Too Fast: Google recently announced an accelerated Chrome release schedule. Now, another major player in the browser market, Mozilla, is doing the same. The company plans to drop two updates for its open-source browser every month across both desktop and Android devices. The new roadmap is expected to begin in September 2026, although Mozilla could change course if unexpected issues arise.

Mozilla Director Sylvestre Ledru confirmed that Firefox will soon adopt a faster release cadence, but said the accelerated release schedule should be considered an “experiment” for now. The final monthly release will be Firefox 154, which is currently scheduled for August 18. After that, Firefox 155 will arrive on September 1 instead of two weeks later. As a result, Mozilla has completely overhauled the Firefox Release Calendar to accommodate the new bi-weekly schedule.

Ledru said the faster release cadence will allow Mozilla developers to ship more “work” on the browser more frequently. Meanwhile, the release process itself should become more predictable, with fewer obstacles affecting rollout plans. Developers will not need to work twice as fast, however, as half-baked features will still (hopefully) receive the additional attention they need.

Ledru’s team will monitor this major change to the Firefox release schedule, adjusting the new approach when necessary. Google announced a new bi-weekly release schedule for Chrome in March, explaining that the change would ultimately benefit all parties by improving the security, reliability, and performance of the web platform.

Last month, Microsoft announced that the Edge browser would adopt the same bi-weekly release cadence as Chrome. This was a much easier decision to make, considering that Edge is essentially a proprietary shell built on top of the Chromium project. Meanwhile, Mozilla continues to use and develop its Gecko layout engine despite Chromium’s growing popularity among browser makers.

Over the past decade, Mozilla has been forced to follow Chrome’s design decisions time and time again. Some of those changes were not exactly popular among Firefox enthusiasts, starting with the more restrictive WebExtensions model for browser add-ons.

Judging by the early reactions to Ledru’s message, the accelerated release cadence could prove to be controversial as well. Firefox users are once again accusing Mozilla of copying Google’s every move, even as the Alphabet-owned company continues to make one questionable decision after another. Other users have pointed out that Firefox is now going to break things twice as quickly while Mozilla attempts to keep pace with Google’s apparent obsession with Chrome’s version numbers.

Everyone my age knows what the stereotypical ‘robotic’ voice is. They changed it because they wanted to hide the fact you were talking to a machine.

We all know that a mouse moving in a perfectly straight line means a machine is controlling it, while humans do something more like a squiggly line. Basically a normal human drawing a line looks like someone with Parkinsons did it as compared to what a machine drawing a line looks like.

Similarly, humans typing have pauses that tend to end after set thoughts. N

Model routing is becoming a key component of the enterprise AI stack, dynamically sending prompts to the right AI model to optimize speed and costs. However, current frameworks mostly treat routing as a static classification problem, which severely limits their potential.

A new open-source framework called Agent-as-a-Router tackles this bottleneck, treating the router as a dynamic, memory-building agent. It uses a Context-Action-Feedback (C-A-F) loop to track model successes and failures and update the behavior of the router. 

The researchers also released ACRouter, a concrete implementation of this paradigm. In their tests, ACRouter significantly outperformed static routers and the expensive strategy of defaulting to premium models, all without requiring teams to train massive models or write endless heuristics.

For real-world applications, this framework provides the option to replace hard-coded AI infrastructure with self-optimizing systems that can adapt to changes in user behavior and foundation models used in the enterprise AI stack. 

The economics of routing and the information deficit

Single-model setups are useful for experiments but detrimental when scaling AI applications. AI engineers use model routing to map tasks to cheaper and faster open models when possible, while reserving expensive frontier models for complex reasoning. 

Currently, developers rely on two main mechanisms for this task. The first is heuristics-based routing, which relies on hard-coded manual rules. For example, a developer might write a rule dictating that if a prompt contains certain keywords, it is routed to GPT-5.5. Otherwise, it goes to a self-hosted open source model like Kimi K2.7. 

The second mechanism is static trained policies. These are machine learning classifiers trained on historical datasets that look at the prompt’s embeddings and predict the best model based on past training data.

Both approaches are static. When the researchers tested these existing mechanisms on real-world coding and agentic workflows, they found a hard ceiling on accuracy. The key finding shows that static routers suffer from a severe information deficit. Because they only evaluate the input text and never see if the model actually succeeded in executing the task, they guess blindly when faced with complex edge cases.

This results in three distinct points of failure. First, static routers suffer from a frozen information state, meaning they cannot accumulate new execution feedback during deployment. Second, they fail in out-of-distribution (OOD) generalization. They break down during day-two operations when enterprise data or user behavior shifts because their training data no longer matches reality. Finally, they are highly vulnerable to model churn. A static classifier trained on today’s models may become obsolete when a better model drops the following week.

