11月9日至18日期间,Anthropic 发布了 Claude Fable 5.2,Google 推出了 Gemini 4 Carbon、Hydrogen 和 Neon,OpenAI 则发布了 GPT-6.2 Astra 和 Sol——三家实验室在十天内推出六款模型。每次发布都在基准测试中匹敌或超越竞争对手,同时将 API 成本降低 20–50%。实际结果是:每项已完成 AI 任务的成本被大幅压缩,以至于过去只有拥有六位数 AI 预算的企业才能使用的智能体工作流,如今连 20 人的公司也用得起。
三家实验室在十天内推出六款前沿及中端模型,每款都降价或提升了效率——Fable 5.2 在智能基准测试中领先,Neon 在单任务成本效率上领先,而 Sol 以 15% 的成本达到了上一代旗舰的水平。
根据模型和努力程度的不同,如今一个典型的智能体业务工作流每项已完成任务的成本在 0.18 美元到 4.00 美元之间——而六个月前,达到同等质量需要花费数百美元。
埃森哲对 750 名高管的调查发现,95% 的人会选择扩大 AI 应用场景,而不是将更便宜的 token 省下的钱存起来——这场价格战是采用的催化剂,而非降低成本的手段。
这个时间点并非巧合。Anthropic 于 11 月 9 日率先出手,Claude Fable 5.2 的定价为每百万输入 token 8 美元、每百万输出 token 40 美元——较 Fable 5.1 降价 20%,输出生成速度最高提升 30%。Google 紧随其后,次日推出 Gemini 4 Carbon,随后于 11 月 12 日推出 Hydrogen,并于 11 月 18 日推出其效率最高的前沿模型 Gemini 4 Neon,定价为每百万 token 输入 3 美元、输出 15 美元。OpenAI 于 11 月 17 日发布 GPT-6.2 Astra 和 Sol,Astra 的定价比前代低 20%,为 8/40 美元,缓存输入可享 60% 折扣。
在 Artificial Analysis 智能指数 v5——一个涵盖知识工作、编程、科学和通用能力的独立综合指标——中,截至 11 月 21 日的排名为:Claude Fable 5.2 以 64.6 分领先,GPT-6.2 Astra 以 62.1 分位列第二,Gemini 4 Neon 以 61.4 分位列第三,Gemini 4 Carbon 以 60.0 分位列第四。第一名与第四名之间的差距不到 5 分。六个月前,这一差距接近 15 分。
据 Artificial Analysis 称,GPT-6.2 Sol 在智能指数上得分 57——与上一代旗舰 GPT-6 Astra 持平——而每项任务的成本仅约为其 15%。这是一个前沿级模型,每个基准任务仅需 1.01 美元。
标价只是故事的一半。智能体工作流不会只发起一次 API 调用——它会规划、调用工具、重新读取不断增长的上下文,并在失败时重试。ProcessRaven 对 78 个常见业务工作流的基准测试发现,在均衡级模型上,智能体结果的中位成本为 0.47 美元,从简单检查的 0.004 美元到 17 步应付账款流程的 3.85 美元不等。这相当于单条聊天消息成本中位数的 20 倍。
这一代产品的悖论在于,更便宜的 token 反而可能产生更昂贵的任务。Artificial Analysis 测得 Claude Fable 5.2 在最大努力下每个智能指数任务的成本为 26.54 美元——大约是 Fable 5.1 的 12.99 美元的两倍——因为当被允许更深入思考时,新模型会多产生 69% 的输出 token。这个模型更聪明,但也更“贪吃”。
这正是努力程度设置和模型路由成为真正杠杆的地方。正如科技分析师 Dan Worth 所指出的,“一个只有在 2/10 美元时才觉得值得的循环,是一个有截止日期的试验”——这指的是 Gemini 4 Carbon 的首发定价,该定价将于 12 月 1 日翻倍至 4/20 美元。相比之下,Google 的 Neon 每项任务成本比 Carbon 低 32%,同时在智能指数上得分更高,而且其定价是永久性的。对于构建生产工作流的企业来说,永久价目表比促销价目表更重要。
这对中型企业究竟意味着什么
智能体 AI 的商业理由一直受到两个限制:模型不够可靠,无法在无监督情况下运行;token 成本使多步骤工作流的费用高得令人望而却步。11 月的发布实质性削弱了第二个限制。当 OpenAI 的中端 GPT-6.2 Sol 在低努力下每项任务成本为 0.18 美元、在最大努力下为 1.01 美元,而 Gemini 4 Neon 以每百万 token 3 美元提供前沿智能并拥有百万 token 的上下文窗口时,像销售线索资格筛选、发票处理和客户支持分流这类工作流的成本算式就改变了。
但如果你经营的是一家 10–50 人的公司,风险并不在于每个 token 多付钱——而在于杰文斯悖论(Jevons paradox)。埃森哲 2026 年 7 月对 750 名高级高管的调查发现,如果 token 价格下降 25%,95% 的组织会选择扩大 AI 应用场景,而不是将节省的钱存起来。接受调查的公司预计,未来 24 个月 token 消耗量将增长 78%,而每 token 价格仅下降 19%。账单会上升,而不是下降。智能体工作流目前仅占企业 token 支出的 11%,但增长最快,因为每一次智能体请求都会针对不断扩大的上下文触发多次模型调用。
麦肯锡的分析进一步印证了这一点:在银行客户服务中,token 成本仅占可变运营成本的 20–25%,而人工监督占 70–75%。AI 智能体中最昂贵的部分不是模型——而是检查模型工作的人。随着可靠性提高,这一比例会发生变化,但就目前而言,建议很直接:从高容量、低风险的工作流入手,在这些场景中你可以摊销固定成本并容忍偶尔的错误。
在 Heygentic,我们一直在密切追踪这种成本压缩。我们在 GPT-6.1 Terra 上看到的使用计算机成本降低 9 倍的趋势已经加速。当 Amazon 通过路由到更小的决策模型展示出 98% 的成本削减时,它预示着这一发展方向:不是用一个前沿模型包打天下,而是在一组与任务复杂度相匹配的模型栈之间进行智能路由。11 月的发布为每家企业提供了更多可路由的层级。
Gemini 定价将于 12 月 1 日重置。Carbon 和 Argon 从 2/10 美元翻倍至 4/20 美元。任何建立在其首发价格基础上的工作流现在都需要重新定价。Neon 的 3/15 美元没有同样的附加条件,可能成为 Google 生产级智能体工作的默认模型。
