The chart is a lie. And so is the benchmark that nobody can verify. Last week, a blockchain-focused outlet ran a piece attributing Thinking Machines Lab—Mira Murati's first post-OpenAI venture—as having unleashed "Inkling," a model they call the "best Western open-source AI model" based solely on its MCP score. The article offered zero standard benchmarks, no architecture details, and no comparison against Llama 3.1, Mistral Large, or DeepSeek-V3. It read like a crypto whitepaper from 2017: bold claims, scant evidence, and a single non-standard metric designed to defy peer review. As someone who spent 2017 dissecting EOS and Tezos ICO whitepapers—tracking how "decentralization fatigue" was being reframed as "developer experience"—I recognize the pattern. The narrative is being constructed before the product. And the arbitrage lies in understanding where the story breaks before the price reacts.
Context: The Murati Effect and the Attention Economy
Mira Murati is not a name that requires introduction. As the former CTO of OpenAI, she was the public face of ChatGPT's safety alignment and the steward of a company that redefined the AI landscape. Her departure in late 2024 was met with speculation about a new venture. That venture is Thinking Machines Lab, and Inkling is its first model. The narrative hook is impeccable: a top-tier AI mind leaving the safety of Big AI to build something "open" and "Western"—a direct counter-narrative to the rise of Chinese open-source models like DeepSeek-V3 and Qwen 2.5. The crypto-native media outlet that broke the story is not an AI publication; it's a blockchain and Web3 news source. This domain mismatch is the first red flag. In 2021, I tracked 15,000 Ethereum transactions to map social capital accumulation in BAYC and realized that when a project chooses an unconventional press channel, it's usually because the conventional channels require scrutiny they can't withstand. The article's focus on "MCP score" as the sole claim-to-fame is reminiscent of how DeFi projects in 2020 touted "total value locked" (TVL) as the ultimate metric—ignoring liquidity mining incentives that made TVL a mirage. Every chart is a story waiting to be corrected. And this one is being written in sand.
Core: The MCP Score – A Metric Built for Ambiguity
MCP stands for Model Context Protocol, a relatively new benchmark designed to evaluate a model's ability to manage context in tool-calling and agent workflows. It's not part of the standard AI evaluation suite—no MMLU, no HumanEval, no GSM8K. It's a protocol-specific test, likely developed by Thinking Machines Lab themselves or by a small consortium. The article never specifies who administered the test, how many questions were asked, or what the error bars are. Based on my experience auditing Compound's governance token distribution in 2020, I learned that when a protocol creates its own metric and claims dominance, you should immediately model the inflationary pressure on that narrative. The MCP score is to Inkling what "impermanent loss" was to SushiSwap: a complex, non-obvious concept that can be spun to sound like a strength while masking underlying weakness. In the Compound case, high APYs were masking solvency risks; here, a high MCP score may mask generic reasoning deficits. The article's claim that Inkling is "the best Western open-source model" is a statement of unspecified geography (why Western?) and unspecified scope (open-source compared to what?). It's a narrative hedge, designed to sound confident while leaving escape clauses. The model's size is not mentioned, nor its training compute, nor its dataset. In the crypto world, this is equivalent to a Layer-2 project claiming "infinite scalability" without specifying the validator set or the bridge security model. Liquidity is a mirror, not a foundation. And right now, that mirror is reflecting a lot of marketing dust.
Let's dive deeper into the MCP protocol. It's an open standard promoted by Anthropic and a few other labs to standardize how models connect to external tools and databases. A model that scores high on MCP is good at following function-calling instructions, maintaining state across multi-turn conversations, and correctly formatting outputs for APIs. That's valuable, but it's a narrow capability. It doesn't tell you if the model can write poetry, solve complex math problems, or generate accurate legal summaries. The article's author—likely a blockchain journalist with limited AI expertise—may have been swayed by the sophistication of the term, much like a retail investor in 2021 was swayed by "TVL" without understanding its composition. The lack of any mention of model size is particularly damning. If Inkling is a 7B parameter model, it cannot compete with Llama 3.1 405B on general knowledge, no matter how good its MCP score. If it's a 70B model, its MCP score might still be inflated by overfitting to the test. The silence on this is not an oversight; it's a deliberate narrative architecture. Decoding the narrative before the price reacts is what I do. And the price here is the attention currency that Thinking Machines Lab is minting. The article's purpose is not to inform but to generate a narrative boon that attracts developers, investors, and talent. It's a liquidity event in the attention economy.