Agent-as-a-Router: A self-evolving system

The core thesis of the Agent-as-a-Router is that a truly effective router must acquire and accumulate execution-grounded information during deployment, essentially learning on the job. 

The researchers achieved this through the C-A-F loop. When a new prompt arrives, the router examines the prompt and task metadata, such as the programming language or difficulty. It then searches its historical memory for similar tasks to see which models succeeded or failed in the past. The router uses this context to select the target model and execute the task. Finally, the system observes the real-world outcome, extracts a success or failure signal, and writes this feedback back into its memory to inform future routing decisions.

Consider an automated enterprise data analytics pipeline. The router receives a SQL generation task and sends it to an open-source model like Kimi. The model hallucinates a column name and fails to compile the SQL. The C-A-F loop observes the compiler error, registers it as feedback, and logs it. The next time a similar obscure SQL query arrives, the router checks its context and routes the task to a more advanced model like Claude Opus 4.8. 

ACRouter

The researchers developed ACRouter as the concrete instantiation of this framework. It is composed of three core components: the Orchestrator, the Verifier, and Memory. This architecture is supported by a tool layer to physically execute the C-A-F loop.

The Memory module powers the context phase. Built on a vector store, it retrieves relevant past interactions and updates the historical database with new outcomes. The Orchestrator handles the action phase. It processes the user prompt alongside the retrieved memory to select the most capable target model from the available pool. The Verifier manages the feedback phase by evaluating the chosen model’s output to generate a clear success or failure signal.

The tool layer hooks the Verifier into real-world execution environments, like a Python code interpreter, an agentic sandbox, or a database engine. The tool layer allows the system to execute the generated code or query and observe the exact outcome, providing the verifiable signal the router needs to learn.

The Orchestrator itself is lightweight. Instead of a massive, computationally heavy large language model, the researchers trained a sub-billion parameter adapter based on Qwen 3.5 (0.8B parameters), which means it can be self-hosted on a device of your choice.

ACRouter in action: Outperforming the frontier baselines

To stress-test the framework, the researchers introduced CodeRouterBench, an evaluation environment comprising roughly 10,000 tasks with verified scores across eight frontier models, including Claude Opus 4.6, GPT-5.4, Qwen3-Max, and GLM-5. The evaluation was split between in-distribution (ID) tests (covering nine single-turn coding dimensions like algorithm design and test generation) and an out-of-distribution (OOD) agentic programming testbed. The OOD tasks were qualitatively different, requiring multi-step planning, file navigation, and iterative debugging to see if the router could adapt to fundamentally new domains.

The baseline results revealed why a single-model strategy is flawed: no single model dominates every category. For example, while Claude Opus 4.6 achieved the highest average performance, it was outperformed in algorithm design by GLM-5 (an 86% relative improvement) and in test generation by Qwen3-Max (a 111% improvement), despite Opus costing roughly 12 times as much as smaller models like Kimi-K2.5. 

In the benchmarks, static routers continuously failed by sending a specific niche coding task to a model ill-equipped for that exact syntax. The static router had no way to know the code was failing to execute. In contrast, ACRouter adjusted its strategy after receiving negative feedback signal from the execution environment. 

According to the researchers’ benchmarking, ACRouter sits firmly at the Pareto frontier of cost and performance. On both the ID task streams and the complex OOD agentic tests, ACRouter achieved the lowest cumulative regret, a metric measuring sub-optimal routing decisions over time. On the in-distribution test set, ACRouter cost $13.21 across the full task run, compared to $34.02 for always defaulting to Opus — a 2.6x savings.

It dynamically matched tasks to the most capable model for that specific niche, suggesting that enterprises can achieve or exceed frontier-level accuracy across diverse workloads without paying a premium price for every query. 

Caveats, limitations, and how to get started

While the Agent-as-a-Router paradigm solves the information deficit, it is not a blanket solution for all AI workflows. 

The framework shines in verifiable tasks where the Verifier gets a clear success or failure signal from the environment, such as coding or data retrieval. It is effective for applications with distribution shifts and domains where different models excel in completely distinct niches. 

Conversely, the setup is overkill for trivial tasks where any model will suffice, or for low-volume applications that do not justify the engineering overhead. It is also unsuitable for subjective domains, such as creative writing, where a correct answer cannot be easily verified and feedback signals are impossible to standardize.

The researchers open-sourced the code on GitHub and released the orchestrator model weights on Hugging Face under the Apache 2.0 license. The router is compatible with Claude Code, Codex, and OpenCode.

Newton’s Cradles have long been staples on office desks, their five steel balls transferring motion in a satisfying chain of collisions. Most versions eventually lose their rhythm as energy dissipates through air resistance, sound, heat, and slight changes in the shape of the balls during impact. A recent project changes the outcome by adding a small, precisely timed push that replaces those losses. The result looks like the classic toy yet keeps swinging for hours on battery power alone.

