谷歌发布低成本Gemini Flash系列模型，但旗舰Gemini 3.5 Pro继续延期

# 美股盘后催化 12/12: Google发布Gemini 3.6 Flash/Flash-Lite/Cyber低成本模型，但旗舰Pro继续延期

- 美股会话: 2026-07-21
- 方向: MIXED；类型: 产品/技术；规划重大性: 7/10
- 映射模式: competitive
- 美股驱动（规划器初判，需独立核验）: 事实：Google发布三款更低成本、更高吞吐、面向agent和网络安全场景的Gemini模型，部分已接入GitHub Copilot、Databricks、Google Cloud和Gemini App；但Gemini 3.5 Pro仍处于伙伴测试，未给出明确发布时间，部分市场评论质疑其前沿能力。推断：低成本推理和agent工作流利好Google Cloud与应用客户，但旗舰延期削弱其相对OpenAI/Anthropic/Microsoft的前沿模型叙事，并可能推动模型API价格战。

## 本催化相关美股真实收盘行情

- 微软 (MSFT.US): 397.75 -1.13%
- 谷歌 (GOOGL.US): 347.15 -1.38%
- 亚马逊 (AMZN.US): 247.55 -0.98%

## 亚太映射假设（仅候选，必须查证）

- 阿里巴巴-W (9988.HK) [HK/竞争替代/同源本地股/mixed]: 低价高吞吐模型竞争可能压低Qwen/阿里云模型API价格，但也验证企业推理需求扩张。；核验问题: 核实阿里云模型API价格、Qwen性能和企业客户采用。
- 京东集团-SW (9618.HK) [HK/客户/mixed]: 模型成本下降有利于电商和企业应用落地，但也改变自研/外采模型策略。；核验问题: 核实京东AI应用对第三方模型和云算力的依赖度。
- 软银集团 (待归一) [JP/其它/mixed]: 全球模型价格和性能竞争会影响其AI投资组合和日本AI服务生态估值。；核验问题: 核实软银对OpenAI、Anthropic、Google生态或本土模型的权益敞口。

