Two years of silence. Then, a whisper. Mira Murati’s Thinking Machines Lab just dropped its first model, Inkling, onto OpenRouter. The only metric? An ‘impressive MCP score.’ No MMLU, no HumanEval, no SWE-bench. Just a protocol test.
In crypto terms, it’s like launching a token with only a social media following and no liquidity. My skepticism radar went off immediately. I’ve seen this playbook before—2017 ICOs with whitepapers that promised world peace but delivered nothing. DeFi Summer protocols shouting APYs without showing audited risk models. Now AI is doing the same: a single ambiguous metric dressed up as a breakthrough.
The Backstory
Mira Murati—former OpenAI CTO, known for her focus on safety and alignment—left in late 2023. Her new venture, Thinking Machines Lab, stayed dark until now. Inkling is their first output. The press calls it the ‘best Western open-source model,’ but the evidence is thinner than a bear market order book.
What is MCP? The Model Context Protocol. It’s not a standard benchmark like MMLU or HumanEval. It’s a protocol for tool calling—how well an AI can interact with external APIs, read files, manage context. Think of it as a test for agentic behavior: can the model book a flight, query a database, execute a trade? On that front, Inkling apparently shines.
But here’s the catch: MCP scores say nothing about reasoning, math, coding, or long-form generation. A model that scores high on MCP could be mediocre at everything else. It’s like a chef who can chop onions perfectly but burns every steak. Useful? Maybe. The best? Absolutely not.
The Core: What We Actually Know
Inkling is now live on OpenRouter, an API aggregator popular with developers. No pricing yet. No code repository. No technical paper. The only claim is ‘best Western open-source model,’ which is a loaded phrase.
‘Western’ deliberately excludes DeepSeek, Qwen, and other Chinese open-source heavyweights that have dominated leaderboards. DeepSeek-V2.5, for instance, matches GPT-4 on several benchmarks. Inkling doesn’t even try to compare. That’s a red flag the size of a flash loan exploit.
Let’s parse ‘open-source.’ In AI 2026, ‘open source’ can mean anything from full Apache 2.0 weights to a restrictive research license that forbids commercial use. We don’t know Inkling’s license yet. If it’s not truly open, the whole narrative collapses. I’ve audited enough protocols to know: the devil is in the license file.
Model size? Unspoken. Based on the quick deployment and focus on MCP, I’d bet it’s a small-to-medium model—7B to 30B parameters. Not the 405B beast that Meta’s Llama 3.1 is. That’s fine for specialized tasks, but calling it ‘best’ stretches credibility.
MCP is not a benchmark. It’s a protocol. The score reflects how well the model handles tool use. For crypto trading agents, that’s critically important. Imagine an AI that can connect to a DEX, read on-chain data, execute a swap, and manage risk. That’s MCP in action. If Inkling excels, it could be the engine for next-gen trading bots.
But I’ve been burned before. During DeFi Summer, protocols with flashy interfaces and high TVL hid reentrancy bugs. The same applies here: a high MCP score doesn’t mean the model is safe, cheap, or scalable. It doesn’t mean it won’t hallucinate and drain your wallet. Agent safety is a completely different ballgame.
The Contrarian Angle: Standard Play, Not Model Play
My contrarian take? Inkling’s real value isn’t the model. It’s the MCP protocol itself. Thinking Machines Lab might be trying to do what Uniswap did for AMMs: create the standard for agent interoperability. If MCP becomes the default way for AI to call tools, they own that narrative. The model is just a marketing vehicle to seed the protocol.
That’s a smart long-term play. But it’s also risky. Meta and Mistral are already working on agent frameworks. OpenAI has function calling built into GPT-4o. The window to own the standard is narrow.
Another unreported angle: the ‘best Western’ tag is a geopolitical signal. Western investors and developers want a counterweight to Chinese open-source dominance. Murati knows that. Her team is a who’s-who of former OpenAI safety researchers. That attracts money and talent. But emotional branding doesn’t replace technical rigor.
Takeaway: Watch for Code, Not Claims
For crypto traders and builders, Inkling matters because it could supercharge the AI agents that will soon run our markets. But right now, it’s a promise, not a product.
Sprint mode: Deactivated until I see the weights. I’ll be refreshing Hugging Face and waiting for standard benchmark scores. If Inkling delivers on MCP and general intelligence, we might have a new contender. If not, it’s just another hype cycle fueled by a single metric.
Real-time alert: Don’t bet your portfolio on a score. Let the code speak. I’ve seen too many projects with great PR and empty backends. The bear market taught me: survival comes before gains. Verify, then deploy.