The solution is based on some cleverness, with an inductive proximity sensor that is activated when one of the end balls begins to descend. The microcontroller then takes over and provides a short charge of current to the electromagnet located immediately below the ball’s path, just enough to give it a slight nudge and restore energy that would otherwise be lost. Timing is key in this situation; if you do it too soon or too late, the ball will either not get any help or will be yanked out of its natural swing. The builder began with two sensors for early experiments to obtain a strong grasp on the motion, but then reduced it to one sensor and some predictive math based on the swing rate, which appeared to function just as well with less hardware.

CERROPI Large Newtons Cradle Pendulum with 7 Balls, Physics Perpetual Motion Desk Toys for Office, Calm…

• SCIENTIFIC DESIGN：CERROPI newton’s cradle adopt scientific design,keep the pendulum balls perfectly lined up,make it possible to swing 50s+. Perfect…
• SUPERIOR QUALITY：Our newtons cradle made of 8mm chromed metal steel bar, 20mm steel Ball,high-impact nylon, ultra-thin nano technology provide…
• WIDE APPLICATION：This newton balls can be used as office desk gadgets toys to relieve stress; as teaching aid,teaching science intuitively; as toys…

The power comes from a rechargeable 18650 battery cell, which receives a voltage increase via a boost converter to give the electromagnet a solid, strong push. A MOSFET does the switching, while a diode protects the circuit from voltage spikes when the coil turns off. When it’s time to charge the battery, simply plug it into a USB-C socket embedded inside the base. There is an issue with ball weight changes, which might cause undesired vibrations in the intermediate balls. So the builder installed some small permanent magnets to provide a modest downward drag, helping to suppress resonance and keep the energy flowing smoothly from end to end.

The frame started off as a normal cradle that you could buy off the shelf. The main custom work is fitting the electromagnet and sensor into the base without changing the outward appearance too much. As it turned out, the build ended up looking like a cute little warship, complete with raised accents that conceal the battery, circuit board, and wires, as well as the charging port, which is tucked away under a magnetic cover.

Assembly begins by arranging the mechanicals by suspending the steel balls from the frame with strings or cables, and then attaching the electromagnet just below one of the end balls. Then you continue on to the electronics, aligning the proximity sensor to detect the ball’s approach, routing the wires to the microcontroller board, and soldering the power components in place. The microcontroller’s firmware detects the ball and calculates the pulse length, and then it’s just a matter of fitting everything into the enclosure, attaching the stabilizing magnets, and giving it a test swing to ensure everything works properly.
[Source]

Almost 60pc of young children across Europe experienced emotional or psychosocial problems online, according to the European Commission.

The EU is considering banning social media for children under the age of 13 years old, as child safety on the internet takes political priority across the world.

The bloc has, over the years, targeted social media giants for their addictive designs, recommender system feeds and child safety measures with its landmark Digital Services Act (DSA) that carries a hefty penalty for offenders.

In 2025, it preliminarily found that Meta and TikTok breached its rules by failing to meet its standards for keeping minors safe online.

Last week, the bloc preliminarily found that Meta did not assess the risks stemming from its addictive design on the physical and mental wellbeing of its users, including minors.

And in April, it found that Instagram and Facebook did not take effective measures to keep minors under 13 from using its platforms, despite Meta’s own terms and conditions stating otherwise.

Social media platforms generally require their users to be at least 13. However, these restrictions can be and are easily circumvented.

Mounting pressure to effectively address wider concerns surrounding the harms caused by social media on children’s developmental growth has led the EU to consider a similar strategy as Australia, Canada and the UK – as well as several of its own member states.

France, Sweden and Greece are currently proposing a social media minimum age at 15 years, while Denmark is considering a limit at 16. Others, including Portugal, Spain, Germany and Poland, are also considering legislating on the topic.

The European Parliament, late last year, called for a harmonised EU digital minimum age at 16, and suggested parental consent for those between 13 and 16.

But in its latest recommendations, an EU special panel on child safety suggests restricting social media for those under 13 until platforms demonstrate that their services are safe by design.

The bloc said it wants to address this issue at the highest level to avoid further fragmentation of the single market and to standardise the level of access children enjoy across the region.

The topic is clearly important to EU citizens, with a recent reporting finding that 92pc of Europeans believe protecting children online should be a priority for its leaders.

“The status quo, a world where we continue to allow Big Tech unrestricted access to our children, will only consign another generation to more mental harm, addiction and misery,” said Commission president Ursula von der Leyen today (13 July).

Her statements come as nearly 60pc of young children across Europe are believed to have experienced emotional or psychosocial problems online, according to the Commission.

Across Europe, leaders are taking a similar stance. Outgoing UK prime minister Keir Starmer said in June that tech giants “failed” to protect children online, while French president Emmanuel Macron told the media in January that “children’s brains are not for sale” after France adopted a bill banning social media for children under 15.