## 必答问题

- Gemini 3.5 Pro何时发布且基准是否达到预期？
- Flash系列能否转化为Google Cloud收入，而不是单纯压低行业token价格？

## 原始证据全文（22 条）

### [E0478] Alphabet IR | 2026-07-21 15:00 UTC | 盘中
Introducing Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber
URL: https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-6-flash-3-5-flash-lite-3-5-flash-cyber/
Introducing Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber
Developers and customers building production AI agents need higher token efficiency, lower latency, and more reliable performance. Our Flash series of models is built to meet the sweet spot of efficiency and quality to enable scaling agentic workflows. Building on Gemini 3.5 Flash, we’re introducing new Gemini models:
- 3.6 Flash: Our workhorse model that delivers better coding, knowledge work, and multimodal performance. According to the Artificial Analysis Index, it reduces output token usage by 17% compared to 3.5 Flash, and in some benchmarks like DeepSWE by Datacurve, we observe up to 65%, all at a lower cost per output token.
- 3.5 Flash-Lite: Our fastest, most cost-effective 3.5-class model, delivering 350 output tokens per second according to the Artificial Analysis Index, also significantly outperforming prior Flash-Lite generations in agentic workflows.
- 3.5 Flash Cyber in CodeMender: Successful cybersecurity applications require careful orchestration of a model alongside an agent infrastructure. We’re introducing a combination of a new, highly efficient, specialized cyber-focused model paired with our CodeMender code security agent that delivers competitive performance at the frontier.
Beyond today’s releases, Gemini 3.5 Pro is currently testing with partners and we plan to make it broadly available as soon as it’s ready. In parallel, our team is already focusing on building the next generation of models. We have started our most ambitious pre-training run yet, for Gemini 4, and are excited by the progress.
3.6 Flash: More efficient and better quality than 3.5 Flash
Gemini 3.6 Flash builds directly on developer and customer feedback from 3.5 Flash. 3.6 Flash not only delivers a step up in coding and knowledge work, but it does this while meaningfully improving token efficiency. For example, on the Artificial Analysis Index, we see 3.6 Flash consuming 17% fewer output tokens than 3.5 Flash. It also takes fewer reasoning steps and tool calls to accomplish multi-step workflows.
This enhanced efficiency is also combined with a lower price than 3.5 Flash. At $1.50/1M input tokens and $7.50/1M output tokens, 3.6 Flash reduces the overall cost per agentic task, making agents more cost-effective to build and run.
3.6 Flash shows better token efficiency and reduced verbosity than 3.5 Flash in an OSWorld verified task (API)
Even while being more efficient, 3.6 Flash sees performance gains compared to 3.5 Flash across use cases:
- 3.6 Flash delivers higher precision with fewer unwanted code edits and reduced execution loops, as seen in DeepSWE (49% vs. 37%), and shows significant improvement in ML Research, as seen in MLE Bench (63.9% vs. 49.7%).
- It has improved computer use capabilities as seen in OSWorld-Verified (83.0% vs. 78.4%). Computer use is now a built-in client side tool via the Gemini API and Gemini Enterprise.
- It outperforms 3.5 Flash in knowledge work, as shown by benchmarks like GDPval-AA v2 (1421 vs. 1349). Customers like Hebbia and Harvey have found it particularly capable at multimodal tasks like document parsing, chart and data analysis, and report drafting.
Customers report 3.6 Flash is a step forward in both cost and quality, balancing token efficiency, accuracy, and speed across complex workflows and knowledge-based tasks:
Built with safety
3.6 Flash is shipping with enhanced Frontier Safety safeguards in the domains of Chemical, Biological, Radiological, and Nuclear (CBRN) and cyber offense misuses. These safeguards make the model substantially more resistant to jailbreaks. At the same time, the model has been trained to minimize refusals for beneficial uses.
For more information, see the 3.6 Flash model card.
3.5 Flash-Lite: Built to scale agentic workflows
Beyond Flash, we’re also releasing Gemini 3.5 Flash-Lite, designed for both low-latency tasks and tasks where high throughput is critical for developers workflows, like agentic search and document processing.
3.5 Flash-Lite is the fastest model in the 3.5 series. As measured by Artificial Analysis, it runs at 350 output tokens/s. Priced at $0.3/1M input tokens and $2.5/1M output tokens and with significantly better quality than 3.1 Flash-Lite, 3.5 Flash-Lite offers a strong price-to-performance ratio for developers and customers running high throughput production traffic.
3.5 Flash-Lite executes high volume tasks at a lower latency than 3.5 Flash.
3.5 Flash-Lite enables efficient scaling for agentic systems. Across thinking levels, the model significantly outperforms 3.1 Flash-Lite. Depending on the workload, developers can configure the model to prioritize low-latency, low-cost execution for high-volume tasks with the minimal and low thinking levels, or engage higher thinking levels to process multi-step subagent workloads. The model now also has computer use as a built-in tool to reliably support these agentic tasks across surfaces.
It’s a significant step up in coding and agentic tasks as seen in Terminal-Bench 2.1 (54% vs 31%), long context as seen in GDM-MRCR v2 (72.2% vs. 60.1%), and real-world task execution as seen in GDPval-AA v2 (1140 vs. 642).
In fact, on many agentic and coding evals, 3.5 Flash-Lite even outperforms 3 Flash, including on SWE-Bench Pro (54.2% vs. 49.6%) and OSWorld-Verified (74.0% vs. 65.1%), making it a faster & more capable option for workloads on both 2.5 and 3 Flash.
Early customers of 3.5 Flash-Lite are highlighting its unique combination of speed, intelligence, and cost efficiency for scaling agentic workflows and data processing tasks:
For more information about the model, see the 3.5 Flash-Lite model card.
3.5 Flash Cyber in CodeMender: finding and fixing vulnerabilities efficiently
AI models have become capable of finding security vulnerabilities faster than current systems can fix them. Tackling this growing threat requires an approach to securing software that is highly capable and efficient.
Flash’s performance and efficiency makes it an ideal foundation to detect, validate, and patch code security issues at scale. Gemini 3.5 Flash Cyber is built on top of 3.5 Flash, and fine-tuned for finding and fixing cybersecurity vulnerabilities at a lower price per token than larger models.
Within CodeMender, which uses multiple 3.5 Flash Cyber agents working together to produce a single combined report, 3.5 Flash Cyber reaches competitive performance at the frontier on the popular benchmark CyberGym.
Given the dual-use nature of this technology, we have taken an intentional approach to deploying 3.5 Flash Cyber. The model will be exclusively available to governments and trusted partners via CodeMender soon as part of a limited-access pilot program. This will give frontline defenders a head start in finding and fixing critical vulnerabilities before they can be exploited, while mitigating against broader misuse.
3.6 Flash and 3.5 Flash-Lite: Get started today
3.6 Flash and 3.5 Flash-Lite are available starting today:
- For developers in the Gemini API via Google AI Studio and Android Studio. 3.6 Flash is also available in Google Antigravity. Get started with the Developer Guide.
- For enterprises in Gemini Enterprise Agent Platform. 3.6 Flash is also available in the Gemini Enterprise app.
- For everyone via the Gemini app. 3.5 Flash-Lite is also rolling out in Google Search.
As you start building with 3.6 Flash and 3.5 Flash-Lite, we welcome your feedback to improve future Gemini models and look forward to releasing 3.5 Pro soon.

