Inkling 2 Small Pricing and API Access: What It Actually Costs
How to access Thinking Machines' Inkling 2 Small via API, what console credits you need, and whether the cost matches the performance.
What is Inkling 2 Small and how do you access it?
Inkling 2 Small is Thinking Machines’ second-generation frontier model, a mixture-of-experts system with 594 billion total parameters and 17 billion active parameters per inference. It reads both text and images, and the lab positions it as a general-purpose model with particular strength in coding and long-running agent tasks. Right now, the only way to use it is through Thinking Machines’ own console. You sign up, generate an API key, load credits onto the account, and call the model from there. There is no open-weight release, so running it locally or through a third-party inference host is not an option.
TL;DR
- Console-only access means anyone who wants to test Inkling 2 Small has to go through Thinking Machines’ own platform, generate an API key, and prepay with credits before sending a single request.
- No open weights exist yet, so self-hosting, quantized local builds, or third-party API resellers are not available options for this model.
- Cost is a real factor, with one independent tester describing the model as expensive to run compared to other frontier options, which matters for anyone budgeting API calls at scale.
- An autonomous fine-tuning demo from the lab reportedly cost around $426 and took about four hours for the model to build its own evaluator, training data, and fine-tuning loop on a forecasting task, a useful reference point for what agentic workloads might cost.
- Benchmark performance is mixed, with Inkling 2 Small trailing DeepSeek V3.1 Flash on coding benchmarks like Terminal-Bench but leading in vision, knowledge work, and finance-related tasks.
- It is not built for casual chat, so teams evaluating pricing should weigh it against the fact that Thinking Machines itself flags the model as less polished for back-and-forth conversation.
Without shadow IT.
How do you get API access to Inkling 2 Small?
Access runs entirely through the Thinking Machines console. The process is straightforward: create an account, generate an API key, and add credits to the balance before making calls. There is no free tier mentioned, no trial credits called out, and no alternative hosting path. If you want to experiment with the model, budgeting for credits upfront is part of the onboarding, not an afterthought.
This console-gated model differs from how many labs now ship frontier releases. Several competitors publish open weights alongside API access, letting developers run smaller variants locally or through cheaper third-party inference providers. Thinking Machines hasn’t done that with Inkling 2 Small. Everything funnels through their own metered system, which gives the lab full control over pricing and usage but also means there’s no way to shop around for a cheaper inference route.
Is Inkling 2 Small expensive to run?
Based on hands-on testing reported by one independent reviewer, yes, the model is pricier than many alternatives. The exact per-token rate wasn’t disclosed in available testing, but the tester’s blunt assessment was that credits burn fast and the model is “not cheap.” That matters because 594 billion total parameters with 17 billion active parameters puts this solidly in large-MoE territory, and compute costs at that scale tend to show up directly in API pricing.
The clearest cost data point comes from an agentic workload the lab itself demonstrated: tasking the model with training a small forecasting system for electricity demand, wind power, air quality, and Wikipedia page views, entirely on its own. The model wrote its own evaluator, generated training data, ran fine-tuning, attempted reinforcement learning, and debugged its own mistakes without step-by-step human guidance. That autonomous run took about four hours and cost roughly $426, and it more than doubled the forecasting score by the end. For teams considering using Inkling 2 Small for similar self-directed fine-tuning tasks, that figure is the closest thing to a real-world cost benchmark currently available. It’s a meaningful number: letting a model manage its own training loop removes a lot of manual SFT work, but the compute bill for that autonomy is not trivial.
What do you actually get for the cost?
Vision handling stood out in testing. In one example, the model was shown a single photo of three glasses of red wine and asked to identify the grape variety and vintage for each glass. It got all three correct in 19 out of 20 trials, reading that level of detail from an image alone. That kind of visual precision is central to why this release leans so heavily on vision capability.