“The more we learn, and the more we see the impact on our children, the stronger the argument becomes for a social media start date,” the Commission president said, adding that the solution might not be foolproof and would take time.

The Commission is working on an “easy to use, privacy preserving, and open-source” age verification app to assist with its goals.

‘Age restrictions alone will not make digital spaces safe’

Commenting on the EU panel’s recommendations, South East Technological University’s Dr Dean McDonnell said: “Children’s digital media use is ecological, shaped by families, schools, peers, communities and the opportunities available offline.

“Any reduction in children’s access to digital spaces should always be matched by investment in safe public spaces, play, sport, arts, libraries, and youth services,” he told SiliconRepublic.com. “Removing one developmental environment without strengthening others can have significant consequences.”

McDonnell, a lecturer of psychology and a member of the Psychological Society of Ireland, stated that the most important thing that the report gets right is that “children are not a single group, and online safety is not a problem parents can solve alone”.

“A toddler, a primary-school child and a 17-year-old have very different developmental capacities, vulnerabilities and rights. The move from protection and supervision towards growing autonomy is sensible, although the broad 3-12 and 13-18 bands will need much finer guidance in practice.

“Its strongest recommendation is to shift responsibility towards the companies designing these environments,” he added. “Safety by default, age-appropriate settings and limits on attention-capturing features such as infinite scroll and recommender systems should not be optional.

“A common minimum age may provide a useful baseline, but age restrictions alone will not make digital spaces safe. Their value will depend on privacy-preserving age assurance, consistent enforcement and independent evaluation.”

Updated, 4.54pm, 13 July 2026: This article was amended to include a comment from South East Technological University’s Dr Dean McDonnell.

Don’t miss out on the knowledge you need to succeed. Sign up for the Daily Brief, Silicon Republic’s digest of need-to-know sci-tech news.

We are absolutely stoked to announce that the Hackaday Superconference is taking place this year November 6th through 8th in glorious Pasadena California, and we want to see you there!

If you’ve been to any of the previous nine Supercons, you know that it’s a fantastic gathering of the most motivated and interesting hackers around — but it’s also been a relatively small gathering. And while we love the very high signal-to-noise ratio of folks who show up, we’re always a little bit sad when the tickets sell out because it represents hackers who couldn’t be there.

So this year, we’re celebrating Supercon Ten by expanding out of our traditional location at the Design Lab so that we can accommodate 20% more hackers, while still keeping the cosy nature of the event intact. So if you’ve been wanting to come to Supercon, but procrastinated the ticket sales every year, this year is looking 20% better.

Call for Proposals

If you want to give a talk to an interested audience of hackers just like you, now is your chance. Fill out the Call for Participation form before Wednesday, Aug 12th to put your hat in the ring. Presenters not only get to share their work with a like-minded audience, but they get in the door free! Presenting really is the best way to attend a conference like this – it’s the ultimate ice-breaker. (Plus, did we mention free?)

We will have two tracks of talks on two stages, and both are a mix of shorter 20-minute talks and longer 40-minute sessions, so whatever the size of your ideas, we have the slot for you. As always, we like to hear about your projects: hardware, software, creation, destruction, or anything in-between. In short, if you have a talk that would interest the readers of Hackaday, it fits. Check out last year’s slate if you’re curious, but bear in mind that we like to see new stuff, so don’t feel constrained by precedent. If you’re into it, there’s a good chance that many of us are too!

All you need is an abstract, a title, and a solid general idea of how the talk is going to go. First time speaker, or grizzled veteran: get your proposal in now.

Plus ça Change…

Supercon Ten starts out as usual with a casual badge-hacking day at Supplyframe HQ on the morning of Friday Nov 6th. We love this day because there’s “nothing” to do! It’s the perfect way to ease into the conference: the doors open, and the food and coffee starts flowing. As the solder melts, brought-along hacks get demoed, friendships form, and plans get hatched. We go on well into the night, with music and festivities to keep you motivated or distracted – the choice is yours.

Saturday and Sunday are chock-full of talks, workshops, challenges, and other events. This year, we’ll be a few blocks south at the ArtCenter South Campus, which means that we’ll be relocating our traditional back-alley ambiance to significantly fancier digs. But of course, we’ll have space for hacking, mingling, and watching the talks.

Sunday evening comes too soon, and at the end of this second day of talks, we’ll let you showcase all of the badge hacks that you’ve been working on before spilling out into the town and falling far too late into bed.

Just because enough is never enough, we’ll probably also meet up informally sometime Thursday night if you’re already in town. And if you’re able to finagle a half-day Monday into your schedule, you’ll find that a bunch of folks have off-schedule side trips that are always popular.

Get Excited!

