AI MODELS

Emerging

Who Controls the Frontier Model?

Who builds the intelligence, who can use it, and on what terms.

Strategic Brief

September 11 – October 11, 2026 · 30-day retrospective · Evidence checked through October 11, 2026

The bigger trend

The leading AI labs cut prices and shipped new models in the same week, while open-weight models keep gaining real work.

On September 22, OpenAI released lower-cost GPT-6 models at half the API price of the versions they replace, and Anthropic launched Claude Opus 5.5 at a 20% lower list price. Google made its real-time Gemini 3.8 Live model generally available, and Mistral announced a new open-weight model. Meanwhile, open-weight models are carrying a growing share of real traffic. One month of launches does not pick a winner; the question is whether open models take important work from closed ones.

Outside → In

What changed — Sense (verified)

September 22: OpenAI released GPT-6 Sol and GPT-6 Luna, lower-cost members of its GPT-6 family, with API prices 50% below the GPT-5.6 versions they replace. Its flagship, GPT-6 Astra, launched September 3.

September 22: Anthropic launched Claude Opus 5.5 with list prices 20% below Opus 5; Anthropic says it costs 40% less to run on typical workloads (company testing).

September 24: Google made Gemini 3.8 Live, its real-time voice and video model, generally available for businesses.

October 6: Mistral announced Mistral Large 4 as an open-weight model, available now as a preview, and said its weights will be released by the end of October.

What it means — Interpret (RAW PEC view)

Closed-model providers compete on capability, convenience and the services around their models; open-weight providers offer control and potentially lower cost. The leaders cut prices this month, but the evidence does not show what drove the cuts. OpenAI's decision to end Cursor's access after SpaceX bought it shows the risk of depending on a single provider's model.

What to consider — Position

Which of your AI tasks truly need the most capable closed model, and which could run on an open model you control?

Inside → Out — suggested enterprise actions

Suggested actions for readers, not a report of actions taken.

Who needs to align — Align

Technology leadership, AI architecture, security, procurement, finance and the business owners of AI use cases.

What the enterprise could execute — Execute

Run two or three representative tasks on a closed model and an open-weight model; compare quality, total running cost, speed, security and the effort to switch.

How to measure and learn — Learn

Task quality, cost per completed task, latency, security incidents and switching effort.

Supporting competitive moves (this period)

  • Mistral · Build ·

    Mistral Large 4 announced as open-weight; weights promised by end of October

    Primary source · mistral.ai

  • Google · Build ·

    Gemini 3.8 Live generally available

    Primary source · cloud.google.com

  • OpenAI · Build ·

    GPT-6 Sol and Luna at 50% below the API price of the GPT-5.6 versions

    Primary source · openai.com

  • Anthropic · Build ·

    Claude Opus 5.5 at 20% lower list price

    Primary source · www.anthropic.com

  • DeepSeek · Build ·

    V4.1 Flash released under the MIT license

    Primary source · huggingface.co

Related: AI Computing (Who Controls AI Compute?) · AI Coding (Who Controls the Coding Workflow?) · AI Economics (How AI Cost, Value and Pricing Connect)

The management question

If an open-weight model can do your important AI tasks well enough, what justifies staying dependent on a closed-model provider, and how easily could you switch?

Stack layers

Contested layers: 4 Models and intelligence

Adjacent layers: 2 Compute, cloud and security; 6 Workflow, agents and orchestration

Contest dimensions

  • Capability
  • Access
  • Price
  • Control

Participants

Closed-model leaders

  • OpenAI (Incumbent): released lower-cost GPT-6 Sol and Luna at 50% below the API price of the GPT-5.6 versions they replace (September 22, 2026); flagship GPT-6 Astra launched September 3
  • Google (Incumbent): Gemini 3.8 Live, its real-time voice and video model, generally available for businesses (September 24, 2026)
  • Anthropic (Incumbent): launched Claude Opus 5.5 with list prices 20% below Opus 5 (September 22, 2026)
  • SpaceX (xAI) (Challenger): owns xAI (February 2026), maker of the Grok models, and Cursor (August 2026)

Open-weight challengers

  • DeepSeek (Challenger): doubled its share of OpenRouter traffic from 9% to 18% in the first half of 2026 and released V4.1 Flash under the MIT license (September 10, 2026)
  • Mistral (Challenger): announced the open-weight Mistral Large 4, with weights promised by the end of October (October 6, 2026)
  • Z.ai (Zhipu) (Challenger): open-sourced GLM-5.2 (June 17, 2026); listed in Hong Kong, with market value reaching HK$1 trillion
  • Moonshot AI (Challenger): its open Kimi K2.5 model is the base of Cursor's Composer 2; valued at about $20 billion

Distribution

  • OpenRouter (Challenger): routes developers across hundreds of models; valued at about $1.3B after a $113M Series B led by CapitalG (May 2026)

Ecosystem Moves

How the players are building, buying and partnering to compete in this battleground. Newest first.

6 Build · 0 Buy · 0 Partner · 1 Compete

Independent and unsponsored: no company pays to appear in these battlegrounds.

Sources
Talk with Ron