### [E0491] CNBC Top News | 2026-07-21 15:00 UTC | 盘中
Google expands Gemini lineup with cheaper models and new Mythos rival
URL: https://www.cnbc.com/2026/07/21/google-gemini-flash-ai-mythos-rival.html
Alphabet is releasing three new Gemini models on Tuesday, including its clearest answer yet to Anthropic's lead in cybersecurity, as the company looks to show progress across a product pipeline that has faced delays and mounting competition.
Gemini 3.5 Flash Cyber is designed to detect and patch software vulnerabilities and will initially be available only to governments and trusted partners through a limited-access pilot. Google said the specialized model runs at a lower price per token than larger models.
That could help Google narrow its cybersecurity gap with Anthropic, which has built an early lead in automated code defense.
Google is also launching Gemini 3.6 Flash, which improves coding, multimodal and knowledge-work performance while using up to 17% fewer tokens and costing less per token than the previous model — a meaningful reduction in the cost of running high-volume workloads.
Gemini 3.5 Flash-Lite, meanwhile, is Google's fastest and least expensive model in the 3.5 family, built for high-volume workloads and smaller tasks within larger AI-agent systems.
The broader lineup reflects Google's bet that price and efficiency can help offset its slower timing in several key product categories.
Artificial Analysis data shows Gemini Flash already undercuts comparable models from Anthropic, OpenAI and Chinese rivals on cost. According to the company, Gemini 3.6 Flash — the stronger of the two new models — is cheaper per task than GPT-5.6 Terra Max, Kimi K3 and Qwen 3.7 Max, while 3.5 Flash-Lite costs just a fraction of that.
The rollout comes on the eve of Alphabet earnings and as Chinese rivals gain momentum. Moonshot AI's Kimi K3 drew enough demand that the company limited new subscriptions and API access because of capacity constraints, while Alibaba is teasing Qwen 3.8 Max, which it said trails only Anthropic's Fable 5 in overall performance.
That demand highlights the other side of the AI race: Building a competitive model is only part of the challenge. Companies also need enough computing capacity to serve it at scale.
Google has a potential advantage through its custom chips, cloud infrastructure and ability to design models and hardware together, although the company has faced capacity constraints of its own.
Tuesday's model launches come as Google is reportedly developing a specialized chip designed to run Gemini up to 10 times more efficiently, part of a broader push to lower the cost of serving AI.
A Google Cloud spokesperson told CNBC in a statement that its teams are "constantly researching and experimenting with new innovations to deliver maximum performance and efficiency for our users and customers" and that "while not every project moves into production, this rigorous exploration is central to our full stack approach."
"By co-designing our hardware and software from the ground up, we ensure our systems are integrated and highly optimized for real-world workloads," continued the statement.
Google is also offering more visibility into its roadmap after questions about delays. Gemini 3.5 Pro is now being tested with partners ahead of broader availability, while the company has begun its largest-ever pre-training run for Gemini 4.

### [E0496] Tech Insider | 2026-07-21 15:00 UTC | 盘中
Google doubles down on cheaper, faster AI — but says Gemini 3.5 Pro still isn't ready
URL: https://www.businessinsider.com/google-gemini-35-pro-flash-cyber-lite-speed-token-price-2026-7
Google is rolling out some new AI models — but not the one that many were expecting.
The company announced on Tuesday three Gemini models that it said are faster and cheaper to run AI agents, as Google continues to double down on efficiency over raw power.
That's not entirely by choice: Google said that Gemini 3.5 Pro, the new frontier model it originally promised for June, is still in testing. Google also teased Gemini 4, the next big flagship model, although that is not expected for several months.
Of the new models rolling out this week, Google describes 3.6 Flash as the "workhorse" that is better than 3.5 Flash at tasks such as coding. Google said 3.6 Flash reduces token usage — units of text processed by AI — by up to 17% compared to its predecessor, according to analysis from Artificial Analysis Index, a benchmark for comparing AI tools.
Google also unveiled 3.5 Flash-Lite, which it says is its fastest and "most cost-effective" version of its 3.5 model yet.
Then there's 3.5 Flash Cyber, which is designed to find and fix cybersecurity vulnerabilities at a performance level similar to rivals but at a "lower price per token than larger models," Tulsee Doshi, senior director of product management in Google's Gemini group, said in a blog post.
That last part is crucial. Anthropic and OpenAI have both rolled out cybersecurity models in recent months, but it seems that, again, Google is hoping to undercut the competition on cost. Google said Flash Cyber will initially be exclusively available to governments and trusted partners.
With the realization that using lots of tokens doesn't always merit the best results, more companies are reining in their AI spending and turning to more cost-conscious models. That gives Google a strategic opening.
"Companies are already blowing through their annual token budgets, and it's only May," Google CEO Sundar Pichai said earlier this year. "If companies used a mix of Flash and other frontier models, they could save a lot of money."
Google says its new models hit the "sweet spot" of efficiency and power for running agents.
No sign of 3.5 Pro
Gemini 3.5 Pro, Google's next big frontier model, is being tested with partners and will ship it "as soon as it's ready," the company said.
Business Insider reported in June that Google had pushed the targeted release date for 3.5 Pro to July. Bloomberg recently reported the model may have been further delayed.
As of Tuesday, Google hasn't got a single model among the top 10 ranking in the Artificial Analysis leaderboard, which uses a composite benchmark to score models across areas including maths and reasoning.
Google also said on Tuesday that it has begun its "most ambitious pre-training run yet" for Gemini 4, its next major model.
Typically, Google launches these big flagship models at the very end of the year, and it's not yet clear whether the 3.5 Pro will be powerful enough to vault Google back to the bleeding edge in the interim.
However, as companies take a more considered approach to their AI spend and cap their token usage, Google is positioning itself for the moment.

### [E0517] Reuters Technology | 2026-07-21 15:02 UTC | 盘中
Google updates lightweight Gemini models, but flagship still delayed
URL: https://www.reuters.com/business/google-updates-lightweight-gemini-models-flagship-still-delayed-2026-07-21/
Alphabet’s (GOOGL.O) Google on Tuesday released a trio of cheaper versions of ‌its Gemini AI model, but shared no timing updates for the flagship Pro model that has already seen its launch delayed by several weeks.

The new model lineup includes Gemini 3.6 Flash and Gemini 3.5 Flash-Lite, updated versions of its lightweight ​offerings , designed for use cases that do not require the most cutting-edge technology. Google also unveiled ​a new lightweight variant, Gemini 3.5 Flash Cyber, which is tailored towards cybersecurity ⁠applications.