We know that we’re announcing late this year. The new venue, combined with a late Hackaday Europe, made for a lot more planning to be done. But now that all of our ducks are in a row, we’re very much looking forward to November. And of course, we can’t wait to see what you all are going to bring with you to Supercon. After all, it’s the Hackaday community that makes it great.

Get your talk proposals in now, and in the next few weeks, we’ll open up ticket pre-sales. Tell your friends, neglect to mention it to your enemies, and start making your Supercon plans today.

Cutting the West’s reliance on China would carry a staggering price. A new study puts the cost of decoupling from China at $23.6tn over 25 years, and warns the bill would land hardest on the industries building Europe’s tech future.

The West has spent three years talking about reducing its dependence on China. A new study tries to price it. According to an exclusive report in the Financial Times, the consultancy EY-Parthenon estimates that the US, the eurozone and the UK would need to invest an extra $23.6tn over 25 years to end their reliance on China in critical industries.

That figure covers manufacturing, technology, research, software and the supply chains beneath them. The wider effort to cut dependence on Chinese chips and rare earths has been building across Western capitals for months.

The burden is split unevenly.

The analysis puts the US bill at $13.7tn, the eurozone’s at $9.1tn and the UK’s at $800bn by 2050. Together that is roughly $940bn in extra spending every year. It would sit on top of what these economies already commit to energy, defence and infrastructure.

For the EU, EY-Parthenon says, the annual sum is close to doubling its entire budget.

A bill the size of the AI boom

The figure is easier to grasp against the technology it would help protect. The US share works out at about $550bn a year. That is close to the $600bn that large American tech groups poured into data centres in 2025.

Put another way, unwinding China from Western supply chains would cost roughly as much each year as the entire American AI build-out. And the materials at stake are the same ones that feed chips, electric cars and those data centres.

Why the materials matter

China’s grip is tightest where it is hardest to replace.

The International Energy Agency expects China to supply more than 60 percent of the world’s refined lithium and cobalt by 2035. It puts China’s share of battery-grade graphite and rare earths at about 80 percent. These are the raw inputs for batteries, magnets and semiconductors.

Europe is already staring at a difficult decade for its chip sector, yet the feedstock still runs through Chinese refineries.

The risk is not hypothetical. Last year Beijing imposed export controls on critical rare earth metals after tariff threats from US President Donald Trump. Car production lines in the US and Europe came close to halting before the two sides agreed a truce.

The scare has already pushed the EU towards a rare-earth stockpile and new bets to break China’s grip on chip materials.

The inflation catch

Money alone would not solve it, the report argues. Alicia García-Herrero, chief economist for Asia Pacific at Natixis, says the West cannot decouple quickly even with heavy investment. Beijing controls too many industrial inputs, she notes, from rare-earth processing to the active ingredients in medicines.

EY-Parthenon adds a second problem. Chinese goods often carry a factory price advantage of 20 to 100 percent, so replacing them would push prices up. The firm estimates European critical sectors could see prices 1 to 2.5 percent higher, keeping inflation above the 2 percent targets of the European Central Bank and the Bank of England.

A partial retreat

The authors do not argue for standing still. Mats Persson, a former Downing Street adviser now at EY-Parthenon, says localising supply chains without saddling taxpayers and consumers with prohibitive costs will be one of the biggest challenges of the coming years.

He expects the annual outlay to start smaller and grow as the effort widens.

A full break, he suggests, is unrealistic. A partial one is not. For Europe, the number sharpens a hard choice. Its push for technological sovereignty now carries a price tag, and the bill for the tools it most wants to control runs into the trillions.

Portable power stations are a great tool for keeping your most important devices — including phones, laptops and routers — charged during an emergency such as a power outage, but what if you need to power your fridge or washing machine? You may be under the impression that you need to spend thousands of dollars on a whole-home backup system such as a Tesla Powerwall. But thankfully, CNET’s lab testing has revealed you don’t need as much wattage to run your home appliances as you might think. 

We tested 19 large and extra-large portable power stations to see how long they could keep our test fridge running in the lab. The results showed that while extra-large models can oftentimes keep your fridge running longer, that’s not universally true, and depending on your energy needs, a smaller portable power station can work just as well for many large appliances. 

Key takeaways: 

• The Fossibot 3600 Pro performed the best overall, keeping the test refrigerator running for an impressive 44.65 hours (1.86 days). The Oupes Mega 5 followed closely behind in second place at 43.72 hours (1.82 days).
• The Oscal Power Max 2400 had the shortest runtime by far, lasting only 12.58 hours (0.52 days).
• The Anker Solix F3800 Plus performed poorly for its size, lasting only 25.68 hours (1.07 days) despite its large 3,840Wh capacity. 
• The Bluetti AC200L unit is the efficiency king for large appliances, lasting 29.27 hours on a 2,048Wh battery. 
• The Anker Solix C2000 Gen 2 is a standout performer, lasting 27.4 hours on a 2,048Wh battery.

