Tinker is Thinking Machines Lab's own fine-tuning and inference platform — and the only official route to Inkling 2 Small in its first week. Its price list is short: $0.30 in, $1.00 out. But Artificial Analysis measured $1.18 per task, nearly 6× the predecessor. This article works out three bills: the unit-price comparison, the per-task cost structure, and the free-vs-self-host trade-off. The headline: the unit price is not expensive; the output volume is. To save money, watch the output side — not the host, since every provider charges the same.
Tinker bills three token meters: training, sampling (output) and prefill (input). Inkling 2 Small's prices:
| Item (per 1M tokens) | Price | Notes |
|---|---|---|
| Serverless input (prefill) | $0.30 | Same for 64K and 256K configs |
| Serverless cached input | $0.03 | 90% off on cache hits |
| Serverless output (sampled) | $1.00 | Includes thinking tokens |
| Fine-tuning (64K context) | $1.50 | Requires the tml-renderers package |
| Fine-tuning (256K context) | $3.00 | Long-context fine-tuning costs double |
Two caveats: first, serverless inference is currently a waitlist beta, not instant-on; second, Tinker serves only 64K and 256K context configurations — the architecture's 1M limit is not available on Tinker, and needs a local deployment.
Together AI, Cloudflare Workers AI and OpenRouter all charge $0.30 in / $1.00 out for Inkling 2 Small, exactly matching Tinker; NVIDIA, Fireworks, Baseten, Nebius and Modal have also added it (prices not individually checked). There is no arbitrage on unit price — choose a host on latency, reliability, free allowances and context options, not price. What actually saves money is caching: in Plotly's test about 96% of input tokens came from the provider's cache, and the whole 43-question exam cost $1.81; multi-turn chats and repeated contexts benefit hugely from the $0.03 cached rate.
Artificial Analysis's independent measurement: $1.18 and 15.7 minutes per index task; the predecessor Inkling Small measured $0.20 and 7.0 minutes. Breaking it down: the $0.30 input price is only slightly above the $0.24 class median — AA called it "somewhat expensive," but it is not the main driver. The main driver is output volume: about 150K output tokens per task, 4× the predecessor's ~37K. At $1.00 per 1M tokens, 150K output tokens is $0.15, and multi-turn tool calls add input and thinking tokens on top. AA's words were "notably slow per task and very verbose" — slowness and verbosity are two sides of the same coin.
But note: $1.18 is the AA index-task figure (long-horizon, high effort), not a universal per-task cost. thoughtfullab's finance-diligence test produced a full memo for $0.69; Plotly's whole 43-question data-analysis exam cost $1.81. Different tasks differ by an order of magnitude — treating $1.18 as a general cost is wrong; treating it as an upper reference for long, high-effort tasks is right.
OpenCode's inkling-2-small-free is free for a limited time, and OpenRouter has a free variant (co-founder Soumith Chintala described it as "free in exchange for data use"). Free entry points suit three things: checking whether the model fits your task, comparing its output quality with your current model, and running small-scale tests. Note "limited time" — free allowances can change at any time, so don't put production traffic on a free variant.
The weights are $0 (Apache 2.0), but the NVFP4 checkpoint needs at least 500GB of aggregated VRAM (2× B300 or 4× H200) and MXFP8 needs 800GB. Rent or buy depends on volume: at low volume, hosted APIs are cheaper (even $1.18 per task does not add up to much); at high volume, long context and high concurrency, self-hosting has lower marginal cost — but factor in hardware depreciation, operations and power. Plotly notes the 4-bit weights need about 356GB of memory, "which fits on one 512 GB Mac Studio" — that is the weight files' footprint; the official 500GB figure presumably includes runtime overhead, so plan against the official number.
Five directly usable tips: first, use caching — repeated contexts and multi-turn chats should hit the $0.03 cached input rate; second, cap output length — output tokens are the bill's biggest part, keep them short; third, lower the effort level — xhigh scores look good but burn tokens fast, and everyday tasks don't need it; fourth, validate task fit on the free variants before paying; fifth, confirm context needs first — Tinker only serves 64K/256K, so 1M needs a local deployment, don't force it on Tinker.
Reference source: this article is an original rewrite produced by this site; facts and figures were checked against kie.ai's Inkling 2 Small specs, costs & benchmarks (around Nov 6, 2026)
← Back to the Inkling 2 Small topic page