Three major events occurred in Silicon Valley last night.
Google released Gemini 3.8 Flash, Meta released Muse Spark 1.3, and Mostik, a previously little-known startup, was featured in an in-depth interview by the authoritative media outlet WIRED.
If you look only at the first two items, this resembles yet another ordinary night of AI topping the leaderboards.
Google recently propelled Gemini to the top of the DeepSWE rankings, only to be surpassed by Meta with an even higher score a few hours later, with the number one title lasting for just a few hours.
By 2026, people have almost become desensitized to this kind of thing.
Models are released, benchmarks are refreshed, celebrations last for several hours on X, and then the wait begins for the next company to continue refreshing them.
But if you look at these three things that happened last night together, I think the things that are truly worth paying attention to are not on those lists.
The economics of AI is undergoing a mutation.
Over the past few years, when we've discussed AI costs, the metric we've most often looked at is the price per million tokens.
However, with the emergence of Agent, this measurement method is becoming increasingly inadequate.
In the future, the truly important question may not be "how much do a million tokens cost," but rather how much it costs to fix a bug, complete a research project, or keep an agent running continuously for three hours.
And last night, the cost of inference was decreasing from several completely different directions at the same time.
Google is packing frontier-level capabilities into increasingly affordable models, Meta is enabling Agents to complete the same tasks with fewer tokens and tool calls, while Mostik is taking it a step further, questioning:
If both parties in a communication are AI, why do models still need to use human-designed languages to communicate with each other?
Gemini Is Back
stopping
First, look at Google.
Google has released its brand-new flagship model, Gemini 3.8 Flash, which the company describes as its most powerful reasoning and coding model to date.
This is also Google's third update to the Flash series in six weeks, from 3.6 to 3.7, and now to 3.8.
Flash has long been perceived as the simpler option: faster and cheaper, but with capabilities slightly below those of the strongest models.
Google is now gradually changing this definition.
This time, the 3.8 Flash upgrade focuses primarily on enhancements to long-horizon coding and agentic workflow.
It scored around 74% on DeepSWE v1.1, which measures AI's long-cycle software engineering capabilities.
The biggest difference between DeepSWE and traditional code benchmarks is that it doesn't just give the model an algorithmic problem, but actually puts the Agent into a code repository.
The model needs to understand the problem, search for code, modify files, invoke tools, run tests, and if it fails, continue to look for the cause.
A complete task can span dozens or even hundreds of steps.
In such a test, Gemini 3.8 Flash once ranked first in the world.
Claude Opus 5 is also around 74%, GPT-5.6 Sol is approximately 73%, while the previously leading Fable 5 is around 70%.
In terms of capabilities, the gap between these top models is no longer particularly large.
What's truly frightening is another figure: the cost.
The API pricing for Gemini 3.8 Flash remains unchanged, with input at $0.75/1M tokens and output at $3.75/1M tokens.
The average cost of completing a full task on DeepSWE is only $2.36.
In contrast, around 74% of Claude Opus 5 had a single-task average cost of $11.84; GPT-5.6 Sol was around 73%, with an average cost of $6.46.
In other words, for software engineering capabilities at the same level, the execution cost can now differ by several times.
This will become extremely important in the era of Agents.
In the era of chatbots, the difference of a few cents in the cost of a response is imperceptible to ordinary users.
However, Agent is completely different, a coding agent may run continuously for 100 steps, a research agent may search through dozens of web pages and read hundreds of pages of materials, and in the future, an Agent in a company may even work continuously for several hours.
Google officials even mentioned that when 3.8 Flash encounters complex tasks, it will proactively work harder, performing more reasoning and tool calls.
Google did not make the model "think less" to save money, but because tokens have become cheap enough that it can be allowed to run longer loops.
So I think what's truly important about 3.8 Flash is not that it has taken the world's number one spot again.
Instead, it is forcing the capabilities that only the most expensive frontier models could previously provide into the price range of Flash.
