The state of the AI frontier, in one place

By raw input price, two models tie for lowest at $0.10 per 1M input tokens: GPT-6 Luna and Llama 4 Scout. But real workloads pay for output too: the cheapest to actually run, on a blended $/1M at a 3:1 input:output mix, is Llama 4 Scout at $0.15 per 1M. Open-weights models can also be self-hosted for the cost of your own hardware.

Eight models share the top spot, all at 1.05M tokens: GPT-6 Astra, GPT-6 Sol, GPT-6 Luna, GPT-5.5 Pro, GPT-5.5, GPT-5.6 Sol, GPT-5.6 Terra and GPT-5.6 Luna. The 1M class is wider still: 30 of the 40 models tracked have a ~1,000K-token window, so very-long-context work is no longer a single-vendor feature.

Not at the ~1M-token frontier any more — there the cheapest proprietary tier now undercuts the cheapest open one. At the ~1M-token frontier the cheapest proprietary model, GPT-6 Luna, runs about 1.2× cheaper than the cheapest open-weights one, GLM-5.3-Flash, on a blended $/1M at a 3:1 input:output mix ($0.20 vs $0.24 per 1M). Open weights still self-host for hardware cost alone, which no per-token rate can match, and the trade-off is that you run the infrastructure. The top proprietary models also still lead on some frontier reasoning tasks.

Every figure is hand-verified against published pricing and cross-checked for contradictions. Each row carries its own verification date and a link to the page the price is printed on (both shown in the full table below) — usually the provider's own pricing page, marked "official". A few models are open weights their maker gives away and sells no API for, so no first-party rate exists to cite; those rows show a named host's list price instead and are marked "hosted". The catalog was last updated 26 Sep 2026, and the last full cross-provider sweep was 3 Sep 2026; newer arrivals are verified on their own dates. It's a single, versioned source of truth (version 2026-09-26) that every BitByteCore tool reads, so a price can't drift between pages. Model pricing moves fast, so confirm with the provider before you rely on a figure.

Yes. The full catalog is published as machine-readable JSON at https://bitbytecore.com/data/frontier.json and as schema.org Dataset markup on this page. Quote or cite it with clear attribution and a link back.