Pathway's 150M Model Proves AI Cost-Efficiency Doesn't Require Massive Scale

Pathway says architecture can matter as much as model scale

* A 150M model reached 29.5% while costing just $0.0007 per task * ChatGPT scored higher, yet its comparable reasoning runs cost substantially more * BDH-CQ performs reasoning internally instead of generating lengthy intermediate text Pathway, an AI lab focused on building Post-Transformer architectures, has released new benchmark results for its BDH-CQ reasoning model. According to the researchers, their 150M-parameter model scored 29.5% pass@2 on the public ARC-AGI-1 evaluation set. It achieved this at a computed inference cost of $0.0007 per task, roughly eleven times cheaper than ChatGPT's underlying GPT 5.6 Luna (Low) model. A cheaper way to reason Today, many AI tools waste computing power because of how they are designed, not because deep reasoning demands it. "Today's AI pays a steep token cost for reasoning, but that cost is imposed by architecture, not by any law of intelligence," said Zuzanna Stamirowska, CEO and co-founder of Pathway. "We show that a different architecture changes the game and opens up a whole new space in terms of how much intelligence per dollar." Amazon Web Services believes that BDH-CQ's result is a promising step toward using advanced AI reasoning in real products more affordably. "Customers are increasingly exploring how to move advanced reasoning from experimentation into production, where performance, efficiency, and scalability all matter," said Nicolas Tarducci of AWS. ARC-AGI-1, a widely used reasoning benchmark for AI systems, checks whether a system can infer an underlying rule from limited examples and apply it correctly to new inputs. In this test, OpenAI's Luna model scored only slightly higher at 34.2%, yet running it still costs significantly more ($0.008 per task). That price gap already includes OpenAI's recent 80% price cut on Luna, which began on July 30th of this year. Further up the chart, Claude Opus 5 and Gemini 3.1 Pro reach 97-98% but cost around $0.5 - $0.6 per task, meaning the frontier's very top costs close to a thousand times more than BDH-CQ for the highest scores. On the cheap end, Qwen3 235B costs over three times more than BDH-CQ while scoring worse than even its Low variant, so it isn't a real competitor on either price or performance. "Pathway shows that model architecture, not just scale, can drive the next leap in AI reasoning," said Łukasz Kaiser, co-author of the original 2017 Transformer paper. Why it costs so much less The efficiency gap stems mainly from a structural difference in how each system actually performs reasoning during inference computations. Many reasoning AI systems generate intermediate text, adding one token after another before producing their final answers. The longer that written reasoning becomes, the more it costs to run and the slower the AI responds to each request. BDH-CQ works quite differently, quietly solving problems inside its own memory instead of writing everything down first as visible text. Pathway also confirmed that early experiments already follow standard Transformer-like scaling laws across model sizes from 1B to 600B parameters. The company also plans to extend this approach toward harder benchmarks, including mathematical reasoning, ARC-AGI-2, and eventually full ARC-AGI-3 evaluations. If these efficiency gains hold across larger and more difficult tasks, cost rather than raw capability could increasingly separate rival reasoning systems. Follow TechRadar on Google News and add us as a preferred source to get our expert news, reviews, and opinion in your feeds.

Did we build the engine before we worked out the physics? What's the pathway to take after the LLM transformer?

The AI boom was built on top of the transformer model first discovered in 2017 for better translation. The stock valuations, planned multi-gigawatt data centers, and the whole AI economy now exist because transformers unlocked the ability to automatically train AI models on human language at scale. Pretty much all of the hype builds on this seminal technology because it opened the ability to process unstructured data in new ways to do pretty much everything people are talking about today - coding, agents, harnesses, digital twins, 'reasoning', content generation, and the list goes on. That said, it's still early days for applying AI to unstructured data because actually building something useful depends on a lot of what is essentially baling wire and duct tape to keep the whole AI vehicle in motion on a rocky road ahead. Context engineering, RAG, harnesses, tokenomics and fine-tuning are some of the more prominent examples. Thus, researchers have been exploring other approaches from first principles for achieving the scale of transformers without all of the baling wire and duct tape. Into this fray steps Pathway, which has recently been valued at $500 million thanks in part to progress on BDH (baby dragon hatchling), a promising post-transformer model that learns in the flow of work, is interpretable, and is smaller and more cost-efficient than transformer-based algorithms. CEO Zuzanna Stamirowska came to AI from complex systems rather than from deep learning. She did her doctoral work on network evolution and co-authored a global trade forecasting model published by the National Academy of Sciences. Pathway started as a real-time data and streaming framework company that stumbled onto an entirely new and practical way to work with natural language and other unstructured data, informed by first principles. Stamirowska explains: My background definitely helped me in addressing the missing theoretical foundations of AI that allowed us to take a step back and rethink the architecture. With brute force scaling of the Transformer (the main piece of math that sparked the AI revolution), the AI community built steam engines for thought before developing the thermodynamics of intelligence. As a result, today's AI is an engineering miracle