Lab awards for power stations:

CNET’s lab testing gives us reams and reams of data that let us determine the most efficient portable power stations across every size category. Our winners for round-trip efficiency, which measures how much energy you get back from the power station compared to how much power you need to draw from a wall outlet to charge it, are listed below. 

Most Efficient Small Unit (<600 Wh): UDPower AC0600 (RTE: 72.92%)
Most Efficient Medium Unit (600-1299 Wh): Bluetti Elite 100 V2 Bio-Based (RTE: 83.44%)
Most Efficient Large Unit (1300-2199 Wh): Bluetti Elite 200 (RTE: 81.08%)
Most Efficient X-Large Unit (>2200 Wh): Jackery HomePower 3600 Plus (RTE: 79.66%)

For charging speed, which measures how quickly a portable power station can charge to 100% from empty, we have only one winner. 
Fastest charging: Oupes Mega 1 Lite 0.73 hours (about 44 minutes)

How CNET stress tests portable power stations for disasters 

Each of the 19 portable power stations in our test batch was used to power a full-size refrigerator (Whirlpool, 60Hz, single-phase, 118V, full-load 7.2A) until the battery completely drained. We measured how long each unit kept the refrigerator running and reported the results measured in hours and days. Since this test reflects real-world use rather than a controlled laboratory measurement of battery capacity, efficiency ratings (round-trip efficiency) are only reported from our standard testing procedure.

While the results showed that larger units last longer for demanding appliances, they also revealed that slim, custom-designed models such as the Bluetti FridgeGuard prove you don’t need huge, bulky home backups. In fact, several of our better-performing models weren’t what we consider “extra-large,” ones with a capacity larger than 2,200 watt-hours, but rather fell in a more modest 1,300 to 2,199Wh range. 

Unique measurements we use for our fridge stress testing

Measured watt-hours (Wh): The total usable capacity the battery has available to give.
Fridge watt-hours (Wh): The actual amount of energy the refrigerator consumed over the duration of the test.
Delta: The difference between the measured Wh and the fridge Wh. This represents the energy the power station consumed itself (idle draw) just to keep its inverter running for that amount of time.
Percentage reduction: The delta divided by the measured Wh. This gives us the exact percentage of the battery’s total usable energy that was wasted as internal overhead.

The best power stations for large appliances  

Below is our list of power stations, ordered by how long they kept our test fridge running. 

As seen in the combined list below, while the units with massive XL battery capacities generally sit near the top, highly efficient large units including the Bluetti AC200L and the Anker Solix C2000 Gen 2 can outrank several bigger systems (such as the Anker Solix F3800 Plus and Pecron E3600LFP) simply because they waste far less energy. 

The extra-large models naturally dominated the absolute runtimes, but performance varied widely by efficiency, with the Fossibot 3600 Pro being our overall winner for both sheer runtime and efficiency at 44.65 hours (1.86 days). In the large category, while they have smaller batteries, the top large units proved incredibly power-efficient, with the Bluetti AC200L as the overall winner, delivering 29.27 hours of fridge runtime (1.22 days). 

Fridge runtime in order from highest to lowest

1. Fossibot 3600 Pro: 44.65 hours (1.86 days)
2. Oupes Mega 5: 43.72 hours (1.82 days)
3. Jackery Home Power 3600 Plus: 38.58 hours (1.61 days)
4. Jackery Home Power 3000: 32.05 hours (1.34 days)
5. Bluetti Apex 300: 30.43 hours (1.27 days)
6. Bluetti AC200L: 29.27 hours (1.22 days)
7. Anker Solix F3000: 28.97 hours (1.21 days)
8. Pecron E3600LFP: 28.15 hours (1.17 days)
9. Anker Solix C2000 Gen 2: 27.40 hours (1.14 days)
10. Bluetti Elite 200: 26.32 hours (1.10 days)
11. Anker Solix F3800 Plus: 25.68 hours (1.07 days)
12. Bluetti FridgePower: 23.43 hours (0.98 days)
13. ETAKER M2000: 22.48 hours (0.94 days)
14. DJI Power 2000: 22.42 hours (0.93 days)
15. BougeRV Rover 2000: 21.10 hours (0.88 days)
16. Dabbsson 2000L: 21.00 hours (0.88 days)
17. Grecell H2400: 17.75 hours (0.74 days)
18. GoalZero Yeedi 1500: 16.50 hours (0.69 days)
19. OSCAL Power Max 2400: 12.58 hours (0.52 days)

After testing eight, there’s only one extra-large portable power station we recommend for fridges

Extra-large (>2,200Wh) 