Please provide the Chinese headline to translate.
Meta Strikes Back After Three and a Half Hours
While everyone at Google was celebrating the release of Gemini 3.8 Flash, Meta made a sudden comeback.
Meta suddenly released the Muse Spark 1.3 model.
According to the evaluation released by Meta, Muse Spark 1.3 achieved a score of 75.4% on DeepSWE v1.1.
This figure surpasses Gemini 3.8 Flash, Claude Opus 5, and GPT-5.6 Sol.
More surprisingly, the Muse Spark 1.2 only scored around 55% on this test.
From version 1.2 to 1.3, a minor version update directly improved performance by about 20 percentage points.
But I think what's truly worth looking at for Meta this time isn't the 75.4%.
There are also two numbers: a reduction of about 20% in tool calls and a reduction of about 25% in token consumption.
This issue is easily overlooked, but it is actually very close to the most important problems Agent will face after it is truly commercialized.
One of the most wasteful aspects of an Agent is that it often doesn't think normally, but instead goes off on a tangent.
For instance, if a coding agent misinterprets the requirements at the 20th step, it may go on to modify five files, run three rounds of tests, and search through a large amount of code before discovering it has taken the wrong path, at which point it has to start all over again.
Dozens of tool calls and tens of thousands of tokens were wasted like this.
So if a model can discover the ambiguity of a task earlier, know it can't proceed earlier, and ask the user earlier, it's actually directly reducing costs.
The Muse Spark 1.3 upgrade has strengthened many capabilities in this area: it can maintain multiple workflows simultaneously in long threads, proactively ask for clarification on ambiguous tasks, request help when encountering unsolvable problems, and confirm before proceeding with irreversible operations, making it less likely to forget initial constraints even after running lengthy tasks for many rounds.
Meta has even begun to emphasize the model's awareness of its own capability boundaries: knowing what it can do and what it cannot do.
These capabilities may not have been as flashy in the Chatbot era as a 20-point benchmark increase, but in the Agent era, they can be directly converted into revenue.
One less mistake made by Agent is an optimization of reasoning in itself.
So when looking at Google and Meta together, the unit of measurement for large model competition is quietly changing.
In the future, people may care less about "how much a million tokens cost" and more about another number:
What is the actual cost of running a task from start to finish.
Mostik Achieves Groundbreaking Innovation
While Google and Meta are still researching how to accomplish tasks with cheaper and fewer tokens, Mostik has started tackling a more fundamental issue:
These tokens are necessary because they serve as a unique digital identifier, allowing for the verification of ownership and transfer of value on blockchain networks, and are essential for various applications such as decentralized finance and non-fungible tokens.
Mostik means "bridge" in Russian.
CEO Sasha Malysheva is the primary developer of the method, and its chief scientist is Stanislav Smirnov, a professor at the University of Geneva and the 2010 Fields Medal winner.
They are attempting something that sounds simple but is in fact extraordinarily difficult: getting two AI models to communicate with each other without using natural language.
Today, the vast majority of Multi-Agent systems work in the following way:
After Model A receives certain information, it first generates several hundred or even several thousand tokens, articulating its conclusions in natural language. Model B then reads all of that text, reinterprets it, builds its own internal representation, and continues its reasoning from there.
But the problem is that the things that actually perform calculations inside large models are not Chinese or English, but rather:
High-dimensional continuous mathematical representation.
This is equivalent to two computers that can clearly transmit data directly, but instead, computer A first prints the file into hundreds of pages, and then computer B uses a camera to perform OCR page by page to retrieve the data.
Everyone has been trying to figure out how to reduce printing costs, but Mostik wants to get rid of printers altogether.
They are trying to build a bridge between the internal representations of different models, allowing them to directly exchange latent representations instead of first generating natural language.
An experiment revealed by WIRED, a Silicon Valley authority, is particularly interesting.
Mostik has bridged the full version of GLM-5.2 753B and Qwen 3.5, which has only 4B and can run on mobile devices.