that lacks an underlying science: it works remarkably well, yet we cannot fully explain why and are puzzled by weird behaviors of the models that suddenly 'escape their creators'. When things go off-script, through things like noisy input, fuzzy goals, long chains of logic beyond the conditions covered by benchmarks, we do not know what to expect and lack a direct way of fixing the model. In Complex Systems, the main thing we study is emergence. When we started working on the new architecture for reasoning, we looked at the elements that biology got right in the brain (local interactions, sparse activations, brain as a physical network system) and wanted to find a small set of laws, in the spirit of statistical mechanics, that derive how a system reasons from how its parts interact. We have done this in the BDH paper, opening a path toward systems whose capabilities, stability, and failure regimes can be characterized before deployment. This ensures we make safety a design property to be analyzed and engineered, rather than an outcome assessed only after deployment. Experience that compounds The seminal insight in the 2017 Google paper introducing transformers was that "attention is all you need." The idea being that previously people required all sorts of prework to teach models to work with natural language and other unstructured data. It took another five years for OpenAI to show that this simple data transformation engine could be trained on large volumes of information to create more capable chatbots like ChatGPT, suggesting that artificial general intelligence (AGI) could eventually be trained on more data. But attention is only part of the story when it comes to something like human intelligence. Another essential aspect is awareness, or the ability to make sense of the focus of your attention in the context of space and over time. Transformers need all sorts of scaffolding to remember the last conversation, the last correction, or the last quarter. Thus, a surrounding stack is required to reconstruct all of it on every call. Stamirowska frames the alternative as the difference between consuming context and accumulating experience. If learning can happen safely during inference, the model can start to accumulate experience, not merely consume context. Less work goes into ever-complicated context management and retrieval, and instead is baked natively into the model. The larger idea is that software becomes less like a tool you periodically update and more like a colleague whose useful experience compounds. Today companies keep paying AI systems to relearn the context of their business, or spend tons of work building extensive Retrieval-Augmented Generation (RAG) systems that never quite accumulate operational intelligence. This presents a long-term opportunity for experience to compound. That changes several things: * Systems can learn an organization's recurring patterns, exceptions, terminology and decision rules through use. * Therefore, every corporation can have its own "sovereign" AI, without fine-tuning open-source models. * An agent that makes a mistake can potentially incorporate the correction rather than requiring that lesson to be engineered back into the surrounding stack. * Personalization can become deeper than "retrieve some facts about this user and put them in the prompt." * Long-running systems can preserve useful states instead of repeatedly re-reading and reconstructing them. * The economics of adaptation change because every improvement does not require another expensive centralized training cycle. Not a single race Pathway is certainly not the only AI lab exploring the frontier beyond transformers. Liquid AI is pushing small models onto phones and laptops. State-space model (SSM) hybrids from NVIDIA, Alibaba and IBM already ship inside mainstream production models. Active inference comes out of neuroscience entirely. Stamirowska sorts her map of this emerging landscape like this: * Transformers: highly capable general-purpose sequence models, with memory and learning mostly handled outside the core inference process. * SSMs / hybrids: change how sequences and state are represented, often targeting efficiency, long context and lower memory overhead. These are less reasoning-focused. * Device-native architectures: optimize intelligence for constrained hardware, latency, privacy and edge deployment. * BDH: Co-occurring reasoning and persistent state. Exploring challenging all of the above. But the primary incumbent we have to beat at the moment is the entire Transformer stack: the model + RAG + long-term context + fine-tuning + agents + external memory + increasingly cheap inference. She explains: Post-Transformer is not a single race. Some architectures are trying to process tokens more efficiently, but we are collectively asking a more fundamental question: can learning, memory and reasoning become part of a model's runtime, or execution phase, rather than through the services wrapped around it?" Thinking without talking BDH-CQ can reason and solve difficult tasks like ARC-AGI and Sudoku Extreme without ever verbalizing its thoughts. In contrast, reasoning is bolted onto transformer models in the form of a verbalized chain of thought. Pathway says BDH-CQ does this while matching the transformer's scaling laws on language, and that early experiments confirm Transformer-like scaling laws hold during pretraining from 1B to 600B parameters while preserving BDH-CQ's latent reasoning. Today's reasoning models think out loud, and every thought has to be turned into words, emitted, and read back before the next one can start. BDH-CQ keeps the thinking internal. One practical implication is increased financial efficiency. For example, on the public ARC-AGI-1 evaluation, BDH-CQ solves 29.5% of the tasks at a computed $0.0007 apiece, against 97.5% for Anthropic's Claude Opus 5 at $0.70. Opus is right a little over three times as often and costs a thousand times more to get there. The BDH approach has not quite caught up with the frontier