Models

Top intelligence score here at 59.4, but restricted to vetted cyber-defence organisations

The most capable closed model you can actually use — first on both long-horizon coding and professional reasoning

OpenAI's flagship: second on science reasoning at GPQA 95.8, and in the same top price tier as Fable 5.1

96% of Astra's intelligence at one fifth of the price — the sweet spot in OpenAI's lineup

Half the price of Opus and only 10% cheaper per task — the best-value general-purpose flagship

A flagship-grade option at $0.96 per index task, strongest on visual reasoning and computer use

First overall on GPQA at 96.4 with a 4% hallucination rate — the strongest Gemini you can actually use

A long-form creative model, top three on both EQ-Bench Creative Writing and LiveBench Language

2.8 trillion parameters and first on Terminal-Bench 4.0, but output runs at just 28 tok/s

The largest open-weight model, with GPQA 93.1 closing in on the closed leaders

Meta's closed flagship, the fastest in its tier at 143 tok/s

A 321B/18B MIT-licensed light flagship with a LiveBench design score above Gemini 4 Argon

The fastest lightweight model here (235 tok/s), close to Opus on LiveBench design tasks

Tuned for creativity: a higher LiveBench design score than Gemini 4 Argon, but only 500K of context

The strongest open model you can run on a Mac: LiveBench design 75.9 beats Gemini 4 Argon

The value champion: $0.26 per index task, with a LiveBench design score of 83.4 that matches Opus

First among open models on Terminal-Bench 4.0 (38.9%) — the pick for long-horizon agent workflows

First among open models at $0.24 per index task, MIT licensed with no revenue threshold

Efficient on light hardware with a 1M open-weight context, but essentially no agent ability

A million-token-context Flash at $0.15/$0.60, $0.42 per task

The 1.6-trillion-parameter former flagship — the vendor is already migrating users to V4.1 Flash

The most permissively licensed local model; 75.8% on science reasoning shows the ceiling for small models

OpenAI's budget tier, with a LiveBench design score of 82.5 that beats Astra

GPQA 95.4, level with GPT-6.1 Sol, at 1/2.7 of the price

The strongest Chinese closed model, first on APEX-Agents at 63.3%

A single-GPU open baseline — 4.4% on APEX-Agents means simple tasks only

LiveBench design 76.9 — a capable document generator among Chinese multimodal models

NVIDIA's 550B/55B open flagship with 1M context, tied into the NIM ecosystem