In October 2026, a wave of open-weight models was released, including Reflection AI’s Beam, DeepSeek V4.1 Flash, NVIDIA Nemotron 3.5 Super VL, and Meta’s Muse Spark Open. Over the past two years, everyone comparing model cost-performance has been accustomed to looking at the price per million input and output tokens. However, as multi-step reasoning and agent workflows have become widespread, executing a single task may require multiple tool calls or iterative thinking and self-correction, making unit token prices alone insufficient for accurately estimating monthly costs in real production environments. Consequently, these new models released in October have begun adopting a new approach: comparing cost-effectiveness by calculating the total expense required to complete an entire task.
The third-party evaluation organization Artificial Analysis (AA) also provides a unified Cost per Task calculation methodology on its Reflection Beam evaluation page. By having models complete the entire suite of Intelligence Index benchmarks, they calculate the cost of all tokens—including inputs, cache reads, reasoning chains of thought, and final answers—and then compute a weighted average based on benchmark weights and task volumes to derive the average cost per task. Although this methodology is not yet an industry standard and only accounts for token consumption during each single evaluation run, it is the first to compare models across diverse architectures and reasoning paradigms on the same per-task cost basis, giving engineering teams a direct horizontal comparison.
Once you lay out this total task bill, the engineering intuition that “whichever model talks more costs more” no longer holds up. AA’s benchmark results show that Meta’s Muse Spark Open and DeepSeek on the DeepSeek evaluation page both achieved an overall score of 39 on the Intelligence Index. Muse was quite restrained throughout the evaluation, generating less than half the total tokens of DeepSeek. Yet once the single-task bill is calculated, Muse—despite emitting less than half as many tokens—costs an average of $0.84 per task, whereas DeepSeek, the most verbose model in the field, leverages its low per-token unit price to push the cost per task down to $0.27—meaning the more token-efficient model actually ends up costing more than three times as much to finish a single task.
If we bring leading proprietary flagship models into the comparison, the price gap across tiers becomes even more striking. Take the highest-scoring proprietary flagship on the Claude Opus evaluation page, for instance: the total number of tokens it emitted across the full evaluation run is actually roughly similar to DeepSeek’s, but its average bill per task reaches $5.98—a 22-fold difference in cost despite comparable output volume. Another high-scoring proprietary flagship on the GPT Astra evaluation page was remarkably sparing with its words during testing, producing only a fraction of the former’s total output, yet its steep per-token list price still pushed its average Cost per Task up to $3.26.
Placing the evaluated models’ overall Intelligence Index scores and average Cost per Task onto the same two-dimensional chart naturally splits the market into two camps along a $1 dividing line. To the right of this line lies the high-price band guarded by proprietary flagships and sub-flagships: alongside Opus and Astra in the $3 to $6 range, mid-to-high-tier models scoring in the 40s on the Kimi evaluation page and Gemini Flash evaluation page also land at around $2 and just over $1 per task respectively, clustering in the upper-right quadrant.
Looking to the left of the dividing line, the four sparse MoE open models released in October do not line up neatly along a diagonal where higher cost buys higher capability; instead, they exhibit an inversion across both score and per-task bill. Beyond the more than three-fold price gap between Muse and DeepSeek at the exact same score line, Beam—which focuses on compressing reasoning chain-of-thought length—lands at a Cost per Task of $0.57 with an overall score stalling at 30, losing out to the most verbose model in the field, DeepSeek, on both capability and per-task expense. As for Nemotron on the Nemotron evaluation page, which pushes the Cost per Task down to a field-low $0.19, its overall score also drops into the lower-left corner of the chart.
Next to this group of low-cost sparse models, Qwen3.8 27B on the Qwen peer model comparison page falls just outside the dividing line. As a compact Dense model, it fails to squeeze into the low-price zone on the left under public cloud per-token billing rules, yet it points directly to another hard financial calculation beyond public cloud API bills: VRAM amortization in self-hosted on-premises data centers.
At its core, the total bill for a single task is a simple multiplication: how many tokens the model emits to finish the task × the execution unit price per token. To drive the total bill down, one path is training the model to emit fewer tokens, and the other is making the compute cost of each token cheaper. Each of these four open-weight models targets a different factor in that multiplication, making distinct trade-offs across model architecture, post-training investment, and capability ceiling.
The path Reflection AI chose for Beam is to directly slash the first factor—reducing redundant steps in the reasoning chain of thought. According to Beam’s official release blog and the Beam model card, this 500-billion-parameter-class sparse MoE model underwent large-scale reinforcement learning during post-training across tens of thousands of next-generation NVIDIA GPUs, incorporating a length penalty into its reward function that officially claims to shorten reasoning chains of thought by roughly 3x at equivalent quality. AA’s benchmark tests did confirm this effect, bringing the average output per task down to 26k tokens. Unfortunately, its list price per token is far from cheap, leaving its average Cost per Task still close to $0.60, and due to the relatively strict length penalty, its overall Intelligence Index score reached only 30.
