Thinking Machines Lab released Inkling 2 Small on November 5, 2026 — a 594-billion-parameter mixture-of-experts model that scores 58.4 on the BenchLeader Index, the highest-ranked new model of the week to November 8. It costs $0.475 per 1 million tokens at blended list price, or $1.18 per task. The Apache 2.0 weights have not yet been released.
Those numbers deserve unpacking, because they tell a story about where the model market is heading: capability is still climbing, but the conversation is increasingly about what capability costs per unit of useful work.
What Is a Mixture-of-Experts Model?
A mixture-of-experts (MoE) model is a large language model built from many smaller specialist sub-models — the "experts" — plus a routing mechanism that decides which experts handle each piece of input. For any given token, only a subset of the experts is activated. The model can therefore be very large in total parameters — 594 billion, in Inkling 2 Small's case — while the amount of computation used for each individual request stays manageable.
The analogy is a hospital: it employs hundreds of specialists, but any one patient sees only a handful. MoE architectures let model builders scale up total knowledge without scaling up the cost of every inference in proportion.
This matters because inference cost — the cost of running the model to answer a question — is the recurring expense that determines whether an AI product is economically viable. A model can be brilliantly capable and still commercially useless if each answer costs too much. MoE is one of the main techniques the industry uses to square that circle.
What the Benchmark Score Means
Inkling 2 Small scores 58.4 on the BenchLeader Index, making it the highest-ranked new model of the week to November 8, 2026. Benchmark indices like BenchLeader aggregate performance across a set of tasks designed to measure a model's general capability — reasoning, knowledge, instruction-following, and related skills — into a single comparable number.
A single number should always be treated with caution. Benchmarks measure what they measure, and a model's performance on a benchmark suite is not identical to its performance on a specific real-world task. A coding assistant, a customer support agent, and a research tool each stress different capabilities, and a model that leads on an aggregate index may not lead on the task a particular founder cares about.
What the score does establish is that Inkling 2 Small is competitive at the frontier of newly released models. Ranking first among the week's new releases is a meaningful signal — and, for a lab still building its reputation, a statement of intent.
Why Per-Task Pricing Matters
The pricing figures are arguably the more important part of the story. Inkling 2 Small costs $0.475 per 1 million tokens at blended list price, and $1.18 per task. Token pricing is the traditional way models are sold: you pay for the amount of text processed. Per-task pricing is a different lens — it asks what a completed unit of useful work costs, which is the number a founder actually cares about when pricing a product.
A model that costs $1.18 per task changes the economics of AI-powered products. If a task is something a customer will pay several dollars for — a drafted document, a research summary, a triaged support ticket — a $1.18 cost leaves a healthy margin. If a task is something a customer expects for pennies, it does not. Per-task pricing makes that calculation explicit, and it is becoming an increasingly important way to compare models, because token prices alone hide differences in how efficiently a model uses tokens to complete a task.
For founders choosing a model, the practical question is not "which model is smartest?" but "which model completes my specific task at a cost my business model supports?" Inkling 2 Small's combination of a leading benchmark score and a $1.18 per-task cost is exactly the kind of trade-off that makes a model worth evaluating for production use.
The Open-Weights Question
One notable detail: the Apache 2.0 weights have not yet been released. Apache 2.0 is a permissive open-source licence, and releasing model weights under it would let anyone download, run, and modify the model. The weights not being released yet means that, for now, Inkling 2 Small is available as a hosted service rather than as something developers can run themselves.
Whether and when the weights appear matters for the open-model ecosystem. Open weights let startups self-host, fine-tune, and build without depending on a provider's API pricing or availability. Until they are released, Inkling 2 Small should be evaluated as an API offering — judged on its hosted price, performance, and reliability.
For now, the model stands as a useful data point: frontier-level capability, priced per task at a level many AI products can afford. Founders building agent products should add it to their evaluation list.
Angeled.com connects founders with investors.
aitoolsrecap.com, November 8, 2026
Read Next
OpenAI Releases Codex Security as Open-Source CLI and SDK
Databricks Unveils Always-On Genie, an AI Agent That Stops Waiting to Be Asked
DeepMind Launches AlphaGenome 2, a DNA Model That Reads Genomes Cell by Cell
ElevenLabs Unveils 13 New Products and Upgrades at Its New York Summit
Google Launches Gemini Decision 1, a Model That Answers Instead of Writing
Runway Open-Sources Praxis-1, Its Robot Model That Learned by Watching Video



