Patreon is reducing its workforce by 20%, which equates to 93 employees, as announced by CEO Jack Conte to staff on Thursday. In a memorandum that the company made public, Conte expressed that while Patreon’s primary operations remain strong, the platform must adapt to market shifts and modify its cost framework to maintain stability, hence the need for these “painful” but essential layoffs.

Conte stated that “AI has drastically altered the technology sector,” and the rate of change has “never been more rapid.” Nevertheless, he emphasized that Patreon is not implementing these layoffs to replace employees with AI.

“To clarify the role of AI in today’s decision: we are not enacting the above changes because we think AI can replace human workers,” he noted. “The more we have embraced these new tools, the more evident it has become that they cannot match the creativity, judgment, attention to detail, or craftsmanship that our team possesses abundantly, nor do they satisfy the human connection that we all value deeply. That’s my personal viewpoint, but more significantly, it forms the basis of Patreon’s strategy: our product vision and business are fundamentally based on the worth of human creativity and connection.”

He added, “AI has indeed transformed the tech landscape, including our modes of work, product development, communication, and beyond. This does influence our operations and organizational structure.”

Alongside the reduction in workforce, Conte stated that Patreon is also reorganizing its operational structure, “flattening” its hierarchy and realigning teams with its primary objectives. Those affected will receive a minimum of 16 weeks of severance pay, plus an extra week for each year of service, healthcare coverage until year-end, and a $1,500 allowance to replace their last company-issued laptop.

Last week, Patreon revealed its collaboration with internet infrastructure provider Cloudflare to proactively prevent AI bots from accessing creators’ work without consent to train their AI models. The company indicated that it needed to bolster its defenses in this area as AI scraping methods have evolved. This action comes as online publishers and creators face challenges from AI companies utilizing their work for training AI models.

The most recent layoffs at Patreon mark the largest reduction since a 17% staff cut in 2022. In that previous round, Patreon also shuttered its offices in Berlin and Dublin.

Corgi, a startup specializing in insurance technology, data room solutions, and coffee shops, is reportedly in the process of raising another funding round that closely follows its previous fundraising and would effectively double its valuation, according to sources cited by Forbes.

This round is anticipated to be a second extension of its Series B funding and has already been finalized. Corgi disclosed its last funding round, the B1 round, which amounted to $106 million at a valuation of $2.6 billion, just eight weeks ago at the end of May.

Amidst the current surge in AI funding, numerous startups are securing consecutive rounds at escalating valuations — yet Corgi has managed to stand out even in that crowded field.

The Y Combinator graduate (summer 2024) secured a $108 million Series A in January at an undisclosed valuation. (PitchBook approximates it at $630 million post-money.) Four months later, it raised its Series B with $160 million at a valuation of $1.3 billion. Just three weeks post that, it revealed a B1 round where the same investors contributed $106 million at a $2.6 billion valuation.

Now, sources inform Forbes of a B2 round emerging eight weeks later. However, Forbes did not disclose the amount that Corgi raised, and the company opted not to comment on the potential funding.

Corgi is supported by TCV and Kindred Ventures. Kanyi Maqubela from Kindred highlighted the startup’s progress to TechCrunch as a rationale for the recent valuation increase, attributing it to the company’s revenue growth.

When Corgi disclosed its Series A seven months back, the founders indicated that the firm had already achieved a $40 million annualized revenue rate. Sources have informed Forbes that it’s now positioned to elevate that run rate to $450 million by the year’s end.

Corgi offers AI-driven insurance solutions, leveraging AI to provide rapid quotes and expedite claims payments. It provides various forms of liability insurance for startups, including general liability, technology-related incident coverage, and employment liability, alongside business renters’ and auto insurance.

While insurance is typically a capital-intensive industry, Corgi faces additional challenges due to its use of a Risk Retention Group (RRG) structure. This method allows individuals in similar sectors or facing like liabilities to pool their resources for collective self-insurance.

As stated on Corgi’s website, RRGs are not governed by all the same state regulations that traditional, rated, underwritten insurance carriers are subject to. According to a spokesperson, Corgi has employed various structures for different insurance types; some policies may utilize state-regulated carriers as an example.

