A report by the Wall Street Journal said Meta had repeatedly delayed plans to make the API available to developers and, as of Tuesday, had not set a launch date. The report cited people familiar with the matter.

However, Meta disputed suggestions that the project had stalled. A company spokesperson said on Wednesday that testing is already underway with a group of early partners and that the company still expects to release the API later this month.

“The muse spark API will be coming soon,” Meta AI Chief Alexandr Wang announced in a post on X in April.

Meta AI Unveils Spark to Power Next-Gen AI across Platforms

The API would allow developers to integrate Muse Spark into their own software and services. An API, or Application Programming Interface, is a software bridge that enables different systems to communicate and work together.

Meta introduced Muse Spark in April as the first model developed under its Superintelligence Labs initiative, which was created to strengthen the company’s position in the competitive AI market.

The model is designed to narrow the gap between Meta and competitors including OpenAI, Anthropic and Google.

While Muse Spark is already available to consumers through Meta’s applications, users can currently access it only through built-in modes such as Instant, Thinking and Contemplating. Developers still do not have access to a public API, and Meta has yet to release documentation, pricing details, rate limits or eligibility requirements.

The lack of information has created apprehension among developers hoping to build products around the model. Without a public timeline, waitlist or technical documentation, companies interested in integrating Muse Spark are unable to plan deployments or assess costs.

The delays also come at a sensitive time for Meta. Investors have been monitoring the company’s AI strategy as it spends heavily on infrastructure, talent and product development.

Questions about execution have grown following reports of an Instagram security incident involving Meta’s AI-powered support system, which exposed weaknesses in automated account management processes.

Earlier on Wednesday, Meta unveiled a new AI agent designed to help businesses handle day-to-day tasks, showing that the company is going beyond consumer chatbots and into enterprise services.

The launch highlights Meta’s goal to compete more directly with OpenAI, Anthropic and Google across multiple areas of the AI market.

Muse Spark is expected to bolster that strategy. It is the first in what Meta has described as a new generation of advanced models from its Superintelligence Labs unit.

However, the repeated postponements have left analysts, developers and investors waiting for evidence that the company can translate its AI investments into products that are ready for global use.

Access is still currently limited to a small group of testing partners, while the developer community is waiting for Meta to open the platform to the public.

GPT-5.4-Cyber is a version of OpenAI’s latest model adjusted for defensive security work and will not be widely available at launch. Instead, OpenAI is giving early access to selected security firms, organisations and researchers.

The release follows Anthropic’s recent launch of its own frontier model, Mythos. That system is being tested under a restricted programme known as Project Glasswing, where only approved groups can use it for cybersecurity tasks.

According to Anthropic, the model has already identified thousands of serious weaknesses across software systems.

OpenAI is taking a comparable route but with a wider rollout plan over time. The company is expanding its Trusted Access for Cyber programme, which it introduced earlier this year. This scheme verifies users before granting them access to more capable tools.

Under the updated structure, more individuals and teams will be admitted, but access depends on how much information they provide to confirm their identity and role. Those in the highest tier will be allowed to use GPT-5.4-Cyber.

The company said the model has fewer restrictions when handling sensitive tasks such as vulnerability research and code analysis. It is designed to support security professionals who need to examine software more deeply, including analysing compiled programmes without access to their source code.

At the same time, OpenAI is carefully monitoring how the system is used. Because the model allows more freedom, the company is limiting its release and adjusting safeguards as it learns from real-world use.

Tools like GPT-5.4-Cyber can be used for both defence and attack, OpenAI acknowledged that risk, noting that threat actors are already experimenting with artificial intelligence to find new ways into systems.

To manage that, the company said access will not just depend on the model itself, but on who is using it and for what purpose. Strong identity checks and clearer signals of intent are being built into the process.

The aim is to make security tools more widely available without opening the door to misuse. OpenAI said it does not want to decide centrally who gets to defend systems, but it still needs controls that can scale.

This latest release builds on earlier initiatives, including its cybersecurity grant programme and tools designed to scan and fix software vulnerabilities. The company said these systems have already helped address thousands of high-risk issues.

OpenAI expects both risks and benefits to grow, saying future models will likely require stronger protection, even as they provide more advanced support for those working to protect digital infrastructure.

The announcements cover hardware, software, cloud services, and partnerships with companies like Lyft, Sony, and Visa.