in terms of accuracy, but then it's much earlier in its learning curve and does so at a much lower cost. But then AGI is a synthetic challenge rather than a practical one that models are optimizing for to achieve bragging rights. A more practical concern is that when transformer models fail, it's hard to figure out why. New economic model Tokenomics is becoming a meme this year as frontier labs and enterprise vendors look for new ways to price AI capabilities. AI sticker shock is already causing pushback and restraint as finance teams start to make honest assessments of the value they are getting for burning more electrons. Stamirowska unpacks how new approaches like BDH might shift the calculus of AI pricing: I would separate the billing unit from the underlying economics. Per-token pricing may survive for a long time because it is simple, measurable and familiar. But strategically, tokens increasingly look like an implementation detail rather than the thing customers actually value. Customers do not want tokens; they want investigations completed, schedules optimized, exceptions resolved, software built, decisions supported, or a persistent agent doing useful work. Perhaps more interesting is the potential for context-aware models to create a new way to think of the long-term value of AI: If models accumulate useful states, the persistent model instance itself becomes valuable. A model that has spent six months learning how a company's operation works is not economically identical to a fresh model with the same base weights. It has a different relationship with the customer entirely. With the result on ARC-AGI-1, we clearly show a new Pareto frontier for AI, meaning that enterprises will be able to get more value per token, and especially, way more value from their AI deployments." One side effect for vendors of these models is a new kind of stickiness. It's one thing to get locked into a particular model out of habit and tooling, and another when value accrues in the context learned over time. The model then becomes a tenured asset and budgeting, depreciation and renewal all change. If value accrues in the instance, switching vendors means abandoning the asset, which is the deepest lock-in yet. It also raises a harder problem in regulated sectors. If a model's behavior changes through use, its validation goes stale continuously. Risk teams will need a way to establish what exactly "the model" refers to at the moment of an audit. Maximum intelligence per watt So how well might a new approach play with the terrawatts of new AI data centers now in the planning phase? Stamirowska says BDH runs on any GPU, so all the buildout is perfectly aligned with the proliferation of AI, no matter the architecture. More interestingly, she believes that BDH's efficiency will help mitigate the physical limits the AI industry is experiencing in terms of compute (chips, energy) and social limits (communities being strongly against data center buildouts). BDH promises the potential for squeezing out more intelligence per watt. She says: Now, taking a more detailed view, memory bandwidth will be the most important parameter. Thanks to the sparsity of BDH, we can pretrain very large models from scratch very cheaply. Speculating a bit, one could imagine a future where, say, one full rack of GB300 or similar would be almost like the new mainframe for a 'corporate-sovereign model.' BDH opens room for everything that better supports sparsity. The small sizes of our models and efficiency, and the fact that we keep state in memory, mean that going to the edge will be not too far ahead. I personally believe that it's reasoning that is key to autonomous systems, so we will increasingly see small, powerful reasoning models deployed at the edge." Sparsity is important because it's what current GPU architectures are weakest at. For example, when Mamba, a popular state space model, first shipped, it needed hand-written CUDA kernels and ran at a fraction of the tensor core utilization a transformer achieves. So its successor was redesigned to map onto tensor cores. BDH benefits most from processors where memory is meshed with compute." The path to Pathway Stamirowska says the first BDH deployment should be bounded, high-value workloads where repeated reasoning over evolving information, large accumulations of states, expensive recurring inference, constraint-heavy problems, or workflows where learnings from prior interactions would materially improve the system. Initially, BDH can sit beside and complement existing models rather than replacing them. She cautions: I would not ask a CIO to replace their Transformer stack. I would ask them to give us the hardest workflow where they are paying a heavy architectural tax today, or transformers simply do not work, and we'd apply BDH and measure performance against standard enterprise metrics. Questions to consider include: * Does it solve the task better (or well, if new?)? * Does it retain useful information across the workflow? * Does performance improve with experience? * Does the total system become simpler or cheaper? For example, some of the direct use cases in financial services may include reconciliation or anomaly detection. If these are successful, the footprint for wider application expands. She explains: What it replaces first is therefore probably not 'the LLM.' It is some combination of repeated context re-construction, excessive inference-time compute, bespoke memory scaffolding, repeated fine-tuning and specialized optimization software around the model. My take James Watt gets credit for the steam engine, which is not quite right. Thomas Newcomen built the first one that did useful work in 1712 to pump water out of coal mines. It was spectacularly wasteful, burning coal by the ton, which was OK because plenty of coal was nearby. That's the current state of transformers and the data centers that run them, which coincidentally all require nearby power plants because there are not enough wires to power them from afar. Watt's