However, in an RRG insurance framework, claims are disbursed from the pooled funds, meaning a significant claim can diminish the resources available for settling other claims. RRGs lack the backing of state guaranty funds; if the pool fails to cover claims, the members absorb the losses. Large claims can even lead to the bankruptcy of the RRG.

Therefore, it’s perhaps understandable that Corgi seeks to enhance its financial reserves.

Despite the expansion of its insurance offerings, the startup has also diversified in other aspects. Corgi has introduced data room software, successfully navigating a recent controversy regarding the software’s coding.

Additionally, the startup operates two 24/7 coffee shops with charming and sometimes sponsored drink names like “Brexspresso.” Located in San Francisco and Atlanta, Corgi plans to open five more locations soon, including several in New York and one in London. Establishing physical cafe locations also entails significant funding.

In the meantime, Corgi has developed a reputation within Silicon Valley for its demanding corporate environment, following remarks from founder-CEO Nico Laqua regarding his expectations for employees to work seven days a week.

U.S. authorities are set to explore the creation of new rules for car manufacturers to guarantee that drivers and passengers can exit their vehicles safely. This newly proposed rule-making process, detailed in a regulatory submission and initially reported by Bloomberg, is a response to several incidents, including fatal ones, where individuals became trapped inside cars featuring flush, electronically controlled door handles similar to those on Tesla models.

The National Highway Traffic Safety Administration (NHTSA) announced this in reply to a petition urging the agency to initiate a safety defect inquiry regarding the emergency mechanical door release of 2022 Tesla Model 3 vehicles. The petition asserted that the door release mechanism does not meet federal motor vehicle standards.

NHTSA dismissed the actions sought by the petitioners, indicating that the matter would be better resolved through rule-making instead of a defect investigation.

Should the agency implement new regulations, all manufacturers will be required to adhere to them. Nonetheless, it is crucial to understand that the act of “commencing” rule-making does not guarantee the formulation of new rules, according to NHTSA.

“NHTSA received a petition for rulemaking to mandate an emergency door egress system to mitigate the risk of occupants being trapped in vehicles during electrical outages, and the agency approved that petition,” the agency elaborated in an emailed statement. “Approval of a rulemaking petition does not signify a decision by NHTSA to issue a new safety standard, but initiates the process of collecting input for an informed decision. NHTSA will compile further information and data, and seek public feedback, to ascertain if new safety standard rule-making is warranted under the Vehicle Safety Act’s guidelines.”

This action follows less than a year after NHTSA began investigating Tesla’s door handles due to nine reports from owners who were unable to access their cars, occasionally with children still inside. The investigation developed after Bloomberg conducted its own inquiry into multiple incidents where Tesla drivers and passengers found themselves trapped in their vehicles post-collision.

Although Tesla vehicles include manual door releases, these are solely accessible from within the car. An initial assessment by NHTSA indicated that the handles might fail if the electronic locks don’t receive adequate voltage from the vehicle’s battery system.

Tesla designer Franz von Holzhausen stated last year that the company is redesigning its door handles. Rivian announced last year that it would be modifying the interior door handle design on its R2 SUV to position the manual release in a more prominent location, nearer to the electrically powered door handles.

This article has been updated to incorporate a statement from NHTSA.

In the last year, Meta has financed the establishment of at least twelve natural gas power stations, including a single project that will utilize enough natural gas to produce as much electricity as the entire state of South Dakota consumes.

Today, Meta disclosed to TechCrunch that it has withdrawn from the RE100 initiative, a corporate renewable energy program, after a decade of involvement. According to a spokesperson from Meta, the decision to part ways was mutual.

This withdrawal concludes months of Meta increasing its investments in fossil fuels to power its AI data centers and raises the crucial question: What does “clean energy” signify for a company that continues to construct gas plants while still identifying as renewable?

The RE100 initiative, a project by the Climate Group, a nonprofit based in the U.K. co-founded by former Prime Minister Tony Blair, offers policy and technical assistance to companies aiming to switch to 100% renewable energy. Competitors of Meta, such as Apple, Google, and Microsoft, remain part of the group’s 444 members. Recharge News was the first to report Meta’s exit.

