels represented by the current generation 58. The conclusions are likely to be outdated if the team still has a comparison screenshot of six months ago. The value of the list lies in its current relative position, not in the historical curve.
In two or nine evaluations, only one or two of them will get what you've got.
Matching the evaluation with the tasks you have is much more useful than scoring the total. The four leading Carbon languages, translated into engineering languages, are roughly this.
It's worth a separate column. Carbon co-ordinates Antigravity CLI with 59 points, which is only 3 points less than the 62 points of Claude Code with Claude Opus 5.5. Second in line: Terminal-Bench 4.0, 64%, FrontierCode v1.1 51%, both 56% behind Opus 5.5. Google has been historically weak in this direction, and it is true that the catch-up was clear this time, but the gap did not disappear.
The same reading is done in intellectual work. Carbon received 1609 Elo in AA-Briefcase, 126 points higher than the previous generation of Argon, and progress was concentrated on the demonstration, with the single item increasing by 293 Elo to 1598 and the analytical quality increasing by 120 to 1671. However, it still ranks behind 1638 in Grok 4.7, 1659 in Muse Wonder and 1678 in Kimi K3.1 Pro. If your job is to plan for clients, this column is more important than the total score.
There are no shortcuts to aligning sub-sections, but only your own historical mission. When the true requests of the past month are marked by type, depending on which assessment each type of task falls, it is usually enough to see the direction of 20 samples. Without sample, the vote was based on the total score, which was tantamount to leaving the responsibility for the selection to the editor of the list.
III. Accuracy rate, hallucinogenic rate, cost per mission Look.
Carbon won 54 points in AA-Omniscence v1.1, 6 points higher than 48 points for Claude Opus 5.5 and 8 points higher than 46 points for Argon. The accuracy rate rose from 48% in Argon to 58% and by 10 percentage points.
At the same time, it's hallucinating from 4% to 11%. Seeing this one alone, it's clear that it's still the second lowest level in all models, only higher than the home, Argon, while Claude Opus 5.5 and GPT-6 Astra have a 50% hallucination rate.
This comparison exposes a common misreading. A lot of teams consider hallucinogenic rates to be absolute, and when they see a change from 4% to 11%, they decide that the model is changing. What really counts is the wrong cost. The 50 per cent hallucination rate means that every two answers are being compiled, and this scale is simply not working in such scenarios as customer service, configuration modification, law enforcement summary, even if it speaks more confidently. For 11 per cent of them, there is a need for a Colonel ' s examination, for example, to return the results of the search and to have key figures from verifiable sources.
The cost column is easier to ignore. The standard price per mission is $7.76, Opus 5.5 is $10.04 and GPT-6 Astra $6.25. Looks like carbon is more expensive than Astra, but the report explains that carbon only uses 123k output token, 29% less than the 173k of Opus 5.5. Prices and discounts cannot be seen only in unit prices, but in terms of the amount of words to be spitted for the same task.
Taken together, the meaning of the score is clear: the total points indicate the capacity ceiling, the hallucinogenic rate indicates the risk threshold, and the cost of each mission indicates the feasibility. None of them meet the standards and the other two are not supposed to be productive. The most common error in the procurement discussions was the taking of a ceiling to justify feasibility.
IV. Making the official price list convert your own unit cost
The official bid is $4 per million input token, $20 per million output token, the same standard price as the previous generation of Argon, and a 50 per cent offline discount, which the report clearly states has no fixed end date. cache hit entered a 5% discount, converted to $0.20 per million.
These three numbers are enough to reverse the exact amount of a task. The following scripts estimate the single cost per mission image, enter a side to remove the cache hit ratio, and the output side to be filled in by your own actual token number. After running, you get your own price list instead of copying the numbers in the report.
PRICE_IN = 4.0 / 1_000_000
PRICE_OUT = 20.0 / 1_000_000
PRICE_CACHE = 0.20 / 1_000_000
def task_cost(prompt_tokens, output_tokens, cache_ratio=0.0, retries=1.0):
cached = prompt_tokens * cache_ratio
fresh = prompt_tokens - cached
per_call = fresh * PRICE_IN + cached * PRICE_CACHE + output_tokens * PRICE_OUT
return per_call * retries
def compare(profiles):
rows = []
for name, p in profiles.items():
cost = task_cost(p["prompt"], p["output"], p.get("cache", 0.0), p.get("retries", 1.0))
rows.append((name, round(cost, 4)))
rows.sort(key=lambda r: r[1])
for name, cost in rows:
print(f"{name:24s} {cost:>8.4f} USD/task")
profiles = {
"gemini-4-carbon": {"prompt": 90_000, "output": 123_000, "cache": 0.7, "retries": 1.1},
"claude-opus-5.5": {"prompt": 90_000, "output": 173_000, "cache": 0.7, "retries": 1.2},
"gpt-6-astra": {"prompt": 90_000, "output": 95_000, "cache": 0.5, "retries": 1.6},
}
compare(profiles)
In the scripts, retries are a loop that many people missed. Once the model fails, retry, retry's input is usually accompanied by the previous round of context, and the real cost deviates from the price. Astra's retries above are given 1.6 because manual review and rerun are factored into unit costs when the low hallucinogenic rate is missing.
