Google has introduced Credentio, an open source C++ library for C2PA (Coalition for Content Provenance and Authority) Content Credentials.

Announced August 13 and available at the mediaprovenance repository, Credentio provides an API designed to run locally within developer applications. This removes the need to send media files to cloud servers for validation, which incurs privacy, latency, bandwidth, and file size limitations.

Credentio is designed to start working with C2PA specification versions 2.2 and 2.4. This is the same code that has powered nearly 40 different conformant C2PA-enabled Google products to scale to tens of billions of generated assets, including images, videos, audio files, and documents across many file formats, Google said.

Through Credentio, media files do not need to be transmitted back to Google or external validation endpoints while offering zero bandwidth overhead, instant validation verdicts, and complete data privacy, the company said. By validating media files locally Credentio eliminates external data transmission requirements, minimizes verification latency, and delivers immediate results even in high-throughput workflows. And media contents remain securely within the local environment.

Credentio is tailored specifically for developers seeking to build performant, enterprise-grade C2PA validator products. Its small memory footprint makes it ideal for integrating into resource-constrained client applications, server pipelines, or high-performance edge software, according to Google.

As data centers evolve into AI factories, compute has shifted from a cost center to a revenue driver.

“Compute is revenue,” said Jensen Huang, co-founder and CEO of NVIDIA. “Without compute, there is no way to generate tokens. Without tokens, there’s no way to generate revenue. So, in this new world of AI, compute equals revenue.”

This reframe changes an organizations’ calculus. If compute is revenue, what do you optimize for? Here are 5 questions to consider:

1. Are you measuring what actually drives AI factory revenue?

Most AI factories are power-constrained, so tokens per watt dictate how much revenue you can generate and the cost per token impacts the AI factory profit margin.

But neither of these metrics should be evaluated at a single operating point. Batch jobs, real-time chat, and agentic workloads demand different points on the throughput-latency curve. AI chips that perform well at only a few points will underserve the full range of workloads.

Additional key operational metrics like time to first token (TTFT), mean time between interruptions (MTBI), and platform useful life are the bedrock of AI factory efficiency. They dictate how quickly an AI factory comes online to generate tokens, the reliability of its revenue streams, and its long-term ability to remain productive as AI workloads evolve.

2. How does agentic AI change what your CPU needs to deliver?

Data center CPUs have historically been optimized for parallel throughput, where more cores improve aggregate capacity. 

Agentic workloads run in loops and make different demands. The model reasons on the GPU, the CPU executes tool calls such as code compilation and data retrieval, and the result returns to the GPU so the model can reason again. Every step runs in sequence, gated by the one before it.

Per-core performance and memory latency determine how fast each step completes, which impacts the quality of service agents deliver and how well the AI factory stays utilized.

3. Is your networking and storage built for AI’s traffic patterns and data volumes?

Peak compute performance means nothing if the network cannot keep every accelerator productive. Networking requirements in an AI factory span three layers with performance demands that off-the-shelf Ethernet cannot deliver.

• Scale-up networking connects multiple accelerators and their memory with high bandwidth and low latency critical for today’s mixture-of-experts models.
• Scale-out networking enables high-speed direct data transfers between GPU memory across tens-of-thousands of servers, sustaining consistently low latency with zero jitter.
• Scale-across networking federates sites into a unified factory as power constraints push capacity across multiple locations, with intelligence and orchestration built in.

Storage must deliver more than capacity and throughput. Agentic workloads require fast, intelligent access to inference state and working memory across long context and multiple sessions. When storage paths can’t keep pace, GPU utilization drops.

4. Does your software stack hold up at scale and improve AI factory economics?

Turning hardware potential into realized performance requires a robust, proven software stack that optimizes every layer from compute primitives to inference frameworks to orchestration.

Open source software gives teams the flexibility to build, customize, and extend on a foundation shaped by a broad developer ecosystem. Strong enterprise-grade software captures that innovation while preserving the reliability for production AI. Moreover, software that delivers continuous performance gains at production scale reduces cost per token and extends the useful life of AI infrastructure.

