Running a single task on Inkling 2 Small costs $1.18, representing a nearly sixfold expense jump over its previous generation that directly challenges the unit economics of deploying open-weight infrastructure. Thinking Machines, the startup founded by former OpenAI chief technology officer Mira Murati, released the mixture-of-experts model containing 594 billion total parameters, with 17 billion active during execution. The release sets up an immediate operational question for engineering leaders: whether benchmark gains justify a sixfold surge in per-task compute costs.
Transitioning to a mixture-of-experts architecture with adjustable reasoning depth delivers higher scores on automated workflows while inflating token volume. On the Artificial Analysis Intelligence Index, Inkling 2 Small reached a score of 31, representing an 18-point jump over its predecessor and matching Google's Gemini 3.8 Flash.
Benchmark Performance and Cost Escalation
The architectural shift shows its clearest effects on multi-step agentic workflows. On AutomationBench-AA, the model's score climbed from 5 to 46 percent, while its FrontierCode 2.0 evaluation rose from 5 to 25 percent. On the knowledge-work benchmark AA-Briefcase, Inkling 2 Small reached 1,527 Elo, which is 615 points higher than the original model.
The model generates roughly 150,000 output tokens per task on average, driving execution expenses up to $1.18.
According to Artificial Analysis data, this token volume represents a quadrupling of output generation, creating the financial mechanism behind the $1.18 price tag. On FrontierCode 2.0, Inkling 2 Small achieved 25.4 percent, while reaching 72.1 percent on DeepSWE and 86.3 percent on SWEBench Pro v2 (Hard). In multimodal evaluations, the model registered 25.2 percent on GDP.pdf, 81.6 percent on MMMU Pro, 64.7 percent on PerceptionBench, and a 1,340 Elo rating on DesignArena Website. In an isolated evaluation, the model identified three wines in 19 out of 20 blind tasting attempts from photos alone.
The Reliability Trade-Off
A critical structural divergence emerges behind the benchmark headline numbers. While the measured hallucination rate dropped from 61 to 27 percent, that improvement coincided directly with a steep decline in query responses. The model's response rate fell from 77 to 53 percent, meaning the apparent reduction in errors stems largely from declining to answer difficult prompts.
Thinking Machines also noted operational limitations, acknowledging that the model sometimes communicates inefficiently, fails to pick optimal tools for large codebases, lacks audio support, and produces cluttered outputs on health topics. The underlying training footprint remained relatively constrained, processing 12 trillion tokens across all phases, with the post-training reinforcement learning run requiring 3 trillion tokens on 2,048 GPUs across nine days.
Proponents point out that at $1.18 per task, Inkling 2 Small remains significantly cheaper than closed competitors like Gemini 3.8 Flash at $7.77 or Claude Sonnet 5.5 at $10.02.
That cost differential places the model closer to low-cost alternatives, though deployment realities dictate whether the savings offset the lower completion rate. Thinking Machines currently distributes Inkling 2 Small as a proprietary offering on its Tinker platform, with plans to release weights under an Apache license, competing directly on price against Chinese alternatives.
Ultimately, the pitch of Inkling 2 Small centers on delivering elite reasoning efficiency to specialized enterprise workflows, yet the operating reality presents a system that charges six times more per task while declining to answer nearly half of user queries.
For comparison
Budgets for automated agent workflows risk a sharp overrun due to the inference footprint of the latest release from Thinking Machines. In mid-May, Thinking Machines focused on interaction models, yet this architecture simply avoids difficult queries in production. With the successful response rate dropping to 53%, the reported drop in hallucinations largely stems from outright refusals on hard prompts. A drop in hallucinations looks far less impressive when the system simply declines nearly half its tasks.
Our stories: Thinking Machines: Mira Murati’s Shift to Real-Time AI (May 12).