模型路由正在成为一项必备准则,而非锦上添花。在同一工作流中,最便宜与最昂贵模型之间 100 倍的中位成本差距——以及更便宜的模型失败和重试更多的事实——意味着路由决策如今是杠杆作用最大的单一成本杠杆。Gartner 警告称,只有不到 10% 的智能体 AI 部署实现了稳健的投资回报,而且由于推理模型消耗的 token 比基础聊天机器人多 5–30 倍,预计到 2028 年每个工作流的推理成本将增长五倍以上。
开放权重模型的差距正在扩大。Artificial Analysis 智能指数目前显示,最佳专有模型(Fable 5.2,64.6 分)与最佳开放权重模型(Kimi K3.1 Pro,46.5 分)之间相差 12 分。在完全开放权重的模型中,差距为 32 分。对于希望通过自托管来控制成本的企业来说,这比六个月前是更艰难的权衡。
价格战尚未结束——它只是变得更加复杂。赢家不会是那些选择最便宜模型的企业,而是那些构建出足够智能的工作流、能够每一次都为每个步骤挑选合适模型的企业。
OpenAI 在智能指数中拥有排名第二的模型——Artificial Analysis
AI 模型:价格正在下降,但某些任务成本更高——The AI News
智能体 AI 的下一个障碍是成本,而非能力——Gartner
本文来源:https://heygentic.com.au/news/november-ai-price-war-three-labs-agentic-workflows-viable
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
Between November 9 and 18, Anthropic shipped Claude Fable 5.2 , Google released Gemini 4 Carbon, Hydrogen, and Neon , and OpenAI launched GPT-6.2 Astra and Sol — six models from three labs in ten days. Each release matched or beat rivals on benchmarks while cutting API costs by 20–50%. The practical result: the cost per completed AI task has compressed so dramatically that agentic workflows previously reserved for enterprises with six-figure AI budgets are now within reach of a 20-person company.
Six frontier and mid-tier models from three labs in ten days, each cutting prices or boosting efficiency — Fable 5.2 leads intelligence benchmarks, Neon leads cost-per-task efficiency, and Sol matches last generation's flagship at 15% of the cost.
A typical agentic business workflow now costs between $0.18 and $4.00 per completed task depending on model and effort level — down from hundreds of dollars for comparable quality six months ago.
Accenture's survey of 750 executives found 95% would expand AI use cases rather than bank savings from cheaper tokens — this price war is an adoption accelerant, not a cost reducer.
The timing was not coincidental. Anthropic moved first on November 9 with Claude Fable 5.2 at $8 per million input tokens and $40 per million output tokens — a 20% cut from Fable 5.1 , with outputs generated up to 30% faster. Google followed the next day with Gemini 4 Carbon, then Hydrogen on November 12, and its most efficient frontier model Gemini 4 Neon on November 18 at $3 input and $15 output per million tokens . OpenAI landed on November 17 with GPT-6.2 Astra and Sol, pricing Astra 20% below its predecessor at $8/$40 with a 60% discount on cached input.
On the Artificial Analysis Intelligence Index v5 — an independent composite covering knowledge work, coding, science, and general capability — the rankings as of November 21 are: Claude Fable 5.2 leads at 64.6, GPT-6.2 Astra second at 62.1, Gemini 4 Neon third at 61.4, Gemini 4 Carbon fourth at 60.0. The gap between first and fourth place is less than five points. Six months ago, that spread was closer to fifteen.
