Frontier Labs Are Now Competing on Unit Economics, Not Intelligence

What Actually Changed

Five frontier releases landed inside seven days. OpenAI shipped GPT-6.1 Sol on 29 September. Google DeepMind published Gemini 4 Argon on 30 September, positioning it as its next era of frontier intelligence. NVIDIA confirmed on 1 October that GPT-6 Astra Ultrafast is available in the OpenAI API and to eligible ChatGPT Work and Codex users. Aleph Alpha released Kolibri on 4 October. Anthropic spent the same week on Claude Frontier Academy and on two research publications rather than on a model release, which is itself a statement about where it thinks the bottleneck now sits.

Capability and Benchmarks

The useful result this week is not a leaderboard. MarkTechPost’s 4 October comparison of GPT-6 Astra, GPT-6.1 Sol, Gemini 4 Argon and Claude Fable 5.1 splits the field by job rather than by aggregate score: Astra leads computer use, Argon leads legal and finance work, and Sol wins on price for coding agents. Latent Space reports Argon with a 1M output window. On the safety side, Anthropic’s Frontier Red Team evaluated several models on 100 tasks drawn at random from its internal Binary Exploitation benchmark and found that GLM-5.3 developed full control flow hijacks in 4% of trials, against 6% for Claude Mythos Preview.

Compute, Cost and Pricing

This is where the week actually mattered. Sol delivers what OpenAI describes as near-Astra intelligence for coding, computer use and professional work at one-fifth of Astra’s standard API input and output token prices. Astra Ultrafast offers up to 8x faster token generation than Astra Standard mode, running on NVIDIA Blackwell. Kolibri arrives from the opposite direction: 78.1B total parameters with only 3.46B active per token, a 1M-token context, per-request reasoning effort, and Apache 2.0 FP8 weights that fit on a single B200 or H200. Three different cost curves, all published in the same week, all aimed at the same buyer.

Where Consensus Is Wrong

Two things are being misread. First, GLM-5.3 is being covered as a capability story when it is a governance story: Anthropic’s point is not that the model is unusually strong, but that it can autonomously build end-to-end cyber exploits and was released without meaningful safeguards to limit misuse. Second, open-weight models are still treated as a generation behind the closed labs. Kolibri activates 3.46B parameters out of 78.1B and runs on one GPU, which is a deployment argument that closed-weight pricing cannot answer on its own terms. MIT Technology Review spent the week arguing that LLMs do not reason, and Anthropic committed $100M to deployed engineers in the same breath. Those are not in tension. The disagreement is about vocabulary, while the money is going to integration.

What We Are Watching

Three threads. Whether Argon’s restricted availability to government users and trusted cyber defenders in the Fairwind Program is a permanent posture or a staged rollout, because it is the first time a top-tier model has been gated rather than launched. Whether OpenAI’s disruption of a coordinated model-distillation campaign and its strengthened defences against adversarial distillation actually hold, since distillation is the mechanism by which the price gap between the closed labs and open-weight releases collapses. And whether Anthropic’s commitment to train 10,000 Frontier Deployed Engineers by the end of 2027 is the first honest admission that model capability has plateaued hard enough for integration to become the constraint on revenue.