While Meta did not elaborate on the reasons for its departure — and the Climate Group did not respond to TechCrunch’s request for comment — the nonprofit has recently revised its guidelines for companies, enforcing stricter reporting on their progress towards renewable energy objectives. Previously, Meta had informed RE100 that it would “operate its entire operations on renewable electricity by 2020.”

Like many technology firms, Meta’s adoption of AI has driven it to secure substantial energy for its data centers, and while the company continues to acquire renewable energy, it has increasingly turned to natural gas.

Meta’s initial move was a 200-megawatt behind-the-meter gas power plant in Ohio, announced in June of the previous year, designed to supply power to one of its data centers.

Two months later, Meta revealed plans to construct three large natural gas power plants in Louisiana to provide electricity to its Hyperion data center. Then in April, the firm announced it would fund seven additional natural gas power plants for the same project. Together, these ten plants will produce 7.5 gigawatts, generating enough electricity to power South Dakota and more.

A spokesperson for Meta informed TechCrunch that the company remains dedicated to aligning its data center energy usage “with 100% clean and renewable energy.”

That’s a significant commitment. While natural gas burns cleaner than coal, it still emits considerable pollution. A single 1-gigawatt data center operating continuously, powered solely by natural gas, will emit 438 metric tons of nitrogen oxides, 149 metric tons of fine particulate matter, 61 metric tons of sulfur oxides, and 298 metric tons of carbon monoxide. These pollutants are linked to various health problems, including asthma, cancer, cardiovascular conditions, and dementia, among others.

Meta can still assert that it is 100% renewable by acquiring environmental attribute certificates. These certificates enable businesses to invest in solar farms in Arizona, for example, while constructing a data center in Ohio. As long as the solar farm generates enough energy in one year to counterbalance the data center’s consumption, Meta reports that as 100% renewable.

Most companies have approached their renewable energy goals through annual matching, but some, like Microsoft, are aiming to align their electricity consumption on an hourly basis. This more rigorous method would better synchronize power generation with data center energy usage. It also motivates companies to invest in projects that combine renewables with battery storage, similar to Google’s earlier efforts in Minnesota, rather than relying on polluting sources like Meta’s Hyperion power plants.

Meta is not alone in its pursuit of natural gas — both Google and Microsoft have recently invested in significant fossil fuel initiatives — but it has made the largest investment. While stepping back from a voluntary industry organization isn’t usually major news, the timing, amidst Meta’s expansion into fossil fuels, makes this shift difficult to overlook.

Following OpenAI’s launch of a new suite of conversational models and enhancements to ChatGPT’s voice functionality, competitor Anthropic is taking steps to enhance Claude’s voice capabilities with a fresh update. The company announced on Thursday that users will have the option to select from Opus, Sonnet, and Haiku models.

Claude’s voice functionality, which debuted last year and was based on the Haiku model, delivered rapid replies but fell short for intricate tasks. The company indicated that with this latest update, the voice mode defaults to the last model used by the user in the text chat, employing its quickest version.

Anthropic mentioned that the revamped voice mode aids users during extended dialogues, such as offering critiques on their communication techniques, discussing a proposal with a client, and conducting brainstorming sessions for product market research.

Additionally, Claude’s voice mode can interface with various applications such as Gmail, Google Calendar, Slack, Canva, and Notion. This allows users to request updates for meetings, compose emails, or generate documents within Notion.

Importantly, this marks a significant distinction from OpenAI’s voice mode, which refreshed its conversational approach but remains unable to utilize external tools for task completion.

Earlier this year, Anthropic introduced multilingual capabilities to Claude’s voice mode in beta form. They stated that users can now converse in multiple languages, though manual selection of the language is necessary. Currently, Anthropic accommodates English, French, German, Hindi, Indonesian, Italian, Japanese, Korean, Portuguese (Brazilian), and Spanish (Latin America/Spain).

The updated voice mode is accessible to all users in beta across different platforms. However, free users will be limited to the Haiku model with only one app connection available.

Anthropic has not altered the voice model with this update, nor have they specified the type of voice infrastructure it employs. Consequently, unlike OpenAI’s release, users may not experience enhancements in conversational aspects such as improved interruption management.