But all eyes have been on the launch of Gemini 3.5 Pro, the top-tier version of the ​model, which Wall Street is waiting for as an indicator of whether Google’s DeepMind AI lab is ​able to keep pace with Anthropic and OpenAI.

Gemini 3.5 Pro was originally slated for a June launch, CEO Sundar Pichai said in his keynote address at the company’s I/O developer conference in May.

The company's launch of the Gemini 3 model family last ​November demonstrated it had returned to the front of the AI race after years as a laggard. ​It caused OpenAI to issue a “code red” internally to concentrate efforts around its competing GPT model.

Google’s main rivals have struck back ‌since ⁠then. Anthropic rolled out its latest Mythos 5 and Fable 5 models, which were deemed so powerful by the U.S. government as to be considered a national security risk .

OpenAI launched GPT-5.6 earlier this month after a similar government-requested delay to vet security concerns.

The delay of Gemini 3.5 Pro is due to the model reportedly falling short of ​internal goals, particularly in ​terms of coding, which ⁠has emerged as a lucrative use case for enterprise AI.

Pichai, though, has highlighted Google's pricing. He has said companies could save upwards of $1 billion per year ​by shifting most of their AI workloads to Gemini models.

Google DeepMind staff delivered ​similar messaging ⁠in press releases on the launches, describing the Gemini 3.5 Flash Cyber as “a cost-efficient and highly capable alternative to large, costly cybersecurity models."

As for the most anticipated model, Google said that Gemini 3.5 Pro is being tested ⁠with ​partners and would be coming “soon.” In one press release, it ​also said training had commenced on Gemini 4.

Pichai is likely to be pressed for more details about Gemini 3.5 Pro when Alphabet ​holds its second-quarter earnings conference call on Wednesday.

### [E0531] 财联社电报 | 2026-07-21 15:03 UTC | 盘中
谷歌发布多款Gemini模型，旗舰Pro版本发布时间仍未确定
URL: https://www.cls.cn/detail/2433110
【谷歌发布多款Gemini模型，旗舰Pro版本发布时间仍未确定】财联社7月21日电，Alphabet旗下谷歌7月21日发布三款成本更低的Gemini人工智能模型，包括Gemini 3.6 Flash、Gemini 3.5 Flash-Lite以及面向网络安全场景的Gemini 3.5 Flash Cyber，旨在满足无需使用最高性能模型的应用需求。不过，市场关注的旗舰模型Gemini 3.5 Pro发布时间仍未明确。该模型原计划于6月推出，但据报道因在内部测试中未达到预期目标，尤其是在代码生成能力方面存在不足，发布时间已推迟数周。谷歌表示该模型目前正与合作伙伴测试，将于“近期”推出。

### [E0532] 新浪财经7x24 | 2026-07-21 15:03 UTC | 盘中
谷歌：Gemini 3.5 Pro目前正在与合作伙伴测试，计划在准备就绪后尽快广泛推出。已开始对Gemini 4进行预训练。推出Gemini 3.6 Flash、Gemini 3.5 Flash-Lite及Gemini 3.5 Flash...
URL: https://wap.cj.sina.cn/pc/7x24/5000642
谷歌：Gemini 3.5 Pro目前正在与合作伙伴测试，计划在准备就绪后尽快广泛推出。已开始对Gemini 4进行预训练。推出Gemini 3.6 Flash、Gemini 3.5 Flash-Lite及Gemini 3.5 Flash Cyber。Gemini 3.5 Flash-Lite定价为每百万输入token 0.3美元、每百万输出token 2.5美元。

### [E0603] 华尔街见闻快讯 | 2026-07-21 15:19 UTC | 盘中
Alphabet旗下谷歌公司推出Gemini 3.6 Flash，3.5 Flash-LITE、以及3.5 Flash Cyber。 目前，Gemini 3.5 Pro处于与合作伙伴的测试过程中。
URL: https://wallstreetcn.com/livenews/3137668
Alphabet旗下谷歌公司推出Gemini 3.6 Flash，3.5 Flash-LITE、以及3.5 Flash Cyber。

目前，Gemini 3.5 Pro处于与合作伙伴的测试过程中。

Alphabet旗下谷歌公司推出Gemini 3.6 Flash，3.5 Flash-LITE、以及3.5 Flash Cyber。

目前，Gemini 3.5 Pro处于与合作伙伴的测试过程中。

### [E0604] 华尔街见闻快讯 | 2026-07-21 15:19 UTC | 盘中
谷歌平价Gemini三箭齐发，网安模型对标Anthropic Mythos，旗舰3.5 Pro仍难产
URL: https://wallstreetcn.com/articles/3777594
Gemini 3.6 Flash提升代码、知识工作和多模态能力，减少输出Token消耗的同时，输出价较前代低近17%；Gemini 3.5 Flash-Lite主打3.5系列中最低价和最快速；Gemini 3.5 Flash Cyber通过CodeMender代码安全智能体，试图以更低成本切入网络安全AI市场。谷歌透露，Gemini 4已启动“迄今最雄心勃勃的预训练工作”。

### [E0644] GitHub | 2026-07-21 15:25 UTC | 盘中
🆕 @GoogleAI’s Gemini 3.6 Flash is now generally available and rolling out in GitHub Copilot.