5. Is security built into your AI data path?

A security breach can compromise customer data, model IP, or the integrity of agent decisions, and also result in downtime and lost token output.

Security must operate inline at AI factory speeds, across data at rest, in transit and in use. Storage must inspect agent behavior, enforce file and network access policies, and protect context memory in real time. At the compute layer, confidential computing with hardware-rooted attestation verifies workload integrity and protects models and data during inference.

How these questions shape extreme co-design at NVIDIA

These considerations from NVIDIA customers have shaped how we build. NVIDIA’s extreme co-design vertically integrates compute, networking, storage, and software to deliver the best performance, efficiency, resilience, and security to optimize AI factory economics. The NVIDIA platform is also horizontally open, ranging from NVIDIA MGX and DSX reference architectures to NVLink Fusion support for third-party XPUs to a broad open source software ecosystem.

The proof is in the performance leadership:

• NVIDIA Vera Rubin NVL72 delivers 10x more tokens per megawatt than NVIDIA GB200 NVL72.
• Vera Rubin’s cableless rack-scale architecture reduces tray assembly from 2 hours to 5 minutes with a 95% first-pass success rate, accelerating bring up and time to first inference. Combined with resiliency software, it sustains uptime. 
• NVIDIA Groq 3 LPX delivers up to 35x higher throughput per megawatt for ultra-low latency inference
• NVIDIA Vera CPU delivers up to 2x higher single-threaded core performance, 3x higher core-to-core bandwidth, and 40% lower memory latency to accelerate agentic AI.
• NVIDIA NVLink, now in its sixth generation, is purpose built for scale-up networking and delivers 3x lower latency and 10x higher packet rate versus off-the-shelf Ethernet.
• NVIDIA Spectrum-X Ethernet delivers up to 1.6x higher performance than off-the-shelf Ethernet and sustains up to 95% efficiency across deployments exceeding 100,000 GPUs.
• NVIDIA BlueField-4 DPU delivers 800Gb/s connectivity and 6x the compute of its predecessor, while the Vera BlueField-4 STX Storage Processor powers NVIDIA CMX to deliver up to 5x higher tokens per second for agentic inference.
• NVIDIA DOCA on BlueField-4 delivers in-silicon security, with runtime threat detection up to 1,000x faster than existing agentless solutions and file and network access policy enforcement.
• NVIDIA Confidential Computing secures models and data in use across every GPU and CPU in the Vera Rubin NVL72.
• NVIDIA software stack powers the world’s largest AI factories. On Blackwell, continuous optimizations reduced token costs for DeepSeek V4 by up to 5x within 1 month.

---

Model routing has emerged as a way to keep AI inferencing costs down by examining prompts and directing them to the most appropriate model — for example, the cheapest one able to respond effectively. This means that requests can be handled more efficiently, giving more accurate results and lower runtime costs, a concern in an era of rising AI expenditure.

Nvidia sees the potential for model routing and has just launched NeMo Switchyard, its take on the technology. Switchyard provides a library for applying multiple routing approaches, enabling developers to apply a system-of-models approach and build more efficient, controllable agents.

Interest in model routing is growing — both as a tool and as an investment: Cloudflare recently introduced a model router as part of a new suite of enterprise AI tools, while payment services company Stripe is looking to buy OpenRouter, according to The Wall Street Journal.

Companies will certainly be looking at routing options more carefully in the future as they work out how best to marry their infrastructures with their AI demands.

Oracle has released a security tool intended to provide organizations with a centralized view of security risk across their database environments.

Oracle Database Security Central will be available free of charge until the end of February 2027.

It arrives at a critical time for Oracle customers, with attackers targeting security flaws to exploit the company’s database products.

Oracle is also concerned about the threat posted by the launch of bug-hunting AI model Mythos. In May it responded by accelerating its patching schedule, switching to monthly releases rather than quarterly. The first monthly batch fixed 35 flaws.

Oracle said Security Central will enable security teams to assess security posture, detect configuration drift, as well as identifying privileged-user and access risks. It will also keep an eye on sensitive data and analyze how it is being accessed, and collect audit evidence accordingly. Finally, it will manage security policies centrally to ensure that policy variance is not putting the company at risk.