Cybercriminals are increasingly harnessing AI to execute attacks on a large scale, with email becoming a key target. AI can swiftly compile personal details — such as data about colleagues, ongoing projects, and recent travel plans — enabling malicious actors to promptly create convincing communications that appear legitimate.

In the previous year, former Google security leaders Cy Khormaee and Ryan Luo, who had previously been involved in creating safe browsing technology and reCAPTCHA, joined forces to establish AegisAI, a startup utilizing AI agents to combat these threats, referred to as spear phishing.

With a decade’s worth of experience in thwarting email breaches, the AegisAI co-founders recognized that current rule-based systems for preventing breaches — which depend on “if-then” mechanisms — are insufficiently fast and are limited in identifying AI-generated malicious emails. Thus, they created AI agents that rapidly assess each message as a human would, focusing on subtle discrepancies that even the most thorough checklists would overlook.

Fewer than twelve months post-launch, AegisAI reports that its technology has been embraced by numerous clients, including the cryptocurrency payments firm Mesh, AI startup LangChain, and the privacy compliance service Lokker. This surge in interest has enabled AegisAI to secure a $36 million Series A funding round led by Battery Ventures, with contributions from current investors Accel and Foundation Capital. The new capital elevates the startup’s total funding to $49 million.

“AI-driven attacks bypass current safeguards more than half the time today, signifying they’re nearly twice as effective as previously seen,” Khormaee disclosed to TechCrunch. “They’ve done their homework on you, they’re aware of everything about you, and they’re executing attacks tailored specifically to you.”

Khormaee asserts that AegisAI’s agents can identify threats that conventional email security systems might entirely overlook. For example, the startup’s AI can detect harmful PDF attachments that initially seem credible, including those with embedded passwords and CAPTCHA, commonly used to deceive standard spam filters.

When Dharmesh Thakker, general partner at Battery Ventures, observed a rise in email attacks, he aimed to invest in a startup capable of countering AI with AI, which aspires to replace traditional email security tools with agent-driven defense mechanisms.

“The adversaries are exploiting email to assault us using AI at a speed far exceeding our ability to respond,” Thakker stated to TechCrunch. “Countering that will become a top priority for numerous companies.”

AegisAI is not the sole startup leveraging AI to examine the context of every incoming email to identify fraud and impersonation attempts. Lightspeed-backed Ocean is also working to challenge established firms like Proofpoint and Mimecast, alongside newer contenders like Abnormal Security.

Nevertheless, considering that AegisAI is directed by specialists who played a role in securing Gmail, the most widely-used email platform globally, Thakker is confident that the startup has the best opportunity to emerge as the leading new hack-prevention entity.

While AegisAI is currently focusing on email, the startup aims to eventually branch out into other defense sectors, such as data security. “The fundamental concept of creating personalized, highly sophisticated agents capable of conducting investigations is going to [influence] who becomes the next dominant security firm,” Khormaee mentioned.

The U.S. government is alerting that hackers supported by the Iranian state are currently infiltrating and disrupting industrial control systems at American water and energy companies. This recent warning comes after federal agencies indicated a rise in hacking from Iranian entities amid the ongoing conflict.

In an advisory released Wednesday, the FBI, NSA, Department of Energy, and CISA reported that Iranian hackers are focusing on programmable logic controllers within internet-connected operational networks, enabling them to alter data displayed, leading to outages and disturbances.

Earlier this year, the Iranian hackers were found to be targeting Rockwell-made controllers, but the advisory has since broadened to include industrial control systems from Schneider Electric and Siemens.

The agencies cautioned that “potentially all internet exposed” industrial control systems might be impacted and encouraged critical infrastructure owners to take preventive measures. According to the advisory, the Iranian-affiliated hackers were “conducting this activity to cause disruptive effects within the United States,” likely as a reaction to the ongoing conflict involving Iran, the U.S., and Israel.

The FBI stated that the hackers gained access to one critical infrastructure provider and altered the programming logic of the controllers to disable processes responsible for crucial shutdowns and alarms. The authorities noted this permitted “systems to enter unsafe conditions without alerting operators to the discrepancies.”

