#ai#finance#team

Four frontier labs: the edge lies in the system around the model

Four labs have frontier models. Why your own benchmark, an LLM gateway and short agent chains can deliver more than choosing a model.

  • What happened this week
  • Why this is a problem for companies
  • Your own benchmark matters more than someone else’s
  • Three tiers, three prices

What happened this week

In one October week, Google released Gemini

Argon into general availability, while Meta introduced Muse. Its flagship Wonder Full outperforms Fable- and Astra-class models on some benchmarks. Four companies now have frontier models, each offering three capability and price tiers. When one lab’s model can replace another’s, AI outcomes depend on the surrounding system: your quality checks, request routing and reliability over long workflows.

According to Latent Space, Gemini

Argon reached general availability, and independent benchmarks, including CursorBench, broadly support Google’s claims. About 20,000 people attended AI Engineer in New York that week. Anthropic explained its AI stack for financial firms, Vals AI’s CEO discussed independent model evaluation, and Coatue’s AI lead showed how the fund turns news into investment decisions. Meta Superintelligence Labs reached frontier capability in fourteen months.

Muse’s Spark, Glimmer and Wonder lineup follows the same pattern as OpenAI’s Sol, Terra and Luna and Anthropic’s Opus, Sonnet and Haiku. Latent Space analysts note that the models have different degrees of post-training, leaving Wonder’s class uncertain.

Why this is a problem for companies

  1. Many companies choose an LLM like an ERP system: once, for the long term.

  2. The team writes prompts for one model, embeds its API calls in code and trains users.

  3. A more capable, cheaper model appears a couple of weeks later.

  4. The team then faces two poor options: rewrite the integration, or keep an outdated model and overpay. Benchmarks are the second problem.

  5. A public benchmark measures tasks chosen by the lab. It says nothing about your company’s tasks.

  6. If industry analysts cannot classify Wonder, an executive cannot use someone else’s comparison table to decide whether it fits the company’s processes.

Your own benchmark matters more than someone else’s

Vals AI earns revenue from independent model evaluation

A company adopting AI needs an internal benchmark: several hundred real examples with verified answers.

We use this approach for catalogue tasks: product descriptions, taxonomy assignment and filter facets.

The test set consists of real SKUs with verified attributes.

We run each new model against that dataset and see the same day how many fields it filled correctly, where it failed and what a thousand product cards cost.

We decide whether to switch using those figures.

Three tiers, three prices

The three-tier lineups at all four labs suggest a practical approach: route requests by task.

A smaller model can classify incoming emails and extract document details for electronic document workflows.

Use a flagship model for reasoning tasks: checking a contract against policy or handling a non-standard claim.

We use LLM & Security Gateway as the single routing entry point between company systems and models.

The gateway selects a model for the task, logs requests and masks personal data.

Switching providers changes configuration rather than application code.

The same gateway also fronts GigaChat, YandexGPT and Qwen for data that must stay inside the required environment.

Assess where AI can deliver impact in your process

clients@kt.team Telegram @kt_team_it

Long horizons: a lesson from robotics

The week’s most useful business presentation concerned robots. Dyna Robotics trained Dyna-2 on a million hours of human video and described scaling laws spanning four orders of magnitude. Jason Ma at AIE focused on another problem: for a robot in the real world, operating autonomously and reliably for even one hour is difficult. Corporate AI agents face the same issue. An agent taking an order through 1C, electronic documents and the warehouse performs twenty to fifty steps.

At 98% reliability per step, a thirty-step chain succeeds without errors only about 55% of the time.

A stronger model that is correct 99% of the time per step gives about 74% overall.

Move deterministic operations into conventional integrations through MCP, Kafka or an ESB and shorten the chain to ten steps. Even the older model with 98% reliability per step then achieves about 82% overall.

In this calculation, engineering the process contributes more than upgrading the model.

Short chains, checkpoints and human review at risky steps make a simple-looking result possible through substantial engineering.

What an executive should do

Build a reference dataset before choosing a model, so you can evaluate a new release in a day.

Connect models through a gateway to avoid depending on a single provider.

Count the steps in every agent workflow and remove operations that ordinary code can handle.

Measure project success through TTU, or time to use: how many days pass before the tool starts moving a business metric.

The number of pilots launched says nothing about outcomes.

Verdict

Gemini 4 Argon and Muse Wonder benefit AI buyers through more choice and lower prices. Companies that can switch models in a day and measure results on their own data will gain. Others will read release announcements while their agents fail at step twenty.

Source

- Latent Space, AINews: AI Engineer NYC - Gemini 4 Argon, Vals AI Benchmarks - Latent Space, AINews: Meta Muse Wonder & Wonder Full - Latent Space, Dyna-2: General Purpose Physical Agents with Hour-Long Autonomy - Jason Ma, Dyna Robotics

Discuss the article: Four frontier labs: the edge lies…

Enter your email or phone number so we can get back to you.

Send via:

By clicking the button, you accept the offer and grant consent to the processing of personal data.

Related Cases

All cases

Related solutions

All solutions
AI Accountant in Your Environment: Implementation CostAI Agent Platforms: Purpose and Open OptionsAI Agent Sprint for LeadersAI agent development: one process in 1–2 weeks

Articles on the topic

AI in Business Collection

Related videos

All videos
Shorts What “KPI” did Amazon set for developers in the AI era? AITeams & hiringManagement
Shorts Can't Keep Up with AI? Build Communication and Accountability AI
Shorts AI does not empower the weak! AI

News on the topic

All news