Google has launched Gemini 3.7 Flash, with updates focused on coding, automation, and agent workflows, alongside lower pricing for production deployments.

The release, just three weeks after Gemini 3.6 Flash, reflects what the company described as rapid iteration driven by developer feedback. Google positioned the model as its “most intelligent workhorse model yet for coding and agents,” aimed at software engineering and multi-step workflows.

Gemini 3.7 Flash is priced at $0.75 per million input tokens and $3.75 per million output tokens — roughly half the cost of its predecessor — signaling a push to make production deployments more economically viable.

“Gemini 3.7 Flash delivers a noticeably improved developer experience over 3.6 Flash,” Google said in a statement. “It better adapts to roadblocks, clarifies intent when needed, and follows instructions with greater fidelity.”

Faster Flash cycle, slower Pro progression

The release comes as vendors are adopting different update cycles across model tiers.

Google’s latest updates are concentrated in its Flash series, which has seen frequent releases. More advanced “Pro” models, typically designed for complex reasoning, continue to follow a slower update cadence. Google has not provided a timeline for its next Pro release, and its CEO, Sundar Pichai, dodged questions related to the Pro release during the company’s recent quarterly earnings call.

A similar split is visible elsewhere. DeepSeek this week introduced its V4-Pro model as a higher-end offering alongside its V4-Flash variant, reflecting a broader separation between cost-efficient and high-capability tiers.

Coding and workflow gains tied to efficiency

Google said Gemini 3.7 Flash improves debugging, issue resolution, and first-pass code generation. In company benchmarks, the model scored 43.6% on FrontierCode 1.1 Main, up from 34.4% in version 3.6, and 65.3% on DeepSWE v1.1, compared with 49.0%.

It also reported gains in workflow automation, with a 30.4% score on AutomationBench versus 17.0% earlier.

“3.7 Flash delivers substantial improvements across software engineering, knowledge work, and web development workflows,” the statement added.

“These remain vendor benchmark claims until the new model accumulates sufficient independent production evidence,” said Sanchit Gogia, chief analyst at Greyhound Research.

For enterprises, these gains are relevant when they translate into operational efficiency, said Amit Chandak, chief analytics officer at Kanerika.

“Benchmark improvements become meaningful in enterprise environments when they translate to fewer correction loops, less human oversight per task, and more reliable multi-step execution,” Chandak said.

He added that production teams are increasingly focused on efficiency. “The more relevant number for production teams is token efficiency,” he said, noting that reductions in token usage can lower both latency and cost at scale.

Google said the model generates more complete web applications with fewer prompts and improved adherence to design inputs. On WebDev Arena, it achieved an Elo score of 1588, compared with 1538 for its predecessor.

In knowledge-intensive domains such as finance, law, and biosciences, Gemini 3.7 Flash scored 34.0% on the GDP.pdf benchmark, up from 22.0%.

The company also said the model “thinks more diligently” and follows instructions with greater fidelity, improving multi-step planning and tool use.

Diverging pricing strategies emerge

Google’s price cut comes as vendors take different approaches to AI pricing.

DeepSeek launched its V4-Pro model at significantly higher price points than its Flash variant, with output token costs reaching about $3.96 per million tokens during peak usage, compared with much lower rates for its V4-Flash model.

“Token cost has been the practical ceiling on scaling AI beyond isolated pilots,” Chandak said. At lower price points, he added, running agent-based workflows at production scale becomes more viable.

He also said enterprises are placing greater emphasis on factors beyond model performance. “The base model layer is commoditizing,” Chandak said, adding that differentiation will increasingly depend on data readiness, governance, and orchestration layers.

Gogia said pricing shifts reflect broader changes in how enterprises evaluate AI systems. “The more important development is the continued compression of the price of useful machine intelligence,” he said. “Capability, latency and cost are becoming inseparable buying criteria.”

“The model becomes an ingredient. The operating architecture becomes the advantage,” Gogia said.

Agent adoption remains measured

Google highlighted agent-based workflows as a key use case, with the model capable of orchestrating multiple sub-agents to generate applications and automate tasks.

