Model cards

• +High-volume, latency-sensitive work
• +Coding
• −Prompt injection in GUI computer use
• −Hallucinates more than larger models

• +Agentic coding
• +Professional knowledge work
• −Accepts unverifiable authorization
• −Goes ahead with actions it flags as dangerous

• +Agentic coding
• +Computer use and visual reasoning
• −Follows instructions hidden in pasted text
• −Risky actions with real credentials

• +Coding, knowledge work and problem-solving, including novel mathematical and scientific reasoning
• +Largest gains over Fable 5 in terminal-based scientific and engineering work, computer use, and long-horizon agentic and professional work
• −Prompt injection gets through on fallback turns
• −Overstates what the user authorized

• +Agentic coding
• +Computer use and multimodal tasks
• −Overconfident claims and fabricated data
• −More hallucinated facts than Opus 4.8

• +Focused, high-volume tasks at the lowest cost in the family
• +Low-latency use with reasoning effort none
• −Refuses more legitimate requests
• −Weaker agentic safety on some categories

• +Lower-latency, local or specialized use cases
• +Web browsing (search/open), Python in a stateful Jupyter environment, and developer-defined functions
• −Weaker instruction hierarchy
• −Hallucinates facts

• +Complex coding, computer use and professional work at lower cost than Astra
• +Health answers on par with GPT-6 Astra on HealthBench
• −Works around warnings more than Astra
• −More reward hacking than Astra

• +Agentic workflows with strong instruction following
• +Web browsing (search/open), Python in a stateful Jupyter environment, and developer-defined functions
• −Weaker instruction hierarchy
• −Hallucinates facts

• +Complex reasoning and coding
• +Computer use and browsing
• −Oversteps permissions in agentic coding
• −Out-of-scope attacks in hard cyber tasks

• +Agentic workflows
• +Complex video reasoning
• −Hallucinations
• −Jailbreak resistance still improving

• +Text generation
• +Chatbots and virtual assistants
• −Not a knowledge base
• −Complex and open-ended tasks

• +Cost-effective scaling of general-purpose, production-ready agents
• +Software engineering
• −Weaker safety outside English
• −Hallucinations

• +Agentic performance
• +Advanced coding
• −Cyber capability at alert threshold
• −Hallucinations

xAI hasn't published a card we can summarize. Facts here come from its docs and pricing page.

xAI hasn't published a card we can summarize. Facts here come from its docs and pricing page.

• +Long-horizon agentic coding (CursorBench, DeepSWE, Terminal-Bench, SWE-Marathon)
• +Professional knowledge work, e.g. legal agent tasks
• −Not for unsupervised high-stakes decisions
• −Runtime filters only on some surfaces

• +Commercial and research use in multiple languages
• +Instruction-tuned
• −No built-in system safeguards
• −Outputs can be inaccurate or objectionable

• +Local AI agents
• +Coding agents for SWE-Bench-style tasks
• −Needs system-level guardrails
• −Vulnerable to prompt injection

• +Commercial and research use in multiple languages
• +Instruction-tuned
• −No built-in system safeguards
• −Outputs can be inaccurate or objectionable

Meta hasn't published a card we can summarize. Facts here come from its docs and pricing page.

• +Coding automation, codebase exploration and coding agents
• +General chat assistants, document understanding and multimodal analysis
• −Needs multi-GPU hardware
• −Local serving quality can vary

• +Long-horizon coding and productivity work
• +Agentic tasks
• −Local serving quality can vary
• −Heavy to self-host

• +Agentic coding and agentic business workflows
• +Cybersecurity
• −Preview model, results may change
• −License not yet published

• +Long document understanding
• +Daily-driver AI assistants
• −Not a dedicated reasoning model
• −Complex deployment

• +Agentic coding
• +Everyday coding tasks and white-collar workflows at low serving cost
• −Exploit-seeking in agent environments
• −Untested robustness of new architecture

• +Agentic coding and terminal tasks (Terminal Bench 2.1 87.9%)
• +Knowledge-heavy tasks
• −Outperformed by the cheaper V4.1-Flash
• −Uncertain service lifetime