The focus this year is on giving businesses better management over AI systems. AWS CEO Matt Garman spoke about how AI agents can drive tangible business results.

“AI assistants are starting to give way to AI agents that can perform tasks and automate on your behalf,” he said during the keynote. “This is where we’re starting to see material business returns from your AI investments.”

Powerful Chips and UltraServers

AWS introduced the Trainium3 chip and UltraServer systems, promising up to four times faster AI training and inference while using 40% less energy.

Trainium4, already in development, will be compatible with Nvidia’s chips, signalling AWS’s intent to bridge proprietary and third-party hardware ecosystems.

AI Agents Evolving

AWS also expanded its AgentCore platform. Features like Policy allow developers to set clear boundaries for agents, while new memory and evaluation capabilities let AI agents remember interactions and be tested against 13 prebuilt evaluation systems.

Among the new “Frontier agents,” the Kiro autonomous agent stands out, writing code and learning team workflows to operate independently for hours or even days. Additional agents focus on security and DevOps tasks, helping teams prevent errors and manage operations more efficiently.

Nova AI and Customisation

Amazon’s Nova family of AI models grows with four new releases, including three text-generation models and a multimodal model that handles text and images. Nova Forge introduces “open training,” enabling organisations to fine-tune pre-trained models with proprietary data.

Companies like Reddit and Hertz are already leveraging Nova to replace multiple specialised models or accelerate development velocity.

Real-World Applications

AWS customers demonstrated practical impacts. Lyft’s AI agent, built with Anthropic’s Claude model via Amazon Bedrock, now resolves driver and rider queries 87% faster and has increased driver adoption by 70%.

Christina Minardi from Amazon noted sustainability applications: “By working with Trane Technologies and the BrainBox AI team, we’re turning our buildings into intelligent systems that learn and adapt, helping us meet both our sustainability and performance goals in real time.”

Other partners showcased broad enterprise use cases. Sony is deploying AWS-powered AI platforms internally and through the Sony Engagement Platform, processing 760 terabytes of data daily to enhance fan experiences.

Nissan’s cloud-based software platform for vehicles has reduced testing time by 75%, while Visa and AWS are enabling AI agents to conduct secure, autonomous transactions.

Data Control and Sovereignty

AWS also introduced AI Factories, which allow companies and governments to run AWS AI in their own data centres. Combining Nvidia GPUs with Trainium3 chips, the system meets regulatory and data sovereignty requirements without sacrificing performance.

Cloud Services and Storage Upgrades

Several AWS services received significant updates. Amazon S3 now supports objects up to 50TB and scales to two billion vectors per index for AI search, while S3 Tables introduces automatic replication and cost-optimising Intelligent-Tiering.

CloudWatch unifies operational, security, and compliance logs for easier insights, and EMR Serverless eliminates local storage provisioning for Apache Spark jobs, cutting costs by up to 20%.

Enhanced Support and Security

AWS also announced upgraded support plans, combining faster AI-assisted responses with expert guidance. Amazon GuardDuty Extended Threat Detection now covers EC2 and ECS environments, while Security Hub offers near real-time risk analytics across multiple AWS services.

Expanding Partnerships

Adobe, Deepgram, BlackRock, and WRITER highlighted collaborative initiatives. Adobe is using AWS for AI-powered creative tools, Deepgram for enterprise voice solutions, BlackRock for Aladdin investment technology, and WRITER for securely scaling enterprise AI agents.

The announcements underline AWS’s strategy to embed AI across infrastructure, enterprise software, and real-world operations.

Starting from autonomous coding agents to sustainability-driven building systems, the AWS re:Invent 2025 conference revealed how businesses are starting to rely on AI agents not just as tools, but as autonomous collaborators.

Users condemned GPT-5 for its “cold,” “mechanical,” and “less empathetic” tone, leading to threats of subscription cancellations. In response, OpenAI restored GPT-4o to the model picker for all paid users.

CEO Sam Altman confirmed that GPT-5’s personality will be revised. “We are working on an update to GPT-5’s personality which should feel warmer than the current personality but not as annoying (to most users) as GPT-4o,” he wrote on X. 

He added that one major takeaway from the rollout is the importance of “more per-user customization of model personality,” allowing individuals to tailor the AI’s tone and style to their preference.