“From a simple text prompt to a fully playable 3D game,” the company said, describing multi-agent orchestration capabilities.

The model also supports multimodal workflows, including transforming static documents into interactive outputs and enabling faster iteration in robotics training.

Chandak said enterprise adoption remains measured. “The organizations making real progress are the ones that started small, picked a single high-volume workflow, proved the outcome, and then expanded,” he said.

He added that governance remains a constraint. “The binding constraint on enterprise agent adoption has always been the governance and accountability layer,” he said, citing challenges around decision ownership, auditability, and data access. The model is available through Google AI Studio, Android Studio, and enterprise platforms including Gemini Enterprise, and is being integrated into Gemini Spark for workflow automation tasks, the statement added.

Enterprises are taking a serious look at neoclouds, the specialized cloud providers built primarily around AI infrastructure, especially GPUs, high-speed networking, and large-scale compute clusters for model training and inference. Unlike traditional hyperscalers that provide broad platforms for almost every kind of enterprise workload, neoclouds tend to focus more narrowly on accelerated computing. CoreWeave, Lambda, Crusoe Cloud, and others are all commonly associated with this emerging AI infrastructure market.

The interest is not difficult to understand. Enterprises are under pressure to move generative AI, machine learning, and advanced analytics projects out of the lab and into production. At the same time, access to large blocks of GPU capacity has become expensive, constrained, and in some cases difficult to obtain from the major hyperscalers. Many enterprises are finding that neoclouds can offer better economics, faster access to capacity, or configurations more closely aligned with AI workloads.

This does not mean AWS, Microsoft Azure, and Google Cloud are being displaced. They remain the default operating environment for most enterprise cloud deployments. They provide mature administrative planes, security tools, compliance frameworks, global footprints, managed services, and operational ecosystems that enterprises have spent years learning how to use.

However, AI has changed the infrastructure conversation. Enterprises are worrying less about which cloud they are standardized on and focusing instead on getting the AI capacity they need, when they need it, at a price that does not destroy the business case.

That brings up one of the most common questions I get from clients: “How different is it to maintain these remote AI cloud systems compared with what we already do on AWS, Azure, or Google Cloud?” My answer is that the fundamentals of cloud operations still apply, but the administrative model does change in important ways. Neoclouds are not simply cheaper hyperscalers. They are specialized infrastructure environments, and specialization always creates trade-offs.

The biggest administrative differences show up in three areas: security, performance, and business continuity/disaster recovery.

Less-developed security

Security in the hyperscaler world is mature because the administrative ecosystem is mature. AWS, Microsoft, and Google have spent years building deeply integrated identity systems, key management services, logging tools, policy engines, compliance programs, network controls, vulnerability management capabilities, and security monitoring services. Enterprises still misconfigure these services all the time, but the building blocks are well known and widely understood.

With neoclouds, security administration may require more direct enterprise ownership. Some providers have strong security capabilities and mature operational practices. Others are still building out the kinds of enterprise-grade controls large organizations expect from the hyperscalers. That means administrators cannot assume that identity federation, privileged access controls, audit logging, encryption, network segmentation, and compliance reporting will behave in familiar ways.

This matters because AI workloads often involve some of the most valuable data an enterprise owns. Training sets, fine-tuning data, prompts, embeddings, model weights, vector databases, and inference outputs may contain intellectual property, customer data, regulated information, or confidential business logic. If an enterprise is using proprietary operational data to fine-tune a model, the administrative stakes are higher than simply spinning up remote compute.

The shared responsibility model still applies, but it must be examined provider by provider. Enterprises need to understand who controls encryption keys, how administrative access is granted and revoked, how logs are exported to the security operations center, how data is isolated between tenants, and how provider personnel access is governed. These are not paperwork questions. They are operating model questions.

Hands-on performance management

The second difference is performance. Traditional cloud administration has trained enterprises to think in abstractions. Administrators select instance types, storage classes, managed databases, autoscaling policies, and observability dashboards. The underlying hardware matters, but it is usually hidden behind a service model.