➡️ It is designed for web and app development, coding and agentic tasks
➡️ In testing, it demonstrated higher task-completion rates and better tok
URL: https://x.com/github/status/2079588576358998350
🆕 @GoogleAI’s Gemini 3.6 Flash is now generally available and rolling out in GitHub Copilot.

➡️ It is designed for web and app development, coding and agentic tasks
➡️ In testing, it demonstrated higher task-completion rates and better token efficiency than Gemini 3.5 Flash across coding and agentic workflows

Try it out in the GitHub Copilot app or @code.

### [E0660] Google DeepMind | 2026-07-21 15:29 UTC | 盘中
We’re rolling out three new models to make AI agents faster, smarter, and cheaper at scale:

🔵 Gemini 3.6 Flash: It uses fewer tokens than 3.5 Flash to deliver higher quality work at the exact same cost.

🔵 Gemini 3.5 Flash-Lite: A fast, co
URL: https://x.com/GoogleDeepMind/status/2079589698490572961
We’re rolling out three new models to make AI agents faster, smarter, and cheaper at scale:

🔵 Gemini 3.6 Flash: It uses fewer tokens than 3.5 Flash to deliver higher quality work at the exact same cost.

🔵 Gemini 3.5 Flash-Lite: A fast, cost-effective option for everyday tasks like processing documents and agentic search.

🔵 Gemini 3.5 Flash Cyber: A cybersecurity model built to find and patch critical software vulnerabilities.

### [E0675] Google AI | 2026-07-21 15:30 UTC | 盘中
Today, we're introducing not one but TWO new models, striking the balance between efficiency and quality to enable you to build production AI agents.

— Gemini 3.6 Flash: Addresses efficiency feedback we received from Gemini 3.5 Flash with
URL: https://x.com/GoogleAI/status/2079589742535118985
Today, we're introducing not one but TWO new models, striking the balance between efficiency and quality to enable you to build production AI agents.

— Gemini 3.6 Flash: Addresses efficiency feedback we received from Gemini 3.5 Flash with upgrades in coding, knowledge work, and multimodal tasks faster, more accurately, and with substantially fewer tokens per task

— Gemini 3.5 Flash-Lite: Our fastest, most cost-effective 3.5-class model yet built for agentic workflows, hitting ~350 output tokens/sec with improved coding and overall quality

Start building with these today via the Gemini API in @GoogleAIStudio or try them out in the @GeminiApp

### [E0855] Databricks | 2026-07-21 16:05 UTC | 盘中
Google’s Gemini 3.6 Flash is now available on Databricks!

In Databricks internal evaluations, it classified documents more accurately than Gemini 3.5 Flash and performed especially well on long-form patents and text-heavy documents. For do
URL: https://x.com/databricks/status/2079598684455551326
Google’s Gemini 3.6 Flash is now available on Databricks!

In Databricks internal evaluations, it classified documents more accurately than Gemini 3.5 Flash and performed especially well on long-form patents and text-heavy documents. For document parsing, medium effort delivered similar quality to high effort while using about 32% fewer total tokens. It also landed on the cost-quality frontier in our OfficeQA Pro agent harness benchmark.

Congrats to the @Google Gemini team on this release. We can’t wait to see what you build with it!

### [E0892] 财联社电报 | 2026-07-21 16:22 UTC | 盘中
谷歌更新轻量版Gemini模型阵容 但旗舰Pro发布再延期
URL: https://www.cls.cn/detail/2433141
①谷歌发布Gemini 3.6 Flash、Gemini 3.5 Flash-Lite两款轻量级模型升级版，以及优化网络安全能力的Gemini 3.5 Flash Cyber； ②谷歌仍未公布Gemini 3.5 Pro具体发布时间，下一代模型Gemini 4的训练工作已经启动。
财联社7月22日讯（编辑 牛占林）Alphabet旗下谷歌当地时间周二发布了三款成本更低的Gemini人工智能模型，但对于已推迟数周的旗舰版Gemini Pro模型，公司仍未公布具体发布时间。美股盘中，谷歌A股价下跌约0.8%。
此次推出的新模型阵容包括Gemini 3.6 Flash和Gemini 3.5 Flash-Lite，两者均为现有轻量级模型的升级版本，主要面向不需要最先进AI能力的应用场景。此外，谷歌还发布了一款全新的轻量级模型Gemini 3.5 Flash Cyber，该模型专门针对网络安全领域应用进行了优化。
据悉，Gemini 3.5 Flash Cyber专为检测和修复软件漏洞而设计，初期将仅通过有限访问测试项目向政府机构及可信合作伙伴开放。谷歌表示，与更大型模型相比，该专业化模型的Token运行成本更低。
这一模型有望帮助谷歌缩小其在网络安全AI领域与Anthropic之间的差距。后者凭借自动化代码防御技术，已经在该领域建立了先发优势。
不过，市场关注的焦点仍集中在Gemini 3.5 Pro的发布时间上。作为Gemini系列中的顶级模型，该产品被华尔街视为衡量谷歌旗下DeepMind人工智能实验室能否继续追赶Anthropic和OpenAI的重要指标。
谷歌首席执行官桑达尔·皮查伊曾在今年5月举行的谷歌I/O开发者大会主题演讲中表示，Gemini 3.5 Pro原计划于6月发布。
去年11月，谷歌推出Gemini 3系列模型后，市场一度认为该公司经过多年落后后重新回到了AI竞赛的第一梯队。该模型的发布甚至促使OpenAI内部启动“红色警报”，集中资源应对GPT系列模型面临的竞争压力。
此后谷歌的主要竞争对手也展开反击。Anthropic推出了最新的Mythos 5和Fable 5模型，美国政府认为其能力强大到可能涉及国家安全风险。
OpenAI则在本月早些时候发布了GPT-5.6，此前该模型也曾因政府要求进行安全审查而推迟发布。
据报道，Gemini 3.5 Pro延期的主要原因在于，该模型未达到谷歌内部设定的目标，尤其是在代码生成能力方面存在不足。而编程能力已经成为企业级AI应用中增长最快、商业价值最高的场景之一。
不过，皮查伊一直强调谷歌Gemini模型在价格方面的优势。他曾表示，企业如果将大部分AI工作负载迁移至Gemini模型，每年最高可节省超过10亿美元成本。
谷歌DeepMind员工在此次新模型发布的新闻稿中也传递了类似信息，称Gemini 3.5 Flash Cyber是“高效节约成本、同时具备强大能力的大型网络安全模型替代方案”。
至于市场最期待的Gemini 3.5 Pro，谷歌表示，该模型目前正在与合作伙伴进行测试，并将在“近期”推出。公司在一份新闻稿中还透露，下一代模型Gemini 4的训练工作已经启动。
预计在Alphabet周三公布第二季度财报并召开电话会议时，皮查伊将被投资者进一步追问有关Gemini 3.5 Pro发布时间及性能表现的更多细节。

