> ## Content Index
> Fetch the complete content index at: https://wire.fourthweb.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# Murati's Inkling Beats GPT-4 on Benchmarks at Half the Cost
- URL: https://wire.fourthweb.ai/muratis-inkling-beats-gpt-4-on-benchmarks-at-half-the-cost/
- Published: 2026-07-26T14:01:03.000Z
- Updated: 2026-07-28T03:05:41.000Z
- Description: The former OpenAI CTO just dropped a model that costs less to run than GPT-4 and scores higher on the benchmark Silicon Valley actually cares about. Thinking Machines Lab released Inkling, Mira Murati's first model after two years of building in silence since leaving OpenAI
- Author: Travis Wright
- Tags: Real World Assets, Agentic Workflows, AI Agents, AI Infrastructure, OpenAI, Anthropic

**The former** [**OpenAI**](https://wire.fourthweb.ai/tag/openai/) **CTO just dropped a model that costs less to run than GPT-4 and scores higher on the benchmark Silicon Valley actually cares about.**

### The Summary

- [Thinking Machines Lab released Inkling](https://decrypt.co/373884/review-inkling-mira-murati-first-open-source-ai?ref=wire.fourthweb.ai), Mira Murati's first model after two years of building in silence since leaving OpenAI
- [The model is available on OpenRouter](https://decrypt.co/373884/review-inkling-mira-murati-first-open-source-ai?ref=wire.fourthweb.ai) with an MCP score that legitimately impresses, outperforming closed models in its class
- Price-to-performance looks good on paper but the real cost story is more nuanced than the headline numbers suggest

### The Signal

Mira Murati left OpenAI in 2024 as CTO, one of the most visible executives in AI. She went quiet. No podcasts, no victory laps, no "stealth startup" LinkedIn updates. [Thinking Machines Lab spent two years building](https://decrypt.co/373884/review-inkling-mira-murati-first-open-source-ai?ref=wire.fourthweb.ai) and now we know what they were building: Inkling, an open-weights model that's live on OpenRouter and scoring higher than it has any business scoring.

The MCP benchmark, which measures multi-step reasoning and context handling, puts Inkling ahead of GPT-4 and competitive with [Claude](https://wire.fourthweb.ai/tag/anthropic/) 3 Opus in specific tasks. That matters because MCP correlates better with real-world agent performance than older benchmarks like MMLU. If you're building agents that need to chain actions across tools, this is the score you watch.

> "The MCP score is genuinely impressive."

But price-to-performance gets complicated fast. Inkling's inference costs look attractive per token compared to frontier closed models. The catch: [the math gets more complicated](https://decrypt.co/373884/review-inkling-mira-murati-first-open-source-ai?ref=wire.fourthweb.ai) when you factor in the full operational picture. Open-weights means you can self-host, which shifts costs from API calls to infrastructure. For high-volume applications, that's a win. For teams without ML ops capacity, it's a hidden tax.

The model architecture details aren't public yet, but the performance profile suggests heavy optimization for reasoning tasks over raw speed. This tracks with Murati's background, where she pushed hard on RLHF and model alignment at OpenAI. Inkling appears purpose-built for applications where getting the right answer matters more than getting any answer fast.

**Key trade-offs for builders:**

- Lower per-token cost than GPT-4, but self-hosting adds infrastructure overhead
- Stronger reasoning scores, potentially slower inference on consumer hardware
- Open weights mean customization and fine-tuning without vendor lock-in

### The Implication

If Inkling's benchmarks hold up in production, Thinking Machines Lab just changed the economics for anyone building agent infrastructure. The open-weights release matters less for hobbyists and more for companies tired of watching their OpenAI bills scale linearly with success. Watch what happens in the next 90 days. If Inkling gets picked up by agent frameworks like LangChain or CrewAI as a default option, that's validation. If it stays niche, the price-to-performance story wasn't compelling enough to overcome OpenAI's distribution advantage.

The real test: do people build businesses on it, or do they just benchmark it and move on.

### Sources

[Decrypt](https://decrypt.co/373884/review-inkling-mira-murati-first-open-source-ai?ref=wire.fourthweb.ai)