AI changes that. With neoclouds, performance administration often gets much closer to the physical infrastructure. GPU type, GPU memory, interconnect design, storage throughput, cluster topology, job scheduling, data locality, and network latency can all have a direct effect on whether an AI workload performs well or wastes money.

GPU economics are unforgiving. An idle or underutilized GPU is a major financial problem. If data pipelines cannot feed accelerators fast enough, if distributed training is misconfigured, or if storage throughput becomes the bottleneck, the enterprise can quickly lose the cost advantage that made the neocloud attractive in the first place.

Administrators therefore need to understand more than basic cloud operations. They need to know how AI workloads behave at scale. They need to understand how training jobs consume storage and network resources, how inference demand fluctuates, how clusters are allocated, and how to measure actual accelerator utilization. This requires closer collaboration among cloud operations, AI engineering, data engineering, platform engineering, and finance.

Capacity planning also changes. Hyperscalers created the expectation of near-infinite elasticity, even though that expectation has always been somewhat exaggerated. In the AI market, it is even less reliable. Neoclouds may provide better access to GPU capacity, but that capacity may come through reservations, fixed clusters, specific hardware commitments, or contractual windows. Administrators need to align training schedules, experimentation cycles, inference growth, and budget controls with the provider’s actual capacity model.

Performance administration in neoclouds is not just about watching dashboards. It is about managing workload economics at the infrastructure level.

Detailed disaster recovery plans

The third difference is business continuity and disaster recovery. Too many enterprises still believe that if something runs in the cloud, resilience is included. That assumption is dangerous in any cloud environment, but even more so when dealing with specialized AI infrastructure.

The hyperscalers provide large global footprints, multiple regions, availability zones, replication services, backup tools, managed failover options, and well-documented resilience patterns. Neoclouds may not offer the same geographic depth or the same range of native continuity services. Administrators must be much more explicit about recovery objectives, failover design, replication, and restoration procedures.

AI workloads complicate this further. Recovering an AI system is not the same as restoring a traditional application server. Enterprises need to protect data sets, training checkpoints, model artifacts, feature stores, vector databases, orchestration pipelines, container images, configuration files, and inference endpoints. If a neocloud environment becomes unavailable, can the business restart training from a checkpoint? Can inference move to another environment? Can the same model run on different accelerators, drivers, frameworks, and networking assumptions?

Those questions need answers before the outage, not during it. Some AI workloads can tolerate delay. A training job may be paused and restarted later without major business impact. Other workloads, especially production inference systems embedded in customer-facing processes, may require much more aggressive recovery targets.

Enterprises should evaluate neoclouds with realistic expectations. The economics may open the door, and the capacity may make the decision urgent. The long-term success of neocloud adoption, however, will depend on how well enterprises administer the differences.

One of AI vendor DeepSeek’s biggest selling points has been its ultra-low price point, but that party’s about to end.

The Chinese model provider is raising API pricing for its V4 model family by notable margins, in some cases by more than 1,100%. The increases may not be that dramatic for all, though; the company is encouraging “more flexible workload scheduling,” with peak rates and half-price off-peak rates.

The news was tucked into the announcement of the general availability (GA) of DeepSeek V4-Pro and upgrades to VR-Flash. The new pricing takes effect for most parts of the world on August 16.

“On paper, at peak, against the right comparator, DeepSeek’s price advantage does disappear, and in places inverts,” said Sanchit Vir Gogia, chief analyst at Greyhound Research. But in practice, “the schedule’s own clock and cache hand most of it back to any buyer paying attention.”

How Flash and Pro compare now

The new API pricing structure is as follows:

• Flash is now $0.22 per million input tokens (cache miss) and $0.66 per million output tokens off-peak; and $0.44 per million input tokens (cache miss) and $1.32 per million output tokens at peak.
  This is up from the flat rate of $0.14 for inputs (cache miss), representing a 57% to 214% increase, and $0.28 per million tokens for outputs, a 136% to 371% increase.
•  Pro is now $0.66 per million input tokens (cache miss) and $1.98 per million output tokens off-peak; and $1.32 per million input tokens (cache miss) and $3.96 per million output tokens at peak.
  This represents an input increase of between 51% and 203% (up from $0.435) and output increase between 127% and 355% (up from $0.87).