GPT-6.2 Sol scores 57 on the Intelligence Index — matching last generation's flagship GPT-6 Astra — at roughly 15% of the cost per task, according to Artificial Analysis. A frontier-grade model for $1.01 per benchmark task.
The sticker prices are only half the story. An agentic workflow doesn't make one API call — it plans, calls tools, re-reads a growing context, and retries on failure. ProcessRaven's benchmark of 78 common business workflows found the median agentic outcome costs $0.47 on a balanced-tier model, ranging from $0.004 for a simple check to $3.85 for a 17-step accounts payable process. That's a median of 20 times the cost of a single chat message.
The paradox of this generation is that cheaper tokens can produce more expensive tasks. Artificial Analysis measured Claude Fable 5.2 at $26.54 per Intelligence Index task at maximum effort — roughly double Fable 5.1's $12.99 — because the new model produces 69% more output tokens when given permission to think harder. The model is smarter, but it's also hungrier.
This is where effort settings and model routing become the real levers. As tech analyst Dan Worth noted , "a loop that only feels worth it at $2/$10 is a trial with a date on it" — a reference to Gemini 4 Carbon's introductory pricing, which doubles to $4/$20 on December 1. Google's Neon, by contrast, delivered a 32% lower cost per task than Carbon while scoring higher on the Intelligence Index, and its pricing is permanent. For businesses building production workflows, the permanent rate card matters more than the promotional one.
What this actually means for a mid-sized business
The business case for agentic AI has always been limited by two constraints: the models weren't reliable enough to run unsupervised, and the token costs made multi-step workflows prohibitively expensive. November's releases have materially weakened the second constraint. When OpenAI's mid-tier GPT-6.2 Sol costs $0.18 per task at low effort and $1.01 at maximum , and Gemini 4 Neon delivers frontier intelligence at $3 per million tokens with a one-million-token context window, the arithmetic changes for workflows like lead qualification, invoice processing, and customer support triage.
But if you're running a 10–50 person operation, the risk isn't overpaying per token — it's the Jevons paradox . Accenture's July 2026 survey of 750 senior executives found that 95% of organisations would expand AI use cases rather than bank savings if token prices fell 25%. Companies in the survey expect token consumption to grow 78% over the next 24 months while per-token prices decline just 19%. The bill goes up, not down. Agentic workflows, currently just 11% of enterprise token spend , are growing fastest because each agent request triggers multiple model calls against an expanding context.
McKinsey's analysis reinforces the point: in banking customer service, token costs represent just 20–25% of variable run costs, while human oversight accounts for 70–75%. The most expensive part of an AI agent isn't the model — it's the person checking the model's work. That ratio will shift as reliability improves, but for now, the advice is straightforward: start with high-volume, low-stakes workflows where you can amortise fixed costs and tolerate occasional errors.
At Heygentic, we've been tracking this compression closely. The pattern we saw with GPT-6.1 Terra making computer use 9x cheaper has accelerated. When Amazon demonstrated a 98% cost cut by routing to a smaller decision model , it signalled where this is heading: not a single frontier model for everything, but intelligent routing across a stack of models matched to task complexity. November's releases have given every business more tiers to route across.
Gemini pricing resets on December 1. Carbon and Argon double from $2/$10 to $4/$20. Any workflow built on their introductory rates needs repricing now. Neon at $3/$15 doesn't carry the same asterisk and may become the default Google model for production agentic work.
Model routing is becoming a required discipline, not a nice-to-have. The 100x median cost spread between cheapest and dearest model on the same workflow — and the fact that cheaper models fail and retry more — means the routing decision is now the single highest-leverage cost lever. Gartner warns that fewer than 10% of agentic AI deployments are delivering robust ROI, and inference costs per workflow are predicted to increase more than fivefold through 2028 as reasoning models consume 5–30x more tokens than basic chatbots.
The open-weights gap is widening. The Artificial Analysis Intelligence Index now shows a 12-point gap between the best proprietary model (Fable 5.2 at 64.6) and the best open-weights model (Kimi K3.1 Pro at 46.5). Among fully released weights, the gap is 32 points. For businesses hoping to self-host for cost control, that's a harder trade-off than it was six months ago.
The price war isn't over — it's just becoming more sophisticated. The winners won't be the businesses that pick the cheapest model. They'll be the ones that build workflows smart enough to pick the right model for each step, every time.
OpenAI has the #2 model in the Intelligence Index — Artificial Analysis
AI Models: Prices Are Falling, But Some Tasks Cost More — The AI News
Agentic AI's Next Hurdle Is Cost, Not Capability — Gartner
本文来源:https://heygentic.com.au/news/november-ai-price-war-three-labs-agentic-workflows-viable
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
