Key points
- Inkling 2 Small scores 31 on the Artificial Analysis Intelligence Index, an 18-point increase over its predecessor.
- The model costs $1.18 per task, nearly six times more than the previous version, due to higher token output.
- Agentic task performance jumps from 5 to 46 percent on AutomationBench-AA, while the response rate drops to 53 percent.
Thinking Machines states that Inkling 2 Small is the strongest Western open-weight model, citing benchmark scores that exceed its predecessor significantly.reported The startup, founded by former OpenAI CTO Mira Murati, released this mixture-of-experts model with 594 billion parameters, 17 billion of which are active at any given time.reportedThe new model performs better on complex tasks but costs much more to run than the previous version.Inkling 2 Small demonstrates significant agentic capability gains at the expense of higher token throughput and cost efficiency.
Evidence: reported
Thinking Machines' Inkling 2 Small claims top spot among western open-weight modelsaifleet-decoder
Evidence: reported
Thinking Machines' Inkling 2 Small claims top spot among western open-weight modelsaifleet-decoder
The claim
Thinking Machines reports that Inkling 2 Small achieves a score of 31 on the Artificial Analysis Intelligence Index.reported This represents an 18-point jump from Inkling Small’s score of 13 and places it roughly on par with Google’s Gemini 3.8 Flash.reported The company highlights substantial gains in agentic tasks, where the model independently executes multi-step workflows.
Evidence: reported
Thinking Machines' Inkling 2 Small claims top spot among western open-weight modelsaifleet-decoder
Evidence: reported
Thinking Machines' Inkling 2 Small claims top spot among western open-weight modelsaifleet-decoder
The response
The performance improvements come with increased operational costs. According to Artificial Analysis, Inkling 2 Small generates approximately 150,000 output tokens per task, about four times more than its predecessor.reported Consequently, the cost per task rises to $1.18, which is nearly six times higher than the previous model.reported
Evidence: reported
Thinking Machines' Inkling 2 Small claims top spot among western open-weight modelsaifleet-decoder
Evidence: reported
Thinking Machines' Inkling 2 Small claims top spot among western open-weight modelsaifleet-decoder
The bottom line for teams
The model is initially available as a proprietary offering through the Tinker platform. Thinking Machines indicates that the weights will be released under an Apache license in the future. This phased availability allows for immediate testing while preparing for broader open-weight adoption.
What the paper shows
The research tests Inkling 2 Small’s performance against its predecessor and competitors using specific benchmarks.reported Reported results show a score of 31 on the Artificial Analysis Intelligence Index and significant gains in agentic tasks.reported Stated limitations include a drop in response rate from 77 to 53 percent and a sixfold increase in cost per task.
Evidence: reported
Thinking Machines' Inkling 2 Small claims top spot among western open-weight modelsaifleet-decoder
Evidence: reported
Thinking Machines' Inkling 2 Small claims top spot among western open-weight modelsaifleet-decoder
Why it matters
- Evaluators can compare Inkling 2 Small’s agentic capabilities directly against its predecessor using AutomationBench-AA scores.
- Teams must account for the sixfold cost increase when planning budgets for high-throughput agentic workflows.
- The drop in response rate suggests a trade-off between thoroughness and coverage in knowledge-work tasks.
- Future Apache licensing may allow independent verification of the claimed open-weight superiority.
Topics: models, benchmarks, agentic AI, open-weight, Thinking Machines
Updated
Then / now / next
Inkling Small scored 13 on the Artificial Analysis Intelligence Index and cost significantly less per task.
Inkling 2 Small scores 31, matching Gemini 3.8 Flash, but costs six times more due to higher token usage.
Teams will monitor the release of Apache-licensed weights to verify open-weight claims independently.
Story timeline
- Then Inkling Small scored 13 on the Artificial Analysis Intelligence Index and cost significantly less per task.
- 06 Nov 2026 Thinking Machines' Inkling 2 Small claims top spot among western open-weight models
- Now Inkling 2 Small scores 31, matching Gemini 3.8 Flash, but costs six times more due to higher token usage.
- Next Teams will monitor the release of Apache-licensed weights to verify open-weight claims independently.
The strongest doubt
Counterpoint from the sources
The hallucination rate drop from 61 to 27 percent is largely attributed to the model answering fewer questions, not necessarily improved accuracy.
For readers
Researchers
- Benchmark scores allow direct comparison with Gemini 3.8 Flash on the Artificial Analysis Intelligence Index.
- The drop in response rate provides data on trade-offs between thoroughness and coverage.
Developers
- Higher token usage requires adjusting infrastructure for models generating 150,000 output tokens per task.
- Apache licensing will enable fine-tuning for specialized applications once weights are released.
Investors
- Cost per task increases sixfold, impacting unit economics for high-volume agentic use cases.
- Thinking Machines positions itself as a leader in Western open-weight models.
Numbers in context
Circle area is proportional to the magnitude; labels retain signs. Figures may describe different measures. Read their contexts above.Ask the article
Local keyword search of this article; no generated answers.Did this article give you the evidence you needed?
Your choice stays on this device. No public tally.Terms in this story
Mixture-of-experts
A model architecture where only a subset of parameters are active for each input.
Agentic tasks
Workflows where the model independently executes multiple steps without constant user guidance.
Artificial Analysis Intelligence Index
A benchmark used to measure the overall performance of AI models.
Questions readers ask
What is the cost difference between Inkling 2 Small and its predecessor?
Inkling 2 Small costs $1.18 per task, which is nearly six times more than the previous model.
How does Inkling 2 Small perform on agentic benchmarks?
It scores 46 percent on AutomationBench-AA, up from 5 percent, and 25 percent on FrontierCode 2.0.
When will the model weights be available under an Apache license?
The weights are expected to follow the initial proprietary release, but no specific date is provided.
Connected stories
Reading receipt
- Inkling 2 Small scores 31 on the Artificial Analysis Intelligence Index, an 18-point increase over its predecessor.
- The model costs $1.18 per task, nearly six times more than the previous version, due to higher token output.
- Agentic task performance jumps from 5 to 46 percent on AutomationBench-AA, while the response rate drops to 53 percent.
Still open: When exactly will the Apache-licensed weights be released?
Still open: How does the model perform on benchmarks not cited by Thinking Machines?