Inputs with cache hits, when apps reuse stored prompts rather than processing similar requests from scratch, have even more dramatic pricing increases of 52% to 1,100%.

Mark Tauschek, VP of research fellowships and distinguished analyst at Info-Tech Research Group, pointed out that the increase does eliminate the price advantage that 4.0 Flash has over OpenAI 5.6 Luna at peak pricing, but not at off-peak pricing, as OpenAI has dropped Luna API pricing by 80%, off-peak.

It also doesn’t eliminate Deepseek 4.0 Pro’s price advantage over Terra, OpenAI’s GPT-5.6 mid-tier reasoning model, even at peak pricing, nor its advantage over GPT-5.6 Sol released in July, Tauschek said.

Greyhound Research’s Gogia noted that, off-peak, V4 Flash is “marginally more expensive” on input and 45% cheaper on output than Luna. Pro at peak, meanwhile, runs close to 5x Luna’s price on a representative coding-agent workload.

DeepSeek’s roughly 98% cache-hit discount, against an industry norm nearer to 90%, is the mechanism that has kept its measured cost per task at about 60% below Luna, even after Luna’s cost cut, he said.

“The schedule re-prices exactly that mechanism,” Gogia said. Flash’s edge over Luna decreases from roughly sevenfold to threefold off-peak, and 1.4 times at peak. “The cache is where the advantage genuinely erodes.”

Encouraging users to rethink their schedules

DeepSeek’s V4-Pro is now generally available, and V4-Flash is in beta. Both models have new flexible reasoning capabilities (low, high, max) and ‘thinking modes’ that use chain-of-thought (CoT) reasoning to improve answer accuracy. V4 Pro is now available on app, web, and via API, and users can try it using “Expert Mode.” V4 Flash is now in beta.

The general availability “completes a two-tier structure in which Flash serves volume and Pro is priced for complexity,” Gogia noted.

DeepSeek’s peak/off-peak pricing is a means to “allocate resources more reasonably,” the company said, to encourage users to “schedule their tasks based on actual usage.”

Gogia pointed out that with the new model, 17 of every 24 hours stay at half price, so timing becomes an economic variable, and work that can wait moves into the cheap hours. In fact, the new pricing schedule hits DeepSeek’s home market hardest and its export market lightest; Western buyers largely pay the off-peak rates.

“Usage is following economics at least as much as capability, and economics can change by schedule,” Gogia noted.

Simple supply and demand

Reading between the lines provides a more nuanced picture, Tauschek noted. “While it’s alarming to see the headlines saying DeepSeek is raising API pricing by 50%-1100%, it doesn’t really tell the whole story.”

Part of that story is demand, which is increasing exponentially. DeepSeek can’t keep up with compute requirements, and Anthropic also had a price increase for the same reason in April. And, while third-party providers have not yet reflected that trend, they’ll eventually have to, Tauschek said.

“This isn’t unexpected at all,” he noted. “It’s simple supply and demand: when demand goes up, pricing goes up, because supply becomes constrained.”

For enterprises that do use DeepSeek (many in the US do not, or can not), the new pricing is not likely to change anything, he said. Cost increases will mostly impact developers, but it will still be less expensive than most alternatives.

He pointed out that enterprises are adapting to model routing, which is critical for developers using agentic workloads. Just a few months ago, organizations were paying per-seat pricing and running up usage as a matter of course, but the market move to usage-based pricing has resulted in sticker shock akin to that of the early cloud days.

“Pricing will continue to be a big deal because CFOs are starting to ask what they’re getting for the massive AI spend,” Tauschek said.

DeepSeek pricing doesn’t change the need for compatibility, multi-modality

CIOs should read the schedule with “relief and unease,” Gogia noted. Relief because the bill is largely schedulable; unease because “a supplier that has learned to price the clock has learned something about its own leverage.”