OpenAI has also introduced three distinct modes for GPT-5: Auto, the default setting for general use; Fast, prioritising speed; and Thinking, designed for deeper reasoning with a context window of 196,000 tokens and a weekly message limit of 3,000 for paid users. After hitting this limit, users can continue on GPT-5 Thinking mini, a lighter version with reduced compute requirements.

In addition to restoring GPT-4o, OpenAI has made several older models accessible via the “Show additional models” toggle in ChatGPT settings. These include GPT-4.1, GPT-03, and GPT-5 Thinking mini. GPT-4.5 is still reserved for Pro users due to its heavy GPU demands.

Altman noted that the quick reinstatement of these models reflects the emotional attachment users have developed to specific AI personalities, an aspect OpenAI had underestimated. 

The company plans to allow even more granular personality customization, with potential options ranging from “Cynic” and “Listener” to “Robot” or “Nerd,” along with the possibility for fully custom tones and behaviours.

“These updates mark a shift in AI development,” Altman added. “It’s no longer just about raw capability; user experience and emotional resonance are becoming central. We’re not always going to get everything on try #1, but I’m proud of how quickly the team can iterate.”

OpenAI is working to balance powerful AI functionality with a more engaging and personally tailored user experience.

The models, Gemini Robotics and Gemini Robotics-ER, are based on Google’s Gemini 2.0 AI framework and are expected to enhance robots’ ability to perform tasks in industrial, commercial, and even domestic environments.

Gemini Robotics is a vision-language-action (VLA) model that allows robots to understand commands, interpret visual data, and execute physical actions. 

This means robots built with this model will be able to respond to spoken instructions, recognise objects, and interact with their surroundings more intuitively. 

The second model, Gemini Robotics-ER, takes this further by incorporating spatial reasoning abilities, enabling robots to understand complex environments and adjust their movements accordingly.

According to Google, these models have been designed to work with robots of various shapes and sizes, from humanoid robots to industrial machines commonly used in factories and warehouses. 

The company has already tested Gemini Robotics on its ALOHA 2 bi-arm robotic platform, demonstrating its ability to handle tasks requiring precision and adaptability. 

The model has also been successfully applied to Apptronik’s Apollo humanoid robot, which is being developed for real-world applications.

Google’s focus on robotics follows the recent decision by Figure AI to end its collaboration with OpenAI after making its own innovations in AI-powered robots. With the introduction of Gemini Robotics and Gemini Robotics-ER, Google is taking hold of the robotics industry.

One of the advantages of these models is their ability to reduce development costs for robotics startups. Companies can speed up the time it takes to bring functional robots to market when they leverage Google’s AI-powered frameworks. 

The tech giant has also stressed the flexibility of its models, stating that developers can customise them for specific robotic applications using Gemini’s advanced reasoning capabilities.

Google has partnered with Apptronik to integrate its AI models into humanoid robots, aiming to create machines capable of performing tasks that require both cognitive reasoning and physical dexterity. 

In August, Apptronik raised $350 million in a funding round led by B Capital and Capital Factory, with Google also participating to support the development of next-generation humanoid robots.

This is not Google’s first move in robotics. The tech giant previously acquired Boston Dynamics in 2013, a company famous for its dog-like and humanoid robots, before selling it to SoftBank in 2017. 

However, the launch of these new AI models is a renewed push into robotics, with Google focusing on AI-driven intelligence rather than hardware development.

With the integration of multimodal AI into robotics, Google aims to bridge the gap between digital intelligence and real-world functionality.

Hence, the increasing demand for automation in manufacturing, logistics, and service industries, will ensure Gemini-powered robots make work even more seamless.

Nonetheless, as AI systems progressively infiltrate essential areas of society, the industry has started taking note that accuracy alone is never enough.

A budding body of research opined that AI models, in spite of their impressive capabilities, can exacerbate and even intensify biases present in the data they are trained on.

These problems are not just random technical difficulties, but an ethical concern. The importance of mitigating bias in AI models to ensure fair and justifiable outcomes in real-case scenarios cannot be overstated.In practical terms, this involves several key steps and methodologies to identify, address, and reduce bias throughout the AI development lifecycle.

One of the pertinent challenges I have tackled  is the problem of data bias. Data bias happens when the training data for an AI model showcases reoccurring inequalities or prejudices, resulting in biassed results.