Contrarian: What If the Model Is Actually Good?
The contrarian angle is that Inkling might genuinely be the best open-source model for agent tasks. Mira Murati's team likely includes top-tier alignment and agent researchers from OpenAI and DeepMind. If they focused intensely on tool-calling alignment—something that many AI labs treat as a secondary concern—they could have achieved a breakthrough. The MCP score might be a real signal, not a mirage. But even if that's true, the way the information is being released is dangerous. In 2022, during the FTX collapse, I mapped the "hubris narrative" that led to the crash by interviewing 30 former executives. I found that FTX's brand story outpaced its financial reality by 18 months. The same dynamic is at play here: the narrative of Inkling as the "best Western open-source model" is being baked into the market's mind before any third-party verification. If the model is eventually proven to be mediocre, the backlash will be fierce—but by then, Thinking Machines Lab will have already captured talent and maybe raised a large round based on that narrative. The arbitrage lies in understanding human fear. Fear of missing out on the next big AI model is driving developers to engage with Inkling prematurely, just as fear of missing out on DeFi yields drove capital into protocols that collapsed. The article's omission of standard benchmarks is not a regulatory violation, but it is an ethical lapse. It's the kind of opaque reporting that crypto news outlets are famous for, where the line between journalism and marketing is thin enough to walk a tightrope. Illusions break; logic remains. And logic says that without verifiable benchmarks across the full AI evaluation suite, the "best" claim is hollow.
Furthermore, the choice to launch on OpenRouter—a platform that aggregates model APIs—suggests a strategy focused on developer adoption over brand creation. OpenRouter is the "Uniswap of AI models": a neutral marketplace where developers can test multiple models without heavy commitment. That's smart for distribution but signals that Thinking Machines Lab is not yet ready to own the customer relationship. It's an outsourcing of trust to a third-party platform, much like a crypto project listing on a DEX before its own exchange. The article does not mention pricing, though it notes "cost-effectiveness is complex to calculate." That complexity is a flag. In my experience with yield farming, when a protocol says "rewards are complex to calculate," it often means the yield is unsustainable and the math only works in the short term. Here, "complex to calculate" may mean that Inkling's API pricing is either artificially low to buy market share or so high that the value proposition collapses. Without transparency, the community is left to gamble. Who owns the attention? Follow the capital. The capital that funded this article—whether direct payment to the news outlet or a reciprocal arrangement—is betting on narrative momentum. The real asset is the attention of AI developers, and Thinking Machines Lab is using a crypto playbook to acquire it.
Takeaway: The Next Narrative to Watch
So where does this leave us? The Inkling story is a textbook example of how narrative precedes substance in high-frequency emerging markets. The next move will be telling. If Thinking Machines Lab releases a paper detailing Inkling's architecture, training data, and full benchmark suite within 30 days, my skepticism will soften. If they release the model weights under a truly open-source license (Apache 2.0 or MIT), the community can validate the claims. If instead they release a non-commercial license or a source-available version, the "open-source" label is another marketing tool. Based on the signals so far, I expect the latter. The crypto native which ran the article will likely follow up with a "part two" when the model is actually available, but by then the narrative will already have done its work. The fundamental question is not whether Inkling is good, but whether the market's attention is being misallocated by narratives that lack verification. In a bull market, euphoria masks technical flaws. The same principle applies to AI models as it does to blockchain projects. The best defense is to read with the eyes of a forensic accountant, questioning every metric and every source. Every chart is a story waiting to be corrected. And the story of Inkling is just beginning—but its ending will depend not on the MCP score, but on the honesty of its creators. As for the crypto news outlet that published this piece? They've made their bet. I'll be watching the liquidity mirrors, decoding the next narrative before the price reacts.
