Know the landscape before you pick. Strengths and tradeoffs, kept current from primary lab sources — printed without vendor theater.
Each family has its own plate URL for sharing and ranking. Use Atlas to choose, then apply a named move from Techniques or practice in Studio. When labs ship new guidance, Pulse carries the primary-source brief.
Fig. 01 — anthropic
frontier
Strengths
Long-context writing, careful reasoning, structured XML prompts, strong instruction following.
Prompting
Prefer XML-delimited role, objective, constraints, and output_format. Multishot examples help.
Watch for
Can be conservative on creative leaps; quality and limits vary by Opus/Sonnet/Haiku tier.
Fig. 02 — google
frontier
Strengths
Multimodal workflows, very long context, Workspace adjacency.
Prompting
Use PTCF. Put document/image context before the ask when multimodal.
Watch for
Best when prompts follow clear Persona-Task-Context-Format framing.
Fig. 03 — openai
frontier
Strengths
Generalist speed, tools and agents, broad ecosystem, strong everyday default.
Prompting
Separate SYSTEM rules from USER task. Prefer explicit schemas for structured work.
Watch for
Instruction hierarchy matters; quality varies sharply by model snapshot and tier.
Fig. 04 — xai
frontier
Strengths
Real-time / X-adjacent context, sharp coding and agentic work, fast iteration on frontier Grok releases.
Prompting
Be direct. State role, goal, constraints, and desired format. Prefer short system rules and explicit tool boundaries.
Watch for
Tone can run hot; treat product claims and live-web answers as needing verification.
Fig. 05 — multiple (flux / midjourney / imagen / gpt image)
specialty
Strengths
Visual generation and edit workflows when subject, style, and lighting are specified.
Prompting
One attribute per line: subject, style, lighting, composition, quality, avoid.
Watch for
Text in images is unreliable; brand/logo requests often fail or infringe.
Fig. 06 — meta + ecosystem
open
Strengths
Local and private deploys, fine-tuning, cost control, full ownership.
Prompting
Be blunt. Show the output skeleton. Fewer implied norms than Claude/GPT.
Watch for
Needs explicit formats; less polish than frontier hosted models out of the box.