DeepSeek took the opposite route: instead of aggressively capping token count, it redesigned the underlying architecture to drive down the unit price of every token. As shown in the technical details of the DeepSeek model repository and our previous architectural analysis, this sparse MoE model employs a Causal Encoder-Decoder (CED) architecture that decouples active parameters between input processing and text generation—activating only 8B parameters during the prefill stage for long-context inputs, and activating 16B parameters only during the token-by-token decoding stage. Because many real-world workloads require stuffing massive background documents into the context window, cutting active prefill parameters in half directly saves inference providers substantial compute overhead on long inputs, pulling both input and output unit prices down to rock-bottom levels. Even though AA’s evaluation notes that it is one of the most verbose models in the tested cohort, its lightweight prefill and ultra-low unit pricing keep its Cost per Task firmly in the twenty-cent range.
NVIDIA’s Nemotron also attacks the per-token unit price by scaling down the overall model footprint and applying full-stack inference optimizations to boost single-GPU throughput. According to NVIDIA’s technical blog, this model uses 120B total parameters and 12B active parameters built on a hybrid Mamba-2 state-space and MoE architecture, incorporating low-bit quantization, distillation, and speculative decoding across training and deployment. Full-stack optimization pulls its Cost per Task down to a field-low $0.19. However, AA’s evaluation also points out that its per-task token consumption remains relatively high—falling short of the top tier in token efficiency—while its overall score drops noticeably and its entire inference acceleration stack relies heavily on NVIDIA’s hardware and software ecosystem.
By contrast, Meta’s release in the Muse model repository takes a completely different path. Rather than fixating on slashing prices, it prioritizes preserving the capability ceiling for multimodal understanding and complex reasoning. Muse is the largest of the four open models at 816B total parameters and 42B active parameters, natively supports multimodal inputs, and defaults to the highest reasoning effort (xhigh), never pitching extreme output compression as its selling point from the outset. Activating over 40 billion parameters for every single token naturally drives up the execution unit price. Even though it kept its output volume admirably restrained across the entire evaluation suite, its resulting Cost per Task is still more than three times higher than that of the identically scoring DeepSeek. Yet the upside is equally clear: workloads demanding intensive multimodal reasoning can still find a high-scoring open-weight foundation capable of handling complex tasks.
After calculating the single-task bill under public cloud API pricing, engineering teams looking to deploy these models into production must reckon with two additional hard costs hidden outside the benchmark tables. One is hardware VRAM amortization for self-hosted on-premises deployments, and the other is the success-rate denominator introduced by failed-task retries in multi-step agent workflows.
Although the four inexpensive open models discussed above all release their weights publicly, their massive footprints—ranging from over a hundred billion to more than eight hundred billion total parameters—still require multi-GPU clusters and complex distributed inference stacks if you want to self-host them on-premises. That is why in the Qwen model repository, Qwen3.8 27B—released under an Apache open-source license—adopts the approach of a 27B Dense model combined with long context and native vision. Its $1.01 Cost per Task on the AA leaderboard looks anomalous partly because a Dense architecture must activate all parameters for every generated token while producing relatively long chains of thought at the highest (xhigh) reasoning setting, and partly because AA defaults to Alibaba Cloud’s relatively high first-party list price; switching to third-party providers like DeepInfra brings its Cost per Task back down to around $0.50. More importantly, in an on-premises data center, such a compact Dense model drastically reduces VRAM footprint and day-to-day operational complexity—substantially lowering GPU cluster procurement and amortization costs for teams with strict data-privacy requirements.
An even more hidden cost than hardware thresholds is that AA’s Cost per Task only counts token expenses from a single evaluation run, leaving out the retry overhead after task failures in production, spinning agent loops, and human engineering fallback and debugging. On Terminal-Bench, a benchmark evaluating complex multi-step terminal execution, combining AA’s in-depth analysis of Beam with model comparison data shows that Beam—despite a single-run bill of just over fifty cents—achieves a pass rate of only 3.0%, whereas the more expensive per-run Qwen Dense model reaches 5.6% and Muse achieves 8.1%. If you plug a model into a low-fault-tolerance multi-step terminal automation workflow, an overly low pass rate will trigger massive retry loops and repeated context accumulation, quickly wiping out the tens of cents saved on a single invocation.
For technical teams evaluating models that pitch a low Cost per Task, there is a clear order of operations for decision-making: engineering selection should first validate pass rates against the company’s most complex real-world tasks to filter out models that fail to clear the bar, and only then align the surviving candidates with the specific bottlenecks of their own systems. If daily workloads are dominated by massive document inputs and high-frequency context prefilling, DeepSeek’s lightweight prefill can effectively hold down the bill; if tasks rely heavily on multimodal understanding and long-horizon, high-complexity execution, Muse, Kimi, or even Gemini Flash, Astra, and Opus remain indispensable fallback options. For fixed, short tasks with clear boundaries and high fault tolerance, teams can consider Nemotron or Beam to trim single-run expenses; however, once fixed GPU-hour amortization in an on-premises data center enters the equation, a compact foundation like the Qwen Dense model that fits comfortably into modest VRAM often proves far more cost-effective on the total balance sheet than ultra-large sparse MoE models with lower nominal API unit prices.