Small model near the leader: when intelligence is enough

On 30 July 2026, Artificial Analysis recorded Thinking Machines' Inkling Small at 40 points on the Intelligence Index, one point behind its frontier sibling Inkling, which scores 41, using less than a third of the total and active parameters: 276B total against 975B, 12B active against 41B. The index's weighted cost per task, on the model's Artificial Analysis page, is US$ 0.07. One point of distance does not, on most tasks, pay the size multiple that separates them. The cost-per-intelligence frontier moved, and that becomes a routing policy.

Key findings

The Artificial Analysis measurement of 30 July 2026 left Inkling Small one point behind the leader on the Intelligence Index, 40 against 41, with less than a third of the parameters and a cost of US$ 0.07 per index task. Parity, though, is selective by task type. Five numbers summarize the shift.

1. One point of difference. Inkling Small scores 40 on the Artificial Analysis Intelligence Index; Inkling scores 41.
2. Less than a third of the size. 276B total parameters (12B active) against 975B (41B active) for the larger sibling.
3. Cost-per-intelligence leadership. The small model's weighted cost per index task is US$ 0.07.
4. Selective parity by task type. The small model ties or beats the leader on coding and frontier reasoning, and loses on factual knowledge and agentic tasks.
5. Token efficiency. About 24 thousand output tokens per index task against 25 thousand for the larger sibling, well below peers at the same intelligence level.

How were the data collected?

The primary source is the independent Artificial Analysis write-up of 30 July 2026, plus the official Thinking Machines model card for Inkling Small. The measurement uses the Intelligence Index, which aggregates nine evaluations with their own weights; cost per task is derived from each provider's token prices.

The evaluations include GDPval-AA v2, Terminal-Bench v2.1, SciCode, Humanity's Last Exam, GPQA Diamond, CritPt, and AA-Omniscience. Cost per task divides each provider's input, output, and cache token prices by the number of tasks and weights the result by the evaluation's share of the index. This is an independent benchmark, not a vendor marketing number, with an honest limit: it measures model capability under index conditions, not the behavior of a live operation or of a specific provider. Price and latency vary by hosting. The sample is the index sample, not your load.

How many parameters does Inkling Small actually use?

The Thinking Machines model card reports 276 billion total parameters and 12 billion active parameters, in a Mixture-of-Experts architecture. The parallel matters because inference bills the active parameters per token, not the total. At each step, only a fraction of the specialists enters the bill. Total is the spec sheet; active is the invoice.

The inference-cost account is not the total parameter count on the spec sheet. It is what each run activates, times the token price, times the tokens the task actually used. The active number explains the small model's cost advantage: for most traffic, that is the real cost, and it is not double, it is a fraction. The active-parameter detail is what turns an architecture difference into a margin difference.

When is a small AI model enough?

A small model is enough when the independent capability measure puts it inside a margin that does not pay the leader's cost multiple, and when the task falls in a round type where it ties or wins. Inkling Small ties on coding and frontier reasoning.

Humanity's Last Exam 32% against 30%, GPQA Diamond 89% against 87%, CritPt 8% against 5%, SciCode 49% against 46%, and Terminal-Bench v2.1 at 55% for both. For those request families, the leader charges the invoice and delivers the same letter. One point does not pay that.

The economic reading is a cost per intelligence point. Taking 3.5 times fewer parameters to score one point behind becomes, in accounting language, the right question: how many of your requests actually require the missing point? A system that sends every request to the leader pays the size multiple across the entire base, including the ping, the greeting, and the registry lookup that the small model resolves on the first try. The leader stops being the default route on the day the point of difference does not show up in the result the customer sees.

How do you promote the small model to the default route?

The recommended policy is routing by cost per resolved task, not by a fixed model choice. That is a margin decision, not a benchmark decision. The conditions for promoting the small model to the default route have names and tests:

• Parity by task. The task has to fall in a round type where the small model ties or beats the leader. Otherwise it is not the route, it is a bet.
• Latency and context. The small model delivers 44.8 tokens per second on Artificial Analysis's reference provider, below the 66.1 t/s median of similar-size open-weights models. The model card and the Artificial Analysis model page report a 1 million token window; the 30 July 2026 article reports 256 thousand for Inkling Small and 1 million for Inkling. The route holds only if the SLO accepts the latency measured on your endpoint and the task fits the window that endpoint actually delivers.
• Cost per resolved task. The decisive metric is not the token price. It is what it costs to finish the right task on the first pass. The small model used ~24 thousand output tokens per index task, against ~25 thousand for the leader.