This is the latest in a wave of cyberattacks executed by Iranian government hackers and their affiliates throughout the region since the conflict began in February. 

The hacks have varied from the typical espionage and hack-and-leak strategies, such as exposing the personal email account of FBI director Kash Patel, to more unusual destructive hacks that have resulted in extensive damage or disruptions. Among the more significant incidents was an attack on the U.S. medical technology company Stryker, which enabled the Iranian hacking group “Handala” to remotely erase tens of thousands of employee devices.

Handala also claimed responsibility for a data breach involving California’s water supplier Cal Water in June, asserting it could have compromised the water supply (without providing proof). The water supplier stated it found no indication of unauthorized access to its operational networks, which manage the water supplies.

Runway aims to transcend being merely an AI model enterprise and is striving to serve as the foundational layer for generative media.

On Thursday, the company introduced Runway Media Router via Runway Dev, its developer platform, which was unveiled earlier this month and offers API access to a growing selection of third-party image, video, and audio models as well as Runway’s proprietary ones. 

The Media Router functions as a tool that automatically identifies the optimal image, video, or audio generation model for a query, depending on whether a developer emphasizes quality, speed, or cost. While model routers have become more prevalent among large language models, Runway asserts this is the first specifically designed for generative media. 

“The routing neatly aligns with our overarching promise of providing the most user-friendly one-stop solution for developers to seamlessly integrate any generative media model,” commented Anthony Maggio, Runway’s chief product officer, to TechCrunch.

The announcement, shared exclusively with TechCrunch, signifies another progression in Runway’s transformation from an AI video startup into a foundational resource for companies engaging in generative media. Through Runway Dev, clients such as Adobe, Cloudflare, ElevenLabs, Expedia, Shutterstock, and Quora can directly embed media generation into their own offerings by utilizing Runway’s API, rather than directing users to Runway’s own application or website.

The router’s release coincides with a significant increase in the availability of generative media models, making it more challenging and time-consuming for developers to assess new offerings. With the Runway Dev platform, developers gain access to the latest media models upon their launch. 

“Most developers aren’t dedicating time to thoroughly comprehend the capabilities of these models or their strengths and distinctions based on various output types across video, image, and audio,” Maggio stated. “Our unique proposition lies in the intelligence regarding the best model for each specific use case, intertwined with the preferences you apply within your business context.”

Maggio mentioned the growing popularity of generative media models from China. However, many businesses developing their own products might hesitate to utilize models originating from China, potentially leading them to favor American model providers — a trend that could gain traction as the Trump administration considers bans and sanctions against Chinese open AI models. 

That’s merely one instance of the preferences developers can establish, though. Maggio notes that clients primarily focus on routing models to consider token pricing and quality. Token pricing has garnered significant attention in 2026 as enterprises that fully embraced agentic AI faced hefty token costs. Within LLMs, model routing based on token pricing has become standard, thus it stands to reason that routing for generative media would follow. The Media Router’s launch also follows weeks after Runway transitioned from unlimited subscription plans to token-based pricing, which prompted some user criticism. 

Determining which models deliver optimum quality for specific tasks is less straightforward in generative media than in language models, according to Maggio. This is where the intelligence layer of the router comes into play. It’s built upon the knowledge that Runway’s internal creative team has cultivated in assessing output across diverse media forms — such as how well video models manage motion, how image models address composition, or how voice models tackle lip syncing. 

Runway had already laid considerable groundwork for that intelligence layer in its agent product, a conversational AI creative collaborator launched in May aimed at transforming text prompts into fully edited multi-shot videos and marketing strategies. 

Maggio states that the Runway Media Router adopts the routing technology that Runway developed for its own products and adapts it for external developers to utilize.

Runway’s current strategy reflects the fragmented and competitive nature of the generative media market and the necessity for the startup to evolve and adapt to maintain its competitive edge. The last release of an AI video model from Runway — Gen 4.5 — occurred in December. At that time, the model topped the leaderboards, surpassing similar offerings from established players like Google. In the same month, Runway also introduced its first world model. 

