ModelPriceWatch.com
Last scan 2026-11-22 Models tracked 322 Providers 42 Cheapest paid pplx-decider-flash-v1.3 $0.010/Mtok in Every price links to its source

Inkling 2 Mini

by Thinking Machines · 26.4B total / 3.4B active parameters

Current new · 2d budget open weights cheap tier

Today's price · per 1M tokens

Input

$0.150

per 1M tokens

Output

$0.500

per 1M tokens

Blended

$0.237

blended $/1M — 3:1 weighted input:output

Cached input

$0.015

10% of input — prompt caching

How this price is scoped: Tinker Models & Pricing table sampling rate (prefill / cached prefill / sample) for thinkingmachines/Inkling-2-Mini, called through the Tinker sampling client. The 64K and 256K training variants print the same inference rates and a 1M sampling context; only the training rate differs ($0.75 vs $1.50 per 1M). Third-party hosts on OpenRouter charge a different rate (PrimeIntellect fp4: $0.10/$0.40); that is not this price.

Source: official Thinking Machines pricing · read MODELPRICEWATCH.COM · 2026-11-22

Price receipt

We read Thinking Machines's own pricing page on and found Inkling 2 Mini (64K train / 1M sample) listed with a price on the same row — that read is where the number above comes from, recorded as a new model. We keep the page text we read; its content hash is c861e64d31.

Overview

The smallest and cheapest Inkling: a 26.4B-total / 3.4B-active Mixture-of-Experts looped transformer that runs its middle layers three times per token (5.4B weight activations per token), with image input. Thinking Machines pruned and healed it from a larger model, then trained it with the same SFT mix and on-policy distillation as Inkling 2 Small, and says it is small enough to run on a laptop. Open weights under Apache 2.0 on Hugging Face. Sampled on Tinker at $0.15/$0.50 per 1M tokens, with cached input at $0.015 — half the Inkling 2 Small rate.

Run it yourself

Deploy this open model on rented GPUs

Open weights mean you can self-host instead of paying per-token API prices. These platforms let you serve it on demand — often cheaper at scale.

Partner links — we may earn a commission if you sign up, at no cost to you. We only list platforms we'd recommend regardless.

Capabilities

struck through = not supported
Input 2/4
Text ✓ Image ✓ Audio Video
Output
Text ✓
Features 5/9
Prompt caching ✓ Reasoning ✓ Coding Fast inference Long context ✓ Open weights ✓ Multimodal ✓ Web search Realtime

Benchmark performance

accuracy % · higher is better
Percentile vs all tracked models 38.4th3 independent measurements
Percentile per $/Mtok 160
SWE-Bench Verified vendor
79.2%
Terminal-Bench 2.1 vendor
71.6%
MCP Atlas vendor
82.7%
BrowseComp vendor
79%

Sources: Thinking Machines Inkling 2 Mini launch post, SWE-bench Verified, bash only (SWE-Bench Verified) · Thinking Machines Inkling 2 Mini launch post, Terminal Bench v2.1, pi v0.83.0 harness (Terminal-Bench 2.1) · Thinking Machines Inkling 2 Mini launch post, MCP Atlas public split (MCP Atlas) · Thinking Machines Inkling 2 Mini launch post, with context compaction (BrowseComp)

Every per-benchmark score we hold for this model is reported by its own vendor, so none is counted toward an accuracy average or any ranking. The percentile above is its standing across the independent composites below.

Independent composite scores
  • AA Intelligence Index: 20.1 (Artificial Analysis composite across reasoning, knowledge and coding evals)
  • AA-Omniscience Index: -8.6 (−100–100)
  • GDPval-AA v2: 1318.5 ELO (Real-world work tasks, human baseline = 1000)

Source: Artificial Analysis (artificialanalysis.ai)

Artificial Analysis evaluations
  • HLE-Diamond: 15.2% (measured 2026-11-22) — A revised 1,000-question set of Humanity's Last Exam, half reasoning, half knowledge
  • CritPt v2, hard subset: 17% (measured 2026-11-22) — The hardest research-level physics problems in CritPt v2, solved with a code sandbox
  • FrontierCode-AA 2.0: 15.1% (measured 2026-11-22) — Coding tasks in real open-source repositories, graded on whether a maintainer would merge the change
  • Terminal-Bench-Science: 1.4% (measured 2026-11-22) — Expert-written science research workflows completed in a terminal
  • GDP.pdf: 16.8% (measured 2026-11-22) — Tasks over long professional documents; share of tasks that pass every criterion
  • AutomationBench-AA: 41.8% (measured 2026-11-22) — Multi-step work across simulated business apps, without breaking the rules set
  • Terminal-Bench 4.0: 5.6% (measured 2026-11-22)
  • Cost to run the AA Intelligence Index: $0.587 per task, 9 min per task
  • Measured by AA: 279.9 tokens/sec output, 1.49s to first token, 9.72s to first answer token

Source: Artificial Analysis (artificialanalysis.ai)

How it stacks up
  • Ranks #112 of 207 comparably-measured models by percentile score, across 3 independent measurements
  • Ranks #16 of 207 comparably-tested models by normalized performance per dollar

279.9 tokens/sec output 1.49s latency to first token (TTFT)

Throughput: Artificial Analysis, measured 2026-11-22 · Source: Artificial Analysis (artificialanalysis.ai)

Source: Artificial Analysis, Thinking Machines Inkling 2 Mini launch post, MCP Atlas public split, Thinking Machines Inkling 2 Mini launch post, SWE-bench Verified, bash only · updated · See full rankings →

Specifications

Context window
1M tokens
Modality
text, image
Parameters
26.4B total / 3.4B active
Open source
Yes — open weights available
Released
Status
Current
Last updated
Tags
budgetreasoningmultimodalopen-weightsmoecaching

Availability verified: — listed on Thinking Machines's own page