Last Night, Silicon Valley Witnessed a Battle of AI Gods

Author: Max, 01 Founder

Editor: Max

Three major events happened in Silicon Valley last night.

Google released Gemini 3.8 Flash, Meta released Muse Spark 1.3, and a startup no one had discussed before, Mostik, was featured in an exclusive interview with the authoritative media outlet WIRED.

If you only look at the first two, it seems like just another ordinary night of AI benchmark-chasing.

Google had just pushed Gemini to the top of the DeepSWE leaderboard, and a few hours later, Meta announced an even higher score. The title of world number one lasted only a few hours.

By 2026, people have almost become numb to this kind of thing.

Model releases, benchmark refreshes, a few hours of celebration on X, and then waiting for the next company to continue refreshing.

But if you connect these three events from last night, I think what truly deserves attention is not on those leaderboards at all.

The economics of AI is undergoing a sudden mutation.

Over the past few years, when discussing AI costs, the most commonly cited figure has been the price per million tokens.

But since the emergence of agents, this measurement is becoming increasingly inadequate.

The truly important question in the future may not be "how much does a million tokens cost," but rather how much it costs to fix a bug, complete a research project, or have an agent work continuously for three hours.

And last night, inference costs were dropping simultaneously from several completely different directions.

Google is cramming frontier-level capabilities into increasingly cheaper models, Meta is enabling agents to complete the same task with fewer tokens and tool calls, and Mostik goes even further, questioning:

If both parties in communication are AI, why do models still need to communicate in a language designed for humans?

PART.01, Gemini Strikes Back

First, let's look at Google.

Google released their new flagship model, Gemini 3.8 Flash, which they officially call their 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.

The impression of Flash in the past has always been very simple: faster, cheaper, but with slightly weaker capabilities than the strongest models.

Google is now gradually changing this definition.

The core improvements in this 3.8 Flash release are focused on long-horizon coding and agentic workflows.

On DeepSWE v1.1, which measures AI's long-cycle software engineering capabilities, it achieved a score of approximately 74%.

The biggest difference between DeepSWE and traditional coding benchmarks is that it doesn't just throw an algorithm problem at the model; it actually places the agent into a code repository.

The model needs to understand the problem, search the code, modify files, call tools, run tests, and continue searching for causes after failures.

A complete task may last dozens or even hundreds of steps.

In such tests, Gemini 3.8 Flash briefly reached world number one.

Claude Opus 5 is also around 74%, GPT-5.6 Sol around 73%, and the previously leading Fable 5 around 70%.

If we only look at capabilities, there is no particularly huge gap between these top models.

The truly terrifying figure is another one: money.

The API launch price for Gemini 3.8 Flash remains at $0.75 per 1M input tokens and $3.75 per 1M output tokens.

Completing a full task on DeepSWE costs an average of only $2.36.

In comparison, Claude Opus 5, also around 74%, has an average cost per task of $11.84; GPT-5.6 Sol, around 73%, averages $6.46.

In other words, for the same level of software engineering capability, execution costs can differ by several times.

This will become very important in the agent era.

In the chatbot era, whether a single response costs a few cents more or less is basically imperceptible to ordinary users.

But agents are completely different. A coding agent may execute 100 consecutive steps, a research agent may search dozens of web pages and read hundreds of pages of material, and in the future, agents in enterprises may even work continuously for hours.

Google officials even mentioned that when 3.8 Flash encounters complex tasks, it will actively work harder, performing more reasoning and tool calls.

Google did not make the model "think less" to save money; instead, because tokens are already cheap enough, they can allow it to run longer loops.

So I think the truly important thing about 3.8 Flash is not that it took world number one again.

Rather, it is forcefully squeezing capabilities that were previously only available in the most expensive frontier models into the Flash price range.

PART.02, Three and a Half Hours Later, Meta Strikes Back

Just as the Google team was celebrating the release of Gemini 3.8 Flash, Meta suddenly struck back.

Meta suddenly released the Muse Spark 1.3 model.

According to Meta's published evaluations, Muse Spark 1.3 achieved a score of 75.4% on DeepSWE v1.1.

This number surpasses Gemini 3.8 Flash, Claude Opus 5, and GPT-5.6 Sol.

Even more astonishing is that Muse Spark 1.2 scored only about 55% on this test.

From 1.2 to 1.3, a minor version update directly improved by about 20 percentage points.

But I think what's truly worth looking at with Meta this time is not the 75.4%.

There are two other numbers: tool calls reduced by about 20%, token consumption reduced by about 25%.

This is easy to overlook, but it is actually very close to the most important issue once agents are truly commercialized.

The most wasteful part of an agent's spending is often not normal thinking, but going off track.

For example, if a coding agent misunderstands the requirements at step 20, it may continue modifying five files, run three rounds of tests, search through a large amount of code, and only then realize the approach was wrong, then start all over again.

Dozens of tool calls and tens of thousands of tokens are wasted just like that.

So if a model can detect ambiguity in a task earlier, realize sooner that it cannot proceed, and ask the user earlier, it is directly reducing costs.

Many of the capabilities strengthened in Muse Spark 1.3 fall into this category: it can maintain multiple workflows simultaneously in long threads, proactively ask when encountering ambiguous tasks, proactively request help when it cannot solve a problem, confirm before irreversible operations, and is less likely to forget initial constraints after many rounds of long task execution.

Meta is even starting to emphasize the model's awareness of its own capability boundaries: knowing what it can do, and knowing what it cannot do.

These capabilities may not be as sexy as a 20-point benchmark increase in the chatbot era, but in the agent era, they can all be directly converted into money.