1. Fossibot 3600 Pro: 44.65 hours (1.86 days) 
2. Oupes Mega 5: 43.72 hours (1.82 days)
3. Jackery Home Power 3600 Plus: 38.58 hours (1.61 days)
4. Jackery Home Power 3000: 32.05 hours (1.34 days)
5. Bluetti Apex 300: 30.43 hours (1.27 days)
6. Anker Solix F3000: 28.97 hours (1.21 days)
7. Pecron E3600LFP: 28.15 hours (1.17 days)
8. Anker Solix F3800 Plus: 25.68 hours (1.07 days)

The Fossibot 3600 Pro is the best power station for keeping your fridge running. It kept our test fridge running for 44.65 hours (1.86 days). That means you can get almost two days of cooling during a power outage if you own the 3600 Pro. It’s also remarkably power-efficient. Even though the Oupes Mega 5 had a much larger measured watt-hour (4,352.86Wh compared to the Fossibot’s 3,293.74Wh), the Oupes wasted an incredible 1,730.22Wh just to keep its inverter running. 

This resulted in a 39.75% reduction in efficiency for the Oupes Mega 5. Meanwhile, the Fossibot only lost 860.03Wh to overhead (a 26.11% reduction), allowing it to power the fridge for almost an hour longer despite its smaller battery. This shows better power efficiency when handling the cycling loads of a refrigerator.

The Oupes Mega 5 comes a close second in the extra-large category, offering 43.72 hours (or 1.82 days) of power. On paper, while these models sound like they’re in a dead heat, it actually speaks poorly of the much larger Mega 5, which only has 4,352.86watt-hours, according to our measurements, despite what it claims on its box.  

Contrast this with the 3,293.74Wh measured wattage of the Fossibot 3600 Pro. That means, despite having a massive 5,040Wh capacity on paper, the Oupes Mega 5 actually loses a huge chunk of it in real-world use, making it inefficient for running the cycling loads of a typical refrigerator. 

After the Oupes Mega 5, our list starts to separate into more distinct winners. The Jackery Home Power 36000 Plus lasts 38.58 hours (1.61 days), while the Home Power 3000 lasts 32.05 hours (1.34 days). This is to be expected as a function of size and capacity, since each subsequent unit, such as the Bluetti Apex 300, had lower measured watt-hours and capacity than the preceding model. 

The large portable power stations that punch above their weight

Large (1,300–2,199Wh

1. Bluetti AC200L: 29.27 hours (1.22 days)
2. Anker Solix C2000 Gen 2: 27.40 hours (1.14 days)
3. Bluetti Elite 200: 26.32 hours (1.10 days)
4. Bluetti FridgePower: 23.43 hours (0.98 days)
5. ETAKER M2000: 22.48 hours (0.94 days)
6. DJI Power 2000: 22.42 hours (0.93 days)
7. BougeRV Rover2000: 21.10 hours (0.88 days)
8. Dabbsson 2000L: 21.00 hours (0.88 days)
9. Grecell H2400: 17.75 hours (0.74 days)
10. GoalZero Yeedi 1500: 16.50 hours (0.69 days)
11. OSCAL Power Max 2400: 12.58 hours (0.52 days)

If you’re looking for a smaller portable power station that can still keep your fridge and other key home appliances running, look no further than the Bluetti AC200L. With a capacity of just 2,048Wh, it ran the fridge for 29.27 hours.

It completely outclassed several XL units with significantly larger batteries, including the Anker Solix F3000 (3,072Wh), which lasted 28.97 hours, the Pecron E3600LFP (3,072Wh), which lasted 28.15 hours and the Anker Solix F3800 Plus (3,840Wh), which lasted 25.68 hours. 

Advertisement

It led the large category by a solid margin, beating out the Anker Solix C2000 Gen 2 (27.40 hours) and the Bluetti Elite 200 (26.32 hours). 

However, the Anker Solix C2000 Gen 2 is worth considering as well. Also offering a 2,048Wh capacity, it achieved 27.40 hours of runtime, comfortably outlasting the XL Anker Solix F3800 Plus despite having nearly half the battery size on paper.

How much power/wattage do you really need to run your key home appliances? 

Advertisement

Watch this: Portable Power 101: Solar ‘Generators’ Explained

04:02

Our lab testing shows that you may not need as much power or wattage in your portable power stations as you might think. It all comes down to what you’re trying to power. 

“In plain terms, 80% of users’ actual emergency power needs fall under 200W,” Sean Tan, PR manager at Anker Solix, said, referencing data from a survey of 759 Anker power station customers. “Fridges, Wi-Fi routers, lights, phone/laptop charging. Not air conditioners. Not power tools. The ‘wattage arms race’ has been optimizing for the 20% edge cases while the 80% majority pays a premium for capability they never use.” 