The result of coming out is usually a reversal of intuition: the one that costs the most, probably not every mission. What really determines the bill is the output length multiplied by the failure rate, not the amount of the unit price.
cache also deserves a special one. The cache input discount in the report was $0.20 per million, and the standard price for $4 was 95 per cent. In the scenes of multiple rounds of dialogue and long-system alerts, the hit rate was 90 per cent from 70 per cent, often saving more money than a different model. Such optimization does not change any capacity indicators, but is the most direct route to value for money.
Five, four models horizontally
The figures in the report are placed in a table to facilitate direct comparison.
The way to look at this table is to put the two blanks aside for the first time, and the remaining rows have the largest gap between hallucinogenic rates. One 11% for one 50%, it's not perfect space, it's scene access.
The second layer looks at the contrast between carbon and argon. At the same price, the score was 9 points and the hallucinogenic rate was 7 percentage points. If the task is to check the information for internal engineers, the additional illusions can be falsely documented on the spot with codes and logbooks, making the selection of the high score more cost-effective. If the task is to export conclusions to external clients, the cooler version is more stable.
The third level is the price label. There is little difference between 7.76 and 7.68, and the real difference is on Opus 's 10.04, which is the three-point and 56 per cent sub-advantage of 51 per cent. If your scene is to change the warehouse, this 30% premium may be worth paying.
Six, it's an indicator you won't write.
There is one easily overlooked phrase in the report: Carbon is greyscale to some users and is not yet publicly available. The previous scores, including 59 points and 51 per cent of the Frontier Code scores, ran with high-intensity slots and matched specific CLIs.
This article was repeatedly criticized in Hacker News discussions. Some users mention that they went Pro subscriptions to the previous generation, Argon, and five messages from the medium slot trigger four hours cooling. Others posted their own replications, saying that Argon, reading a file name, created an entire conversation that simply didn't exist, and that it was completely off-limits to a novel. On the same day someone added a 13 per cent drop in the stock price after the round of publish. These are user and capital market perspectives and do not necessarily represent technical judgements, but they point to the same reality: publish and usability are two things.
These discussions also included a more practical message. It was mentioned that it had used a previous generation model through an unofficial client and compared its restoration effects to others, concluding that the difference was real, but that it needed a single client to get it. Such manoeuvres are not less costly to small teams because they shift the maintenance burden from suppliers to themselves.
For the engineering team, this should become a clear rule of decision-making: no structural migration based on a model that you can't reach. The greyscale model is suitable for placement in the assessment pool and is recalculated once a month; production flows must run in a version that is already stable and has clear SLA and quota commitments. It is no problem to use scores as a basis for procurement, and to use scores as a basis for getting on line today is an accident.
VII. Quoting instead of a vote
The typical outcome of the voting was a one-month team fight, a single single family at the end and no way out when there were bottlenecks. A more durable approach would be to route requests by type of task.
A reference is included in the July 2026 Copilot update. Copilot CLI has an automatic model selection online, and simple queries do not take up the budget for complex multi-step missions. The same updates were returned to the limit of AI quota for the session level, which can be configured from the CLI and SDK sides. Taken together, these two things are a minimum set of cost management: routers are responsible for handing over cheap jobs to cheap models, and quota gates are responsible for running out of control.
The same month, Kimi K2.7 Code entered the Copilot Model Selector, the first open source weight model to choose from. Open source weights mean that the side of reasoning can be built on its own and that unit costs and data boundaries can be controlled on their own. This route is much more realistic than waiting for another greyscale model for teams that are budget sensitive or subject to compliance requirements.
Attach an action that must be scheduled earlier. GitHub Models retired from service on 30 July 2026 and stopped receiving new clients in June. Any system that hangs business processes at this end without making abstract layers will be forced to temporarily change codes on the day of retirement. This explains why the route floor is abstract: The model should be reconfigured, not changed business logic.
Eight, four things to do today.
First, the model that you can move is presented in a table with three columns: can it be adjusted, measured cost per mission, failure rate? It's all three. The selection will float.
In the second, the official price list is replaced by a script above. Enter token, do not estimate. The cache hit rate was taken from the bill, and this item was often dropped by 30 per cent.
Third, there is an access line for each category of mission. For example, the output to external clients must be below 15% hallucinogenic, programming tasks can be relaxed, but must be filled back with tests. Once the access line is written down, the model does not need to fight again.
Fourth, the greyscale model is thrown into the assessment pool and remeasured on a monthly basis to avoid production. The assessment pool does not require a complex platform, but a script that can reset missions, record costs and failure rates is sufficient, and it is critical that comparable data be available for each change. The scores will change, the ratings will change, and the top 10 may not be able to get in after the first three months today, and your system should not be shaking with the list.
The rankings address information asymmetries and do not address decision-making. Carbon takes 59 points, and it's true for every mission, and it's true for most people now. The three sentences are written into the decision-making document at the same time, and the selection is done.
MODEL_POLICY = {
"cheap_qa": {"model": "gemini-4-carbon", "max_output": 4_000, "retry": 1},
"code_edit": {"model": "claude-opus-5.5", "max_output": 32_000, "retry": 2},
"client_report":{"model": "gemini-4-argon", "max_output": 8_000, "retry": 1},
}
def pick(task_type, budget_left):
rule = MODEL_POLICY[task_type]
if budget_left < rule["max_output"] * PRICE_OUT:
Rice SystemsExit ("quota is inadequate and refused a high-price job" + task type)
return rule
This strategy sheet is so small that it is not worth calling a structure, but it's a practical problem: When quota came to the bottom, the most expensive task was often waiting in line to run, and it happened to be the last one to run.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.