For example, a credit scoring model might be unintentionally biassed towards social groups from socio economic background if the training data majorly includes such groups.

I have been useful in the aspect of identifying these biases and integrating remedial actions. I advocate for robust methods for data collection, guaranteeing diverse data for AI training. This entails not only curating a variety of data sources while evaluating their quality and relevance.

In addition to data bias, I discussed extensively the role algorithms play in fairness. Despite having impartial data, AI models have the potential to yield biassed outcomes if the algorithm is not developed with fairness considerations. Even with unbiased data, AI has the ability to produce results if the algorithms themselves are not tailored with fairness in mind.

As a front runner in the tech industry, I have championed the use of equity focused algorithms, which are specifically designed to reduce uneven effects on diverse groups. This is important in the fintech industry because they make decisions based on AI system output for example loan approvals or fraud detection can have deep impacts on individuals life.

Another important aspect of my work is the consistent monitoring and evaluation of AI models after deployment. I genuinely believe that an AI model’s fairness cannot be fully achieved at the development stage only.

Once deployed, these systems engage with intricate and real world settings, possibly leading unanticipated biases. The team I oversee at interswitch have integrated complex monitoring frameworks to identify the performance of AI models in real time.

They evaluate key metrics to identify any signs of bias and take swift corrective actions when needed. This methodology ensures that the models maintain fairness and equity throughout their lifespan.

My commitment towards ethical AI transcends technical driven solutions, I am an advocate of accountability and transparency in AI development.

I believe that organisations have an ethical duty to explain how their systems operate and decision making processes behind them.

This transparency promotes trust among individuals and stakeholders, making it simple to address concerns related to fairness and bias.

Moreover, I actively engage academic institutions, industry leaders and regulatory bodies to ensure best practices in AI are maintained.

I often share my knowledge in conferences, workshops and panel sessions on how to reduce AI bias. My contributions to the tech industry have been broadly acknowledged and I have also become a revered authority in the discussion of ethical discussion.

My role at Interswitch showcases a comprehensive strategy in AI development, one that transcends beyond conventional accuracy benchmarks.

My commitment to evaluating and reducing bias in AI models guarantees that these systems yield fair and just outcomes in practical implementation.

The groundwork has been laid for a future where AI serves as a tool for positive social change,rather than a perpetuator of existing differences.

This effort highlights the potential of AI to bridge gaps, enhance accessibility, and drive equitable outcomes across diverse communities. The idea has shifted towards leveraging AI technologies to empower marginalised groups, ensuring that the benefits of AI are shared widely and equally.

About the writer:

Folashade Oluwatosin is a Senior Data Scientist with expertise in advanced data analytics, machine learning, and statistical modeling. She has successfully implemented data-driven solutions in various fintech and automobile companies, enhancing operational efficiencies and customer experiences. Known for her proficiency in scientific tools like Python, R, and SQL, Folashade excels in transforming complex data into actionable insights. Her strong leadership abilities have enabled her to lead cross-functional teams, driving innovation and fostering a culture of continuous improvement.

In a statement titled “Pause Giant AI Experiments: An Open Letter,” 1,332 signatories, including Musk, and Apple’s co-founder co-founder Steve Wozniak, called on AI labs to pause the training of advanced AI systems for at least six months.

The letter, which was published via Future of Life Institute, sounded a note of warning regarding the possible costs of the continued creation of “AI systems with human-competitive intelligence”

The signatories raised the alarm that AI systems were currently poised to automate human jobs, flood information channels with untruths, and ultimately replace the human race.

The statement read in part, “Contemporary AI systems are now becoming human-competitive at general tasks, and we must ask ourselves: Should we let machines flood our information channels with propaganda and untruth? Should we automate away all the jobs, including the fulfilling ones?

“Should we develop nonhuman minds that might eventually outnumber, outsmart, obsolete and replace us? Should we risk loss of control of our civilization? Such decisions must not be delegated to unelected tech leaders.”

They further argued that the development of powerful AI systems should only be considered when their creators “are confident that their effects will be positive and their risks will be manageable.”

“This confidence must be well justified and increase with the magnitude of a system’s potential effects.

“Therefore, we call on all AI labs to immediately pause for at least 6 months the training of AI systems more powerful than GPT-4.

This pause should be public and verifiable, and include all key actors. If such a pause cannot be enacted quickly, governments should step in and institute a moratorium