Model catalogue
A living catalogue, refreshed as models ship and specifications move. The durable analysis lives in the decision guides: choosing a model and deploying models.
Catalogue
The heavy-reasoning tier: best goal coherence over long agentic runs and the hardest code. Price reserves it for the steps that genuinely need it, not the default node.
The production workhorse: reliable structured output and strong long-horizon coherence at a fraction of frontier cost. Our default for most agent nodes.
The fast, cheap router: sub-second latency at a fraction of the workhorse price. Ideal for classification, triage and the cheap leg of a routing strategy.
A gated frontier model with exceptional cybersecurity capability, released to vetted defenders only (Project Glasswing). Page-worthy: the case study for when a model earns its own page.
The main frontier alternative to Claude: very reliable structured output and tool calling, broad general capability. Azure gives EU-region hosting.
Native multimodal and the longest context window in the closed tier. Vertex AI gives EU regions for residency-sensitive workloads.
The open-weight default with the largest ecosystem and tooling. The license is not OSI open-source: it carries an active-users cap and EU conditions, read it before you build.
The strongest permissive open-weight line: Apache-2.0, excellent coding, very broad language coverage, and sizes spanning frontier down to small. The clean commercial choice.
Open-weight reasoning and agentic-coding leader under an MIT license, the cleanest terms for unrestricted commercial use. Strong on SWE-bench-style tasks.
The EU-native option: open models under Apache-2.0 plus hosted flagships, EU hosting with no third-country transfer. The default when sovereignty drives the choice.
A clean MIT license and a low active-parameter footprint make it one of the cheapest fully open self-hosted options at sustained scale.
A strong agentic and long-context open-weight contender that has topped neutral capability indices. Heavy to self-host, often consumed managed.
Google's small open-weight family, notably RAM-efficient. The go-to for on-device and multimodal mobile, where a few gigabytes is the whole budget.
Small models that punch above their size on reasoning, MIT-licensed. Quantized to a few gigabytes, they run a capable assistant on a laptop or phone, fully offline.