### [E1028] TechCrunch AI | 2026-07-21 17:11 UTC | 盘中
Google releases three new Gemini models — but no 3.5 Pro
URL: https://techcrunch.com/2026/07/21/google-releases-three-new-gemini-models-but-no-3-5-pro/
On Tuesday, Google DeepMind released Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber. Gemini 3.6 Flash is Google’s “workhorse model” that promises improved capabilities in coding, knowledge work, and multimodal performance while reducing token usage by up to 17%, making it cheaper than its predecessor 3.5 Flash.
Gemini 3.5 Flash-Lite is the most cost-effective model in the class, and 3.5 Flash Cyber is a specialized model that was fine-tuned for finding and fixing cybersecurity vulnerabilities at a decent price point. This model will be exclusively available to governments and trusted partners as part of a limited access pilot program, according to Google.
Google says the focus on these releases is to deliver efficiency, latency, and reliability to customers that are building AI agents at scale.
The launch is notable not just for what Google shipped — cheaper, faster models optimized for coding, efficiency, and cybersecurity — but for what it didn’t. The update doesn’t include the long-anticipated update to Google’s flagship model, Gemini Pro, which was last updated in February.
In the time since that launch, OpenAI has released GPT-5.5 and begun rolling out GPT-5.6, while Anthropic has launched Claude Opus 4.8, Claude Sonnet 5, and expanded access to its frontier Fable 5 model, highlighting the intense release pace of the rival labs.
Google teased the release of Pro as part of the 3.5 Flash release in May, saying the Pro version was “already being used internally, and we look forward to rolling it out next month.” Last week, Bloomberg reported that Google was facing internal delays in launching the 3.5 Pro as it struggled to meet internal performance goals.
Gemini Pro models are generally Google’s highest-capability offerings for complex reasoning and coding tasks, while Flash models prioritize lower cost and faster response times for production applications.
Google DeepMind product lead Logan Kilpatrick said Tuesday that the company is currently testing Gemini 3.5 Pro with partners and hopes to “land soon.” He also noted that the team has started its most ambitious pre-training run yet for Gemini 4.

### [E1037] The Information | 2026-07-21 17:17 UTC | 盘中
Google on Tuesday announced three new versions of its more efficient “Flash” Gemini models, including one focused on cybersecurity.

But Google offered no update on the timing for its long-awaited 3.5 Pro model, expected to be its next flag
URL: https://www.theinformation.com/briefings/google-releases-new-gemini-flash-models-flagship-still-delayed
Google on Tuesday announced three new versions of its more efficient “Flash” Gemini models, including one focused on cybersecurity.

But Google offered no update on the timing for its long-awaited 3.5 Pro model, expected to be its next flagship model. At Google’s developer conference in May, CEO Sundar Pichai had said that the model would be released in June. “Gemini 3.5 Pro is currently testing with partners and we plan to make it broadly available as soon as it’s ready,” a Google DeepMind director [wrote](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-6-flash-3-5-flash-lite-3-5-flash-cyber/) in a blog post on Tuesday. Google has also started its pretraining run for Gemini 4, according to the post.

Google has struggled to catch up to Anthropic and OpenAI in creating AI that can generate code, seen as the most lucrative application of the AI boom. Last month, Google [expanded](https://www.theinformation.com/articles/google-revamps-new-ai-coding-strike-team-amid-struggle-catch-anthropic?rc=n9lbpq) the scope of its “strike team” focused on coding to try to improve on the capabilities of its models.