Coding and generative tasks were also tested directly through the API. A single text prompt asking for a self-contained HTML simulation of a rotating doner kebab, with no external libraries or pre-made image assets, produced a visually coherent result: a heating element with rendered flames, a spit with layered meat texture, and rotation animation, all generated from scratch. The flames looked convincing; the meat texture and cone shape had room for improvement.
Multilingual capability showed up strongly too. Given one English proverb (“don’t count your chickens before they hatch”), the model produced a clean, searchable table translating it into a long list of languages, including lower-resource and regional ones like Seraiki and Nepali, with scripts and writing directions rendered correctly across most entries.
The most demanding test involved a hand-drawn circuit diagram fed to the model through the API, with a request to simulate the circuit and fix any errors in the original drawing. The model returned a working interactive simulator with an input frequency slider, a live waveform, and a magnitude plot, correctly calculating a cutoff frequency based on the resistor and capacitor values shown. It also identified errors in the original hand-drawn circuit, explained why each one mattered, proposed fixes, and flagged which parts of the design were already correct, without being asked to do that last part.
How does Inkling 2 Small compare to other models on benchmarks?
The benchmark picture is mixed rather than uniformly strong. On the Artificial Analysis Index and on several coding benchmarks, including Terminal-Bench, DeepSeek V3.1 Flash leads by a wide margin: Inkling 2 Small scored 13.1% versus DeepSeek’s 33.8% on Terminal-Bench specifically. That’s a significant gap in agentic coding performance, and it suggests Inkling 2 Small is not yet the strongest option for pure coding workloads despite the attention this release gives to long-running agent tasks.
Where Inkling 2 Small pulls ahead is vision, knowledge work, and finance-related benchmarks, topping several of those categories. Thinking Machines itself has also been upfront about the model’s weaknesses: it’s not the most concise communicator, it has no audio support, and it’s less polished as a conversational chat model. For anyone planning to use it as a general question-and-answer assistant, that’s a direct signal from the lab that other models will likely serve that use case better.
Is Inkling 2 Small worth the cost?
It depends heavily on the workload. For vision-heavy tasks, multilingual content generation, or STEM-style reasoning (like diagnosing and simulating circuit designs from a sketch), the model performs well enough to justify testing it despite the price. For straightforward coding agent work, the benchmark gap against DeepSeek V3.1 Flash is hard to ignore, especially given DeepSeek’s lower reported cost in general industry comparisons. For casual conversational use, the lab’s own stated limitations suggest looking elsewhere.
Anyone evaluating this model for production use should treat the console-only, credit-based access as a cost planning exercise from day one: estimate token volume, factor in the model’s apparent premium pricing, and weigh that against the specific tasks (vision, finance, autonomous fine-tuning) where it has actually shown an edge.
Frequently Asked Questions
How many parameters does Inkling 2 Small have?
It has 594 billion total parameters in a mixture-of-experts architecture, with 17 billion active parameters used during any single inference pass.
Can I run Inkling 2 Small locally or through other API providers?
- Patient records✓ Granted
- Surgery schedule✓ Granted
- Billing export✕ Denied
No. There are currently no open weights for the model, and access is limited to Thinking Machines’ own console and API, which requires an account, an API key, and prepaid credits.
Is Inkling 2 Small good for coding tasks?
It’s mixed. On Terminal-Bench, it scored notably lower than DeepSeek V3.1 Flash (13.1% versus 33.8%), suggesting it lags behind leading models on coding-specific agentic benchmarks despite the release’s emphasis on agent work.
What was the real-world cost of an agentic task using this model?
In one demonstration, the model autonomously built an evaluator, generated training data, and fine-tuned a forecasting model across electricity demand, wind power, air quality, and Wikipedia traffic data. The process took about four hours and cost approximately $426.
What is Inkling 2 Small best suited for?
Based on available testing, it performs strongest on vision tasks, multilingual content generation, and STEM-style reasoning such as circuit analysis. It is explicitly weaker as a conversational chat model and lacks audio support.