For fridges specifically, you need to consider that they’re intermittent loads. That means they cycle on and off, drawing power heavily for short periods and then entering a low-power standby mode. Because of this, a power station’s inverter efficiency and self-consumption (idle draw) matter more than just battery size. That’s why a portable power station such as the Fossibot F3600 Pro can beat something like the Oupes Mega 5, which looks larger on paper in terms of capacity. 

Advertisement

“We generally recommend a 1kw or 2kw power station for common kitchen appliances such as refrigerators, microwaves and coffee makers,” said Brian Essenmacher, head of North American business development for EcoFlow. “These mid-sized units can typically power a fridge for a half day or full day.” 

“A smaller 10- to 15-cubic foot refrigerator during a short four-hour outage may only require around a 1kWh power station,” said Marc Deng, R&D manager at Bluetti, concurring with the power estimate, though he pointed out that for US consumers, the requirement may be on the higher end. “But the reality for many American households is different. Larger French-door and side-by-side refrigerators are often in the 18- to 28-cubic-foot range, with typical power consumption around 80 to 190W. For families preparing for longer outages caused by hurricanes, winter storms or grid failures, around 2kWh becomes a much more practical starting point.” 

What you need to know before buying a power station for a fridge

“The biggest trap is looking at capacity alone,” said Tan. “Capacity tells you how much energy is stored, not how much actually reaches the device.” He points to two things that matter more: light-load efficiency and idle consumption. “A home fridge runs at 100–200W, cycling on and off. Most of the time the power station is operating well below its rated output. Efficiency at these light loads varies dramatically between units.” As more idle consumption, “When the fridge compressor is off, the power station is essentially sitting idle, but still burning power. Industry typical idle draw is 15–20W. That adds up fast.” 

Those looking for a reliable emergency fridge backup should prioritize low idle power consumption and high small-load efficiency over buying the largest, heaviest battery available. A well-optimized 2,000Wh unit can easily outlast a poorly optimized over 3,000Wh unit, as our data shows. By default, all portable power stations waste some energy; the key is to find the one that wastes the least, while also providing stable power. 

Ideally, you also want one with an uninterruptible power supply (or UPS), which helps maintain stable, continuous power. “Modern refrigerators often rely on electronic controls and variable-frequency compressors, which are designed to operate continuously under stable power conditions,” said Marc Deng, R&D manager at Bluetti. “A good rule of thumb is that the UPS transition time should be below 20 milliseconds, with around 10 milliseconds being ideal. If the transfer is too slow, the refrigerator’s internal electronics may reset, causing the compressor and control system to restart unnecessarily. Over time, repeated interruptions can create additional stress on the appliance.” 

It’s also important not to overlook the surge power stat that many portable power stations advertise. Deng pointed out that depending on the refrigerator model, the startup surge can range from around 600W to well over 1,200W. “Surge power matters a lot,” Tan said. “Compressor startup pulls several times the running wattage. A station needs 2,500–3,000W peak to reliably handle most residential fridges.” 

Advertisement

You also need to consider the percentage of energy wasted as overhead. The Anker Solix F3800 Plus wasted 46.63% of its energy just keeping itself running, which explains its short runtime. The Pecron E3600LFP had the lowest overhead loss with a tiny 15.75% reduction, while other highly efficient units included the Anker Solix C2000 Gen 2 (18.21% loss) and the Bluetti AC200L (22.16% loss). These low-loss units ensure that power actually reaches the refrigerator rather than being lost as internal heat. 

In the XL category, our undisputed winner is the Fossibot F3600 Pro, which is $2,600 at full price, but frequently on sale for less. It lasted the longest and wasted the least energy in the XL category. For the large power stations, our winner is the Bluetti AC200L. It punched above its size and power capacity, offering excellent efficiency and the longest fridge runtime of any large portable power station. 

Do you need a portable power station specially designed for fridges?

Advertisement

It’s worth noting that there’s a new category of portable power stations, with slim models designed to sit on or be mounted behind a fridge. Examples include the new Bluetti FridgePower and the Jackery FridgeGuard, both of which I’m testing at home, while Anker’s Solix S2000 is a more traditional take on the capability. 

Of these, we’ve only been able to run lab tests on the FridgePower, and the results are impressive. It falls into the upper tier of large portable power, keeping our test fridge running 23.43 hours (0.98 days), so it lasts nearly a full day. Of its 1,621.09Wh measured capacity, the FridgePower only lost 328.08Wh (a 20.24% reduction) to internal inverter overhead. This compares favorably with other, more efficient models. 

While it won’t keep the fridge running for multiple days, even the bulky Fossibot 3600 Pro is hard-pressed to do that. The design is its biggest selling point: The FridgePower is slim, lightweight and can be out of sight and out of mind, unlike hooking up your fridge to a traditional power station. So while it’s not required to power a fridge, the FridgePower won’t look out of place in your kitchen and it provides sufficient power to get you through an outage.