Thinking Machines Lab, the startup led by former OpenAI technology chief Mira Murati, released Inkling 2 Small on Thursday, its second model generation, and made it available for fine-tuning and inference on its Tinker platform. The company’s launch post describes a Mixture-of-Experts model with 594 billion total parameters and 17 billion active, which accepts image inputs. It scores 31 on the Artificial Analysis Intelligence Index, against 12 for the first Inkling model released in July, according to Artificial Analysis.
Inkling 2 Small is listed at $0.30 per million input tokens and $1.00 per million output tokens, while the first Inkling (at its Xhigh setting) is listed at $1.00 and $4.05, Artificial Analysis shows. By our arithmetic, that is a 75% lower output price and a 70% lower input price, for an index score 2.6 times as high.
The key insight: Thinking Machines Lab released Inkling 2 Small, a 594B-parameter Mixture-of-Experts with 17B active, on its Tinker platform. It scores 31 on the Artificial Analysis Intelligence Index, up from 12 for the first Inkling, and by our arithmetic its listed output price is 75% lower.
Thinking Machines Lab’s X post at 23:43 UTC on 5 November: Inkling 2 Small scores 31 on the Artificial Analysis Intelligence Index and is available for fine-tuning and inference on Tinker, with full weights planned.
Inkling 2 Small is available today on Tinker, with weights coming soon. It's a capable small model. We look forward to seeing what you make with it. https://t.co/fjgRc364Ii
Mira Murati, a minute later: the model is available on Tinker today, with weights coming soon.
The model: a Mixture-of-Experts with 594B total and 17B active parameters that supports image inputs and variable effort, per the launch post. Tinker lists the first Inkling at 975B total and 41B active parameters.
Access: initially through Tinker only, “with closed weights.” The company says it intends “to make the full weights available under an Apache 2.0 license next week after the completion of external safety testing and red-teaming.”
Training: the first pretraining phase covered 8.4 trillion tokens and “ran for 68 hours on NVIDIA GB300 NVL72 racks,” natively in NVFP4, the post says.
Its own benchmark table: on SWE-Bench Pro v2 (Hard) it reports 86.3%, second to Kimi K3’s 88.2% in its open-model comparison set; on its kernel-optimisation test it solved 54.25 of 69 tasks on average, against 63 for GPT-6 Astra and 60 for Claude Opus 5.5.
Limitations the company lists: it is “not the most succinct communicator,” does not support audio, and is “less optimized for the chat experience.”
Artificial Analysis Intelligence Index scores for Thinking Machines Lab’s first Inkling model (July 2026, Xhigh setting) and Inkling 2 Small (November 2026), with Artificial Analysis’s median for comparable models. Source: Artificial Analysis model pages, read 6 November 2026.
Artificial Analysis ranks Inkling 2 Small fourth of 29 comparable models on its index, above the median of 23, and calls it “notably slow per task and very verbose”: it took 20.8 minutes per index task and generated 154 million output tokens across the evaluation, against a median of 89.8 million. It cost $1.18 per index task to evaluate, against $1.35 for the first Inkling, by Artificial Analysis’s count.
The first Inkling was also among the 22 models in Visa’s agentic-commerce benchmark, where it passed 28.9% of shopping runs, the lowest pass rate of the 22; we covered the results in Visa: AI Shopping Agents Fail 42.8% of Runs, Often Still Buy. Visa did not test Inkling 2 Small.
Business Pill · Usage-Based Pricing
A short explainer on usage-based pricing and how it differs from per-seat pricing. It teaches the general idea only and does not discuss Thinking Machines.
Update, 6 Nov, 00:58Z: how Artificial Analysis ranks it. In its own write-up of 5 November, Artificial Analysis said Inkling 2 Small “matches Google DeepMind’s Gemini 3.8 Flash (high) and surpasses Xiaomi’s MiMo-V2.6-Pro” on its Intelligence Index v5.0.1, and that Thinking Machines’ “plan to release its weights would make it the highest-scoring open weights model developed outside China based on current rankings.” That ranking is conditional: Artificial Analysis lists the model as proprietary at launch, with an Apache 2.0 weights release planned and its date unconfirmed.
Artificial Analysis also put the cost of running its index at $1.18 per task for Inkling 2 Small, against $7.77 for Gemini 3.8 Flash (high) and $10.02 for Claude Sonnet 5.5 (medium), while Inkling 2 Small used about 150,000 output tokens per task against about 99,000 for Gemini 3.8 Flash (high). By our arithmetic, at a matching index score, Gemini 3.8 Flash (high) cost about 6.6 times as much per task, even though Inkling 2 Small used about 1.5 times as many output tokens.
As we read it, the “Small” label is relative: at 594B total parameters it is smaller than the 975B first Inkling, but runs 17B active parameters per token against 41B.
As we read the launch, Tinker comes first: the model is sold for fine-tuning and inference there before the weights are released.
As we read Artificial Analysis’s data, the gain came with more tokens per task: 154 million output tokens across the index, against a median of 89.8 million.
Thinking Machines Lab, 5 November 2026
“We intend to make the full weights available under an Apache 2.0 license next week after the completion of external safety testing and red-teaming.”
Three Implications
FOR TINKER CUSTOMERS The model is available for fine-tuning and sampling now; checkpoints become downloadable with the weight release, the company says.
FOR OPEN-WEIGHT RIVALS The launch post compares it with GLM-5.3-Flash, DeepSeek-V4.1-Flash, Muse Spark Open and Kimi K3 on coding benchmarks.
FOR BUYERS COMPARING PRICES Artificial Analysis lists it at $0.30 per million input tokens and $1.00 per million output tokens.
The framework says: “The winning strategy in every major platform shift has followed a consistent pattern: Close the SCARCE layer (keep proprietary) + Open the ABUNDANT layer (commoditize)”
As we read it, Thinking Machines plans to open the model weights while selling the fine-tuning and serving service, Tinker. The framework itself does not discuss Thinking Machines.
What Is Not Established
The weights are not yet public; the Apache 2.0 release is planned for next week and depends on safety testing finishing, per the company. Thinking Machines has not disclosed revenue, Tinker usage or customer numbers in the launch post. The benchmark tables in the launch post are the company’s own selection and comparison set. The price comparison uses Artificial Analysis’s listed prices for each model; the 75%, 70% and 2.6x figures are our arithmetic.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
Thinking Machines’ second model generation, Inkling 2 Small, scores 31 on the Artificial Analysis Intelligence Index, up from 12 for its first Inkling, and is listed at a 75% lower output price by our arithmetic. It runs on Tinker now, with closed weights, and the company plans an Apache 2.0 weight release next week.
A note on sourcing. Model details, training facts, benchmark tables, limitations and the weight-release plan are from Thinking Machines Lab’s 5 November 2026 launch post; parameter counts for the first Inkling are from Tinker’s product page. Index scores, rankings, speeds, token counts and prices are from Artificial Analysis’s model pages for Inkling 2 Small and Inkling (Xhigh), read 6 November 2026. The Visa result is from Visa’s published benchmark data. The 75%, 70% and 2.6x figures are our arithmetic.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