Aside from an enhancement to its video editing model, Aleph 2.0, in May, Runway has not introduced a new dedicated frontier video model in several months. (TechCrunch has inquired about the startup’s timeline for the Gen 5 release.) Currently, while Aleph 2.0 is ranked among the leading video editing models according to Artificial Analysis, the company’s text-to-video and image-to-video models no longer hold top positions. The top 20 rankings feature models from major players like Google and China’s ByteDance and Alibaba. 

Rather than requiring developers to rely on the consistency of a single model, Media Router operates on the premise that the optimal model will continually evolve — keeping Runway actively engaged so that it can persist in advancing the frontier. If not as the best new AI model, then as the premier orchestration layer. 

Anastasis Germanidis, Runway’s co-founder and co-CEO, recognized that the startup has long been primarily associated with “that end user piece,” yet it needed to construct a comprehensive infrastructure to achieve that, which encompasses a developer platform, a creative toolset, and an inference layer beneath it all. He indicated that there has been a growing demand for Runway to be integrated throughout every segment of that stack. 

“Great models are essential, but orchestration is increasingly significant as people create entire campaigns with those models or generate complete multi-scene productions utilizing those models,” Germanidis relayed to TechCrunch. “This is an intelligence layer we increasingly had to build — one that complements the pixel models. The router is one method through which the advantages of that layer are made accessible to users.”

Or as Maggio more broadly articulated: “If you zoom out at the one thing Runway has been dedicated to since 2018, it’s that we’re profoundly focused on research while simultaneously building for the future of this space.”

Have a confidential tip or sensitive documents? Rebecca Bellan is delving into the inner workings of the AI industry, from the companies shaping its future to the individuals affected by their decisions. Reach out securely, and off the record, on Signal at rebeccabellan.491 or via email at [email protected] from a personal device.

OpenAI announced today the launch of ChatGPT Health, a feature designed to assist users with health-related inquiries, now accessible to all U.S. users aged 18 and over across all plans. This news follows a lawsuit by a pastor in Florida against the company for providing a potentially harmful suggestion to avoid consulting a doctor.

The company began pilot testing the feature through a specialized hub in January, enabling users to consolidate data from various services along with personal information from platforms like Apple Health, Function, and MyFitnessPal.

Previously, OpenAI reported that users submitted 230 million health-related inquiries each week. That figure has now risen to 300 million.

Users are also able to integrate their medical records from healthcare systems such as Epic and Oracle Health, along with platforms like One Medical and Function Health.

OpenAI previously stated that users were required to utilize the health hub for inquiries regarding health. Now, they have the flexibility to gain insights from connected data across all queries in the health section. During testing, the company found that 70% of health-related inquiries occurred outside the designated hub. With this new feature, users can leverage their health information to gather details about food or allergies in the general chat.

The company acknowledged that its models have improved in addressing health inquiries. It noted that the smallest model from its latest release, GPT 5.6-Luna, exceeds the performance of GPT 5.5 on the HealthBench evaluation, an open-source benchmark designed by the company to assess large language models (LLMs) on health queries. OpenAI reaffirmed that it does not utilize any user data for model training and collaborates with healthcare professionals to enhance its models in handling health inquiries.

Despite these advancements, the company emphasizes in its terms that its services are “not intended for diagnosing or treating any health conditions.” It referenced these clauses in response to the aforementioned lawsuit, additionally informing The New York Times that it is working on ensuring safer health or medical-related responses. With the recent rollout, OpenAI intends for individuals to verify information and make medical decisions based on professional guidance.

Numerous studies have pointed out that AI tools are unreliable for medical counsel. Nevertheless, this has not prevented companies such as Anthropic and Google from introducing health-oriented AI functionalities.

Health in ChatGPT is being released to logged-in U.S. users with free, Go, Plus, and Pro plans on the web and iOS this week.

According to a chart released by the company on Wednesday, Tesla’s emerging “Robotaxi” network provided fewer miles for paying customers in the second quarter compared to the first.

This decline from quarter to quarter contradicts Tesla’s narrative and efforts over the past year. The company has largely tied its future to the concept of an extensive, affordable, revenue-generating Robotaxi fleet — or going “balls to the wall for autonomy,” as articulated by CEO Elon Musk in 2024. The drop in Robotaxi miles also coincides with declining profits in Tesla’s main businesses, which fell short of Wall Street expectations, as shown by data released on Wednesday. Tesla’s shares plummeted over 13% in early trading on Thursday.