An agent making one fewer mistake is itself a form of inference optimization.

So when you put Google and Meta together, the unit of measurement in large model competition is quietly changing.

In the future, people may care less and less about "how much does a million tokens cost," and more and more about another number:

How much does it cost to run a real task from start to finish.

PART.03, Mostik, a Disruptive Breakthrough

If Google and Meta are still studying how to complete tasks with cheaper and fewer tokens, then Mostik is starting to touch on a more fundamental question:

Why do these tokens need to exist at all?

Mostik means "bridge" in Russian.

CEO Sasha Malysheva is the main developer of this approach, and its chief scientist is Stanislav Smirnov, a professor at the University of Geneva and the 2010 Fields Medal winner.

They are attempting to do something that sounds simple but is actually extremely difficult: enabling two AI models to communicate without using natural language.

Today, the vast majority of multi-agent systems work like this:

Model A receives some information, first generates hundreds or even thousands of tokens, expressing its conclusions in natural language; Model B then reads all that text, reinterprets it, builds its own internal representation, and continues reasoning.

But the problem is that what large models actually compute internally is not Chinese or English, but:

high-dimensional continuous mathematical representations.

This is equivalent to two computers that could directly transfer data, but Computer A first prints the file into hundreds of pages of paper, and then Computer B uses a camera to OCR each page back.

Everyone has been studying how to reduce printing costs, but Mostik wants to simply throw away the printer.

They are trying to build a bridge between the internal representations of different models, allowing models to directly exchange latent representations instead of first generating natural language.

An experiment disclosed by the authoritative Silicon Valley media WIRED is very interesting.

Mostik bridged the full version of GLM-5.2 753B with Qwen 3.5, which has only 4B parameters and can run on mobile devices.

The resulting hybrid system has capabilities between the two models, but inference costs are only 1/20 of the full GLM-5.2.

Of course, it is definitely too early to say that Mostik has solved model communication.

Latent communication itself did not first appear yesterday; over the past few years, there have been many studies attempting to exchange internal representations such as embeddings, hidden states, and KV caches.

The real difficulty is that different models have different architectures, parameters, training data, and internal coordinate systems, so how to make two models truly understand each other's latent space is itself a very difficult problem.

Smirnov himself admits that there is currently a lack of mature mathematical language to describe the common representations between different models.

But the 1/20 figure still excites me greatly.

Because it makes me seriously consider a completely different future AI architecture.

PART.04, In the Future, You Only Need a 0.xB Model

Over the past two years, when discussing on-device AI, we have been studying how to cram larger models into phones: 7B, 4B, 3B, 1B, constantly distilling, quantizing, and compressing.

But if Mostik's approach can truly work out, I think future phones may not need a "know-it-all" large model at all.

Your device may only need to run a small model of 0.xB or a few B parameters.

It is cheap, can run continuously, is responsible for understanding which app you are currently in, what you just did, and the device's status, and handles the vast majority of simple, high-frequency tasks.

When truly difficult problems arise, it can call a remote large model through latent space communication.

But the key difference is that it does not need to resend a 100,000-token context to the cloud for the remote model to read from scratch, as is done today.

It may only need to transmit a highly compressed latent state.

After the remote large model completes complex reasoning, it may not need to generate thousands of tokens of natural language explanation for the local model; instead, it can directly send back the new internal representation.

In this way, the local small model handles high-frequency, cheap, continuous operation, while the cloud frontier model only handles low-frequency but truly difficult reasoning, with some kind of bridge for efficient communication in between.

If this can truly be achieved in the future, the reduction in inference costs may be far more than just a 30% or 50% price cut in model APIs today.

It may change the way we organize AI computation itself.

PART.05, Conclusion

So looking back at last night, on the surface there were three completely different pieces of news.

Google released Gemini 3.8 Flash, squeezing frontier-level capabilities into the Flash price range;

Meta released Muse Spark 1.3, enabling agents to accomplish more with fewer tokens and tool calls;

Mostik went even further, attempting to prevent some tokens between models from being generated in the first place.

They all point to the same change:

Intelligence is becoming cheaper at an astonishing rate.

And this price reduction is no longer just about API unit prices.

Model prices are falling, the amount of computation required to complete tasks is falling, and now even the way models exchange information is being redesigned.

The ultimate impact of this may be far greater than a few percentage points increase on benchmarks.

Because many technologies truly explode not when they "first become usable," but when they "finally become cheap enough to use casually."

Today, having an agent spend dozens of dollars to complete a small task for an ordinary person may be completely uneconomical.

If in the future it only costs a few cents, many applications that no one would consider today suddenly become viable.

Today we cannot have dozens of agents running continuously around one person 24 hours a day, but if inference costs drop by another one or two orders of magnitude, it may become the default state.

It is now 7 a.m., and I should have gone to sleep hours ago.

But after Gemini was released, Meta caught up a few hours later, and then I saw Mostik's 1/20 experiment.

After lying in bed, my mind kept thinking about these things: the 0.xB model in the phone, the large model in the cloud, and how they might not need to talk to each other in natural language at all.

The more I thought, the more I couldn't sleep, and finally I got up and finished writing this article.

Gemini's 74% and Muse's 75.4% may be overtaken by new numbers in a few days.

But I truly find it hard to calm this excitement right now.

Because more than who took world number one again, I increasingly care about another question:

How cheap can such intelligent AI ultimately become?

Once powerful intelligence truly becomes cheap enough to be summoned at will, I think many AI products that seem crazy today may just be getting started.