innovation sixty years later was a separate condenser that cut fuel requirements to a quarter. The transformer is Newcomen's pump. It got the water out of the unstructured data to unlock access, and it's also wasteful in a lot of ways in terms of energy, tooling, complexity, and misplaced trust. This is probably why Dominion Energy has approved seventy gigawatts of data center projects in Northern Virginia, with another fifty in the queue, compared to the five gigawatts built today. It's also why roughly 30% of all planned US data center capacity now comes with its own on-site generation, with more than 80% of that shiny new gas plants (if a new gas turbine could ever be called shiny). Nobody runs a reciprocating steam engine anymore, and the world is trying to wean itself off of coal. But we never abandoned the principle. Most of the world's electricity is still made by boiling water to spin a turbine, including the electricity going into the data centers. In the long run, transformers might experience the same trajectory. Foundational somewhere in the stack in the same way the steam cycle is foundational and invisible. A lot of the pushback people currently hold against AI mirrors the pushback of coal and steam and industrialization, particularly in the early days. The modern variant is the energy and associated rate hikes & pollution, the slop, and the extraction. Only time will tell if these are intrinsic to AI, the human operators or the reliance on an inefficient engine for transforming the context of unstructured data into value. Much of the tooling around frontier models exists because transformer models have attention but no awareness to make sense of the situation they are tasked with or what they learned last time. A good share of the electricity now being wired into these models will go to the baling wire and duct tape rather than the work we want done. William Jevons pointed out in 1868 that Watt's efficiency gains increased coal consumption rather than reduced it. Cheap AI could mean more people will find ways to put it to work. Hopefully the gains will turn out to be higher than the environmental and social costs of whatever AI nirvana is being promised today. The trillion-dollar question is not whether the transformer is the right architecture. It is whether we work out the thermodynamics before or after the carnage.

Pathway Claims Major AI Cost-Efficiency Breakthrough with New 150M-Parameter Model

The cost gap comes from a structural difference. Many transformer-based reasoning systems externalize their work as a chain-of-thought, generating extra tokens sequentially and feeding them back into later steps. Pathway, an AI lab building a Post-Transformer architecture and models, today published benchmark results for BDH-CQ, a 150-million-parameter reasoning model. BDH-CQ scored 29.5% pass@2 on the public ARC-AGI-1 evaluation set at a computed inference cost of $0.0007 per task. Thanks to a new approach to reasoning, BDH-CQ runs approximately 11 times as cheaply per task as GPT 5.6 Luna (Low), even after accounting for OpenAI's 80% price cut of 5.6 Luna on July 30th. Luna scores 34.2% against BDH-CQ's 29.5%, a modest accuracy gain at 11 times the cost. ARC-AGI-1 is a public reasoning benchmark that tests whether a system can infer an underlying rule from a small number of examples and apply it to a new input, a capability often associated with human-like intelligence. The cost gap comes from a structural difference. Many transformer-based reasoning systems externalize their work as a chain-of-thought, generating extra tokens sequentially and feeding them back into later steps. As the trace grows, so do inference cost and latency, significantly. BDH-CQ instead performs this work in a recurrent latent state, learning from examples and refining a solution without generating an intermediate text trace. It thinks natively rather than using an intermediate-language scratchpad. "Today's AI pays a steep token cost for reasoning, but that cost is imposed by architecture, not by any law of intelligence. Currently, every reasoning step consumes context, adds latency, and burns compute. We show that a different architecture changes the game and opens up a whole new space in terms of how much intelligence per dollar. A 150M-parameter model, built on Pathway's BDH architecture, reasons recurrently in latent space, and sets a new state of the art in cost efficiency on ARC-AGI-1. The bottleneck was never intelligence. It was design," said Zuzanna Stamirowska, CEO and co-founder of Pathway. Scores and comparison costs are drawn from the ARC Prize Foundation's public leaderboard, which plots each submitted system's benchmark score against its cost per task, as of July 2026. Pathway's cost is computed from measured hardware time. Comparison costs are those reported to the leaderboard and may reflect API pricing for generalist models. BDH's ARC-AGI-1 results were evaluated by Łukasz Kaiser, a co-author of the 2017 paper that introduced the Transformer architecture, who said, "I've followed Pathway closely and replicated their ARC-AGI-1 results myself. Pathway shows that model architecture, not just scale, can drive the next leap in AI reasoning." The results were also reproduced by Richard Zhong, an NYU researcher focused on model evaluation and benchmark robustness and a co-author from Bielik. Pathway plans to scale the architecture next and extend the approach to more challenging reasoning benchmarks such as mathematical reasoning, ARC-AGI-2, and ARC-AGI-3, and to build a latent reasoning Large Language Model. When its efficiency and state-tracking capabilities extend to those domains, BDH will support a wide range of applications that must reason reliably as information and constraints change, from cybersecurity incident response to real-time industrial operations. Early experiments confirm Transformer-like scaling laws apply during pretraining at scales from 1B to 600B parameters, while preserving the latent reasoning capabilities specific to BDH-CQ.