Instead, as the relative performance of its biggest models has slipped, it has tried to position itself as the [cost-effective, workhorse option](https://www.theinformation.com/newsletters/ai-agenda/google-pitches-ai-coding-tools-cost-effective-option?rc=n9lbpq). In its announcement on Tuesday, Google said that its new 3.6 Flash and 3.5 Flash-Lite models were more token-efficient—a measure of AI consumption—than previous versions. The 3.6 Flash model was also stronger on coding and knowledge work than 3.5 Flash, Google’s previous Flash model, Google said.

### [E1208] The Information | 2026-07-21 18:45 UTC | 盘中
Anthropic’s Opus 4.7 model took more than four seconds to generate the first token among customers whose data NiCE analyzed recently, while Google’s Gemini 3 Flash took about one second to do so. Google’s model was also generally speedier t
URL: https://www.theinformation.com/articles/anthropic-best-powering-customer-service-new-data-show
Anthropic’s Opus 4.7 model took more than four seconds to generate the first token among customers whose data NiCE analyzed recently, while Google’s Gemini 3 Flash took about one second to do so. Google’s model was also generally speedier than Opus 4.7, generating four times the number of tokens per second.

Despite the speed difference, Anthropic’s model costs 18.5 times *more* than Google’s Gemini 3 Flash. Anthropic has a much speedier model in Opus 4.6 Flash, which has a lower time to first token than Gemini 3 Flash, but it typically costs nearly 100 times more than the Google model, NiCE said. (Anthropic’s cheaper Sonnet and Haiku models are slower and slightly costlier than Gemini 3 Flash, the data show.)

OpenAI’s small, fast models, such as 5.4 mini and 4.1 mini, are also cheaper and increasingly popular options among NiCE customers, Verma said.

And OpenAI recently upped the ante. Earlier this month it launched a state-of-the-art audio model for ChatGPT, although that model isn’t yet sold through an application programming interface, according to OpenAI’s website.

The OpenAI audio model is known as bidirectional, meaning it continuously processes the speaker’s voice so it can immediately adjust its response when it’s interrupted, similar to a conversation between people.

There’s still one big advantage to using Anthropic: accuracy. Anthropic’s Opus 4.7 and 4.6, for example, make mistakes at half the rate of Gemini 3 Flash, NiCE said. That can be meaningful when customers expect excellent service.

“We offer a family of models because different jobs have different needs,” an Anthropic spokesperson said. While models like Fable and Opus are for frontier reasoning tasks that take minutes to days, they said, “Sonnet and Haiku balance high intelligence with speed for more high-volume work. What's consistent across our models is the quality of their completed work.”

**Atlassian Warns Against Anthropic**

Atlassian, like other SaaS providers, markets itself as an AI “orchestrator” that helps businesses switch between models and coding agents to develop software.

Last week, the company launched a bunch of AI updates to Jira, its software project management app. They give teams the ability to assign tasks, such as designing a website login page, to different AI applications including Anthropic’s Claude Code, Microsoft’s Github Copilot, and Atlassian’s own Jira coding agent. It also launched a more advanced version of its Jira AI agent for Slack that bears resemblance to Anthropic’s recently debuted Claude Tag, [an AI agent we covered in detail](https://www.theinformation.com/articles/salesforce-employees-worry-anthropics-invasion-slack).

Similar to Claude Tag, members of a team can tag the Jira agent in their Slack channels, which are essentially group chats, and ask it to analyze messages and create work items for team members. (The Jira agent will also be available in Microsoft Teams soon.)

“I want to have flexibility to use any model, and I don‘t want all of my company’s information and contacts just going to Anthropic,” said Tamar Yehoshua, Atlassian’s chief product and AI officer, on the advantage of using the Jira agent over Claude Tag. For more on the trillion-dollar question of whether top AI labs could use their customers’ data to develop competing products, [see our column from last week](https://www.theinformation.com/newsletters/applied-ai/ai-providers-threat-customers).

### New From Our Reporters

[https://www.theinformation.com/articles/mercors-fast-growth-relies-biggest-ai-companies-documents-show](https://www.theinformation.com/articles/mercors-fast-growth-relies-biggest-ai-companies-documents-show)Exclusive

#### [Mercor’s Fast Growth Relies on Biggest AI Companies, Documents Show](https://www.theinformation.com/articles/mercors-fast-growth-relies-biggest-ai-companies-documents-show)

By Cory Weinberg and Julia Hornstein

[https://www.theinformation.com/articles/google-plans-new-frozen-chip-run-ai-models-efficiently](https://www.theinformation.com/articles/google-plans-new-frozen-chip-run-ai-models-efficiently)Exclusive

#### [Google Plans New ‘Frozen’ Chip to Run Its AI Models Much More Efficiently](https://www.theinformation.com/articles/google-plans-new-frozen-chip-run-ai-models-efficiently)

By Erin Woo and Qianer Liu

[https://www.theinformation.com/articles/trumps-ai-agenda-collides-reality](https://www.theinformation.com/articles/trumps-ai-agenda-collides-reality)The Big Read

#### [Trump’s AI Agenda Collides With Reality](https://www.theinformation.com/articles/trumps-ai-agenda-collides-reality)

By Leo Schwartz

[https://www.theinformation.com/articles/new-kimi-k3-model-means-u-s-china-ai-race](https://www.theinformation.com/articles/new-kimi-k3-model-means-u-s-china-ai-race)AI Agenda

#### [What the New Kimi K3 Model Really Means for the U.S.-China AI Race](https://www.theinformation.com/articles/new-kimi-k3-model-means-u-s-china-ai-race)