The condition that holds the other three is confidence that the cost is real. In a Brazilian operation, FX conversion and the effective cost of contracting the token abroad raise the invoice relative to list price, which makes shifting light traffic onto the light route more valuable, not less.

Published price is the list; cash leaving the account is another bill.

A model router with intent classification applies this policy per request: it normalizes, classifies, picks the model by cost, performance, latency, and context, and leaves the leader as a configurable reserve route. Nexforce Router is a practical example of that layer, with automatic failover, per-key spend limits, and traceability of every call. It occupies only the routing-infrastructure seat; it is not an agents product. The policy lives in the router.

When is the small model NOT enough?

The point of distance does not vanish. It moves to specific task types. The small model is not enough on sensitive factual knowledge, on agentic tasks, and on volume spikes with a latency SLO it does not deliver. Technical honesty requires naming them, one by one.

Inkling Small loses on factual knowledge and agentic tasks: AA-Omniscience -9 against the leader's 2, pulled by 31% accuracy against 40%, and τ³-Banking 15% against 24%. It is also slower at inference, at 44.8 tokens per second against the 66.1 median of similar-size models.

For a company, that means the small model is not enough when a wrong answer about a fact is expensive, when the task requires long interdependent action chains, when a volume spike demands speed the model does not deliver, or when context explodes beyond what the small model holds with margin. Keeping the leader as reserve is not conservatism. It is the other half of the same cost policy: pay the multiple only where the point of difference shows up in the result. Not every task fits the small model.

How much does cost per resolved task change the bill?

The difference shows up in the volume moved. An operational hypothesis, not a number measured on the index, is to shift most of the load off the leader's route onto the small model and keep the leader only where the point of difference appears. That drops cost per resolved task by an amount the token price table does not show. The table below completes the decision.

That table is a routing hypothesis, not a promise. The rule that holds it is the same on any invoice: sending all traffic to the most expensive model is paying for the point of difference across a base that does not use it. A multi-model proxy only forwards the request and bills the route; a router with a cost policy is what applies this table per request and records each choice for the auditor. The savings show up in the volume moved.

Frequently asked questions about small models versus the leader

When is a small AI model enough?

When the independent capability measure reaches the leader inside a margin that does not pay the cost multiple, and the task falls in a round type where it ties or beats the leader, such as coding and frontier reasoning. For sensitive factual knowledge, very long context, or agentic tasks with expensive failure, the leader remains the reserve route.

What is cost per intelligence for AI models?

It is the cost per intelligence point: the weighted cost per index task divided by the capability level reached. Inkling Small scores 40 points at US$ 0.07 per task, while the leader scores 41. The routing decision uses that gap to send most tasks to the light route and reserve the leader where the point of difference appears.

Is Inkling Small better than the frontier model?

On some measurements, yes. It ties or beats the larger sibling on Humanity's Last Exam, GPQA Diamond, CritPt, and SciCode, and loses on factual knowledge and agentic tasks. There is no single better model for everything; there is the model-and-task-type combination with the best cost per resolved activity.

How does Nexforce Router enter this decision?

Nexforce Router applies the cost policy per request: it normalizes, classifies intent, and picks the model by cost, performance, latency, and context, with the leader as the reserve route, automatic failover, per-key spend limits, and traceability of every call. It is a routing-infrastructure layer, not an agents product.

References and Further Reading

The Inkling Small parity analysis against Inkling comes from Artificial Analysis, published on 30 July 2026, with the Artificial Analysis model page for price, speed, and context, and the official Inkling Small model card at Thinking Machines for the parameters and the Apache 2.0 license. The whole account starts there. On the blog, the broader routing-as-cost-cut argument is in Token price falls, but AI cost rises, the request-complexity reading is in Complexity-based routing: quality without waste, endpoint variation is in Endpoint evaluation and routing policy, and gateway-layer choice is in How to evaluate and choose an LLM gateway. The structural shift in the frontier index is in Intelligence Index 2026: the end of the AI model duopoly.

The next step is a cost policy, not a scoreboard

A model that scores one point behind with a third of the size is not an architecture headline. It is a bill headline. The company that treats model choice as a single decision, taken once and forgotten, pays the leader's multiple across the entire base, while the light route, which resolves most tasks on the first pass, sits switched off in the code. The leftover intelligence point does not pay the invoice it costs. It pays only when the task requires it, and routing by task is exactly where it enters. The next decision is yours, and it is a budget decision, not a benchmark decision.