At first glance, the chart seems to indicate consistent growth in paid Robotaxi rides from August 2025 to June 2026. However, the figures presented are cumulative; when analyzed by quarter, it reveals that Tesla’s Model Y SUVs with paying passengers covered approximately 1.1 million miles in the first quarter, which decreased to about 700,000 miles in the second quarter, representing a drop of nearly 36%.

This occurs even as the firm has broadened its fledgling operation to six cities in Texas and Florida, utilizing a combination of supervised and unsupervised vehicles.

It is probable that Tesla is including the paid miles logged in the San Francisco Bay Area as well, even though these branded Robotaxis lack state-mandated permits for autonomous operation and have a safety driver present. Tesla has referred to this operation as part of its “Robotaxi coverage.”

The reduction in miles traveled also occurs as Tesla made a notable acknowledgment during a conference call on Wednesday regarding its second-quarter findings. When asked about the slow scaling of the Robotaxi service, Musk stated that the company must “gather driving data that is specific to the Cybercab” — the firm’s specially-designed, two-seater sedan that is anticipated to constitute the majority of its autonomous vehicle inventory — “before we can deploy a large number of them on the roads.”

“In contrast to, say, Model 3, Model Y, and our other vehicles where we have a considerable number on the roads, millions of vehicles on the roads, we don’t have that for Cybercab. Therefore, we need to accumulate miles with Cybercabs that are modified with steering wheels and accelerator and brake pedals, among other things, to calibrate to the Cybercab chassis,” he explained. “As we gain confidence with that, the count of Cybercabs in cities will rise significantly.”

This reflects a deviation from the company’s long-standing assertions about how its fleet of nearly 10 million customer vehicles has been quietly gathering data in the background to prepare for future robotaxis (alongside training the driver assistance software for consumers, which Tesla dubs Full Self-Driving).

During the call, Tesla’s executives framed the sluggish progress as a pressing need for safety caution.

“Our ambitions for Robotaxi are quite high, but we must exercise caution to avoid any accidents or harm to individuals,” Musk remarked.

He then expressed his concern that he does not want Tesla Robotaxis involved in accidents, as he believes negative media coverage could prompt a regulatory response.

“Although there are, I believe, 30 to 40,000 automotive fatalities each year in the United States, most do not make the news; you rarely hear about any of them. However, if we harm just one person, it will dominate the headlines globally, and regulators will act swiftly against our operations,” he stated.

Ashok Elluswamy, Tesla’s VP of AI, claimed that Tesla’s Robotaxis have experienced “zero notable incidents” while traveling “over 380,000 miles” without a safety operator present. He did not clarify what the company defines as “notable incidents,” although he asserted that “any reports were of other actors impacting us while we were stationary.”

Tesla has reported 22 incidents to the National Highway Traffic Safety Administration since it began trialing its Robotaxi service. While most of these incidents involve other vehicles colliding with Tesla’s Robotaxis, the company has noted three instances attributed to its teleoperators maneuvering the vehicles remotely, along with several cases of the cars striking objects at low speeds, including curbs, utility poles, and a tow truck’s bed.

This marks yet another shift in the company’s story. Tesla spent years insisting that regulatory barriers were the primary obstacle to full-scale Robotaxi rollout — although the company never specified which regulations were considered prohibitive.

Now, the company claims that proving safety is what is hindering Robotaxis’ progress. And although it is still in the very early phases, Tesla has opted to take this opportunity to celebrate its decision to develop an autonomy system that does not rely on radar or lidar sensors, unlike the industry leader Waymo.

“Historically, so-called experts have always asserted that you need lidars, radars, HD maps, and an entire array of tools to drive safely. Here, we demonstrate that this is not the case. You can achieve safe, comfortable, and affordable autonomy using only cameras,” Elluswamy stated.

Both Musk and Elluswamy pledged that growth is on the horizon. They noted that the amount of unsupervised miles driven has increased by approximately 10% each week since Tesla began offering them at the end of last year.

“We will continue to scale, I believe, very, very rapidly,” Musk remarked.