By Alix Coutures, Rocket Drew and Aaron Holmes

[https://www.theinformation.com/articles/exclusive-oracle-data-centers-face-multibillion-dollar-cost-surprises](https://www.theinformation.com/articles/exclusive-oracle-data-centers-face-multibillion-dollar-cost-surprises)AI Infrastructure

#### [Exclusive: Oracle Data Centers Face Multibillion-Dollar Cost Surprises](https://www.theinformation.com/articles/exclusive-oracle-data-centers-face-multibillion-dollar-cost-surprises)

By Ann Davis Vaughan

### Upcoming Events

#### [Wednesday, July 22 — Creative Dealmaking in the Age of AI](https://www.theinformation.com/events/lvs-creative-dealmaking-in-the-age-of-ai)

Don't miss The Information's Dakin Campbell moderating an exclusive conversation on financing the AI infrastructure boom. Join leaders from Goldman Sachs, Blue Owl, and Skadden to unpack how Wall Street funds global computing.

[More details](https://www.theinformation.com/events/lvs-creative-dealmaking-in-the-age-of-ai)

#### [Wednesday, July 29 — Intimate Dinner with The Information and EY in New York City](https://www.theinformation.com/events/new-york-city-private-dinner)

Request an invitation to join The Information and EY for an exclusive dinner bringing together senior leaders for candid conversation on closing the trust gap in enterprise AI.

[More details](https://www.theinformation.com/events/new-york-city-private-dinner)

#### [Wednesday, September 23 — AI Agenda Live SF 2026](https://www.theinformation.com/events/ai-agenda-live-2026)

Get ahead of the story. Purchase your ticket to AI Agenda Live, where The Information's reporting team convenes the leaders driving the next era of AI. Early Access pricing available now.

[More details](https://www.theinformation.com/events/ai-agenda-live-2026)

#### [Tuesday, October 27 — The Information’s 2026 WTF Summit](https://www.theinformation.com/events/wtf-summit-2026)

Tickets on sale now: The Information returns to Napa Valley October 27–28 to convene senior women across tech, media, and finance. Join us for two days of intimate, candid conversations with the leaders navigating today’s global shifts.

[More details](https://www.theinformation.com/events/wtf-summit-2026)

### [E0600] Tae Kim | 2026-07-21 15:18 UTC | 盘中
Google releases smaller Gemini 3.6 models before 3.5 Pro? Yikes.
URL: https://x.com/firstadopter/status/2079586924847616452
Google releases smaller Gemini 3.6 models before 3.5 Pro? Yikes.

### [E0609] Tae Kim | 2026-07-21 15:20 UTC | 盘中
Google: "Beyond today’s releases, Gemini 3.5 Pro is currently testing with partners and we plan to make it broadly available as soon as it’s ready."

Gemini 3.5 Pro is NOT READY
URL: https://x.com/firstadopter/status/2079587296186081565
Google: "Beyond today’s releases, Gemini 3.5 Pro is currently testing with partners and we plan to make it broadly available as soon as it’s ready."

Gemini 3.5 Pro is NOT READY

### [E0620] Tae Kim | 2026-07-21 15:22 UTC | 盘中
"Gemini 3.6 Flash is worse than grok 4.5 in all coding task"

"Sonnet 5 > Grok 4.5 > GPT 5.6 Luna > Gemini 3.6 Flash"

"At this point I genuinely want google to be serious"
URL: https://x.com/firstadopter/status/2079587705738969433
"Gemini 3.6 Flash is worse than grok 4.5 in all coding task"

"Sonnet 5 > Grok 4.5 > GPT 5.6 Luna > Gemini 3.6 Flash"

"At this point I genuinely want google to be serious"

### [E0652] Tae Kim | 2026-07-21 15:27 UTC | 盘中
Why even release this? This is embarrassing for Google. You are late already. Who cares. They should have just moved on to Gemini 4.
URL: https://x.com/firstadopter/status/2079589193274061101
Why even release this? This is embarrassing for Google. You are late already. Who cares. They should have just moved on to Gemini 4.

### [E0809] Tae Kim | 2026-07-21 15:58 UTC | 盘中
Google getting beaten by @alexandr_wang this early, when Meta's new AI team is tiny in comparison, is absolutely embarrassing for Google.
URL: https://x.com/firstadopter/status/2079596813229666619
Google getting beaten by @alexandr_wang this early, when Meta's new AI team is tiny in comparison, is absolutely embarrassing for Google.

### [E0895] Chubby | 2026-07-21 16:24 UTC | 盘中
"Gemini 3.6 Flash scores 50 on the Artificial Analysis Intelligence Index, matching Gemini 3.5 Flash"

I really appreciate all they are doing, but..
URL: https://x.com/kimmonismus/status/2079603356125839519
"Gemini 3.6 Flash scores 50 on the Artificial Analysis Intelligence Index, matching Gemini 3.5 Flash"

I really appreciate all they are doing, but..

## 研究交付要求

这是一条已经从整场材料语义拆出的独立催化，不要重新把它压回大盘总评。核实美股异动原因和传播时序；用结构化供应链、内部 KB 与公司披露坐实每个亚洲候选；逐腿查真实行情和本地市场 price-in；候选不成立也要逐条给出有证据的 AVOID。