Going forward, he predicted, Flash keeps the volume usage, Pro handles complexity, and interface compatibility lowers the cost of adoption and departure. The real question becomes whether lower economic floors, open weights, and compatible interfaces, when taken together with multi-model routing, make foundation model intelligence materially easier to substitute.

Capable inference can be produced “far below the price structures that once surrounded frontier AI,” Gogia noted, and open weights mean model developers become one of just several parties able to serve inference requirements. “The traditional software dependency changes shape when that happens,” he said.

The vendor still matters, as do capability and support, but once a workload can move between providers, and enterprises manage their own orchestration and governance, the vendor no longer owns the whole dependency, Gogia said.

The most lasting effect of DeepSeek is unlikely to be that it stayed cheapest, he noted. “It is that every provider must now explain why intelligence should command a premium once near-equivalent capability is available through several technical and commercial routes.”

The MSVC Build Tools Preview is updated regularly with the latest features and fixes from the MSVC (Microsoft C++) development team. The MSVC team’s updates for August 2026, targeting the Visual Studio v14.52 release, bring numerous improvements to the C++ front end, modules, code generation, debugging, static analysis, the linker and assembler, and more.

The updates were announced in an August 13 blog post. Developers can get the MSVC Build Tools Preview through the Visual Studio 2026 Stable Channel or (more quickly) through the Insiders Channel.

C++ compiler front end improvements include added support for default-initialized new[] expressions in constant evaluation when used with modules and improved diagnostics for calling-convention mismatches, incompatible compiler options, and unusable compiled-library files. C++ module improvements include improved type merging when header units are shared across multiple modules, and improved module serialization of data members and template information. Optimizer updates include improved optimization of equality-test loops with loop-carried dependencies, improved jump-to-jump optimizations, and improved optimization around restricted pointers and function calls.

Debugging improvements include better diagnostics produced by debug-information inspection tools and improved type merging for programs that use header units and modules. Work continues to support larger debug-information streams, including groundwork for streams up to 4 GB. And a way has been added for debugger clients to release debug-database file handles without immediately discarding loaded data, allowing the file to be rebuilt sooner. Static analysis improvements include expanded code-analysis coverage across more components of the toolset.

The MSVC Build Tools Preview updates also include dozens of bug fixes. Developers are encouraged to try out the MSVC Build Tools and let Microsoft know what they think. Feedback can be shared on the Visual Studio Developer Community site.

Microsoft has released Visual Studio Code 1.133, an update to its code editor that brings more flexibility to Anthropic Claude sessions. The update also allows users to open the Agents window without signing in to GitHub.

VS Code 1.133 was released August 12. It can be downloaded for Windows, Linux, and Mac from code.visualstudio.com.

With VS Code 1.133, users now can mix Anthropic and Copilot providers in Claude sessions. Previously, a Claude session ran entirely through either a GitHub Copilot subscription or Claude’s existing configuration, such as an API key. Switching providers required reconfiguring the agent host. Now, the model picker displays both groups, so users can switch providers between turns. The model selected is used for the next turn. Models under Anthropic bill the API key, and models under Copilot use the Copilot subscription.

A new experimental setting, chat.agentHost.allowSignedOutWhenUsable, allows the Agents window to be opened without requiring GitHub sign-in. Previously, the Agents window opened with a GitHub sign-in prompt that could not be dismissed, blocking users whose machine could not reach github.com and users who do not interact with GitHub. Enabling this setting associates GitHub authentication with individual agents or models instead of the Agents window. In this release, this behavior only supports Claude. Support for Copilot with the user’s own model keys and Codex is planned for future releases.

VS Code 1.133 also brings auto-reload to HTML files opened in the integrated browser, refreshing them automatically when the file changes on disk. Developers can toggle automatic reload for individual browser tabs and configure the default with the workbench.browser.autoReloadOnFileChange setting.

Microsoft also introduced a setting to enable sticky scroll for prompts in the chat window. The chat.stickyScroll.enabled setting allows users to pin prompts to the top of the chat, similar to sticky scroll in the editor. Users then can select the pin to jump back to the prompt, or use the previous and next buttons beside it to step through prompts.

VS Code 1.133 follows VS Code 1.132, released August 5.