The final hybrid system had capabilities that fell between the two models, but its inference cost was only 1/20 of the full GLM-5.2.
Of course, it's too early to say that Mostik has already solved the problem of model communication.
Latent communication itself is not a new concept, as there have been numerous research attempts in recent years to exchange embeddings, hidden states, KV caches, and other internal representations.
The real challenge is that different models have different architectures, parameters, training data, and internal coordinate systems—making it inherently difficult for two models to truly understand each other's latent space.
Smirnov himself acknowledges that there is even a lack of mature mathematical language to describe the common representations between different models.
But the figure of 1/20 still excites me greatly.
It made me start thinking seriously about a completely different future AI architecture.
The Future Only Needs a Sub-Billion Model
Over the past two years of discussing edge AI, we've been exploring how to cram larger models into mobile phones: 7B, 4B, 3B, 1B, through continuous distillation, quantization, and compression.
But if Mostik's approach ultimately proves viable, I believe future smartphones may not need an all-purpose large model at all.
Your device may only need to run a small model of 0.xB or several B.
It is cheap, can run continuously, is responsible for understanding which app you are currently in, what you just did, and the status of the device, and handles the vast majority of simple and high-frequency tasks.
When genuinely encountering difficult problems, it calls upon remote large models through latent space communication.
But the key difference lies in that it doesn't need to resend the entire context of 100,000 tokens to the cloud like it does today, allowing the remote model to read from the beginning.
It may only need to transmit a highly compressed latent state.
After the remote large model completes complex reasoning, it doesn't necessarily need to generate thousands of tokens of natural-language explanation for the local model—it can simply transmit the new internal representation directly back.
This way, local small models are responsible for high-frequency, low-cost, and continuous operation, while cloud-based frontier models only handle low-frequency but truly difficult inferences, with some kind of Bridge enabling efficient communication in between.
If this can be achieved in the future, the reduction in reasoning costs may be far more than the 30% or 50% decrease in model API prices seen today.
It may change the way we organize AI computing itself.
So looking back at last night, on the surface these were three completely unrelated news items.
Google has released Gemini 3.8 Flash, bringing frontier-level capabilities into the price range of Flash.
Meta has released Muse Spark 1.3, enabling agents to accomplish more with fewer tokens and tool calls.
Mostik went further, attempting to prevent some tokens between models from being generated from the start.
They actually all point to the same change:
Intelligence is becoming cheaper at an extraordinary pace.
Moreover, this price reduction is no longer just about a decrease in API unit prices.
Model prices are decreasing, the amount of computation required to complete tasks is decreasing, and even the way models exchange information with each other is being redesigned.
The eventual impact of this could be far greater than a few percentage points on the benchmark.
Many technologies truly take off not when they "become available for the first time," but when they "finally become cheap enough to be used casually."
Having an agent spend tens of dollars to complete a trivial task for an ordinary person may not be cost-effective at all.
If in the future it only costs a few jiao, many applications that nobody would consider now would suddenly become viable.
Today, it's impossible for us to have dozens of Agent running around a person 24 hours a day, but if the cost of inference drops by one or two orders of magnitude, it may become the default state.
It's already 7:00 a.m., and I should have gone to sleep several hours ago.
After Gemini was released, Meta caught up within a few hours, and then I came across Mostik's 1/20 experiment.
After lying down in bed, my mind kept thinking about these things, thinking about the 0.xB model on my phone, thinking about the large models in the cloud, and thinking that they may not need to communicate with each other in natural language at all.
The more I thought about it, the more I couldn't sleep, so I finally got up and finished writing the article.
74% of Gemini and 75.4% of Muse may be overtaken by new numbers in a few days.
But right now, I really can't contain my excitement.
Because I'm becoming increasingly concerned with another issue, rather than who has achieved world number one again:
How cheap can such intelligent AI ultimately become?
Once powerful AI becomes cheap enough to be casually invoked, I think many AI products that seem crazy today may have only just begun.