The fastest way to pick an AI model is to map the task to the right archetype. The LLM cards spec does exactly that with five models: Claude Sonnet 4 (Artisan), Gemini 2.5 Pro (Sorcerer), o3 (Warrior), GLM 4.5 (Apprentice), and Kimi K2 (Apprentice). These five cover most enterprise and developer workflows if you match them to the job instead of arguing about benchmarks.
The Short Version
- Artisan — Claude Sonnet 4: Tool-using, agent-friendly, great at coding/writing/design. Higher cost. Not the best at heavy reasoning chains.
- Sorcerer — Gemini 2.5 Pro: Very long context, creative, smart. Reliability and steerability are weaker. Tool use is middling.
- Warrior — o3: Daily driver for research and tools. Reliable. Weaker at creative writing and front-end coding polish.
- Apprentice — GLM 4.5: Cheap, capable for coding/writing/tool tasks. Weak at heavy reasoning. No vision.
- Apprentice — Kimi K2: Extremely fast via Groq, very cheap, good for agents and coding. Hallucinates more, struggles with long context, no vision.
Why These Five Cover Most Work
Most teams need three things: dependable tools/agents for automation, long-context creativity, and a cost-speed option for scale. The cards map cleanly:
- Automation and tools: Claude Sonnet 4 or o3
- Long-context creative work: Gemini 2.5 Pro
- Cost-speed at scale: GLM 4.5 or Kimi K2
Claude Sonnet 4 — The Artisan
Claude Sonnet 4 is built for getting real work done: coding, writing, design, and actual agent workflows. It supports short, near-instant replies or extended reasoning when you need depth. It also plays well with developer stacks: strong tool-use, code execution, file operations, and integrations across major platforms, including Anthropic, Amazon Bedrock, and Vertex. Vision support is currently in public preview. If I need sub-agents coordinating code generation, retrieval, planning, and synthesis, this is my first stop.
- Strengths: Tool use, agent patterns, coding from plan to refactor, instruction-following, error recovery, and sub-agent orchestration.
- Limits: Not the top choice for heavy multi-step reasoning chains compared to the biggest Anthropic models; priced above budget models.
- Where it fits: End-to-end automation, developer copilots, agent workflows that need consistent action control and error handling.
Agentic context: cheap tokens vs. expensive tasks
Tool calls, file I/O, build pipelines, and human-in-the-loop checks are the real cost centers, not tokens. That’s why a model like Sonnet 4 wins when the throughput involves external tools. I break down this dynamic here: Cheap AI Tokens, Expensive Tasks.
Most cost sits in tools and retries, not in tokens. Pick models that handle actions reliably.
Gemini 2.5 Pro — The Sorcerer
Gemini 2.5 Pro’s edge is a very long context window and generative creativity. If I need to keep thousands of tokens of briefings, transcripts, or design notes in a single pass and riff on them, this is a solid pick. It’s also useful for speculative ideation that benefits from extended memory, like creative direction documents or design comps with narrative.
- Strengths: Long context, imaginative output, good at coding/writing/design when the project benefits from extended memory.
- Limits: Reliability and steerability can wobble. Tool-use skill trails models like Sonnet 4 or o3.
- Where it fits: Long-context creative writing, synthesis of large document sets, speculative prototyping.
o3 — The Warrior
o3 is the steady daily driver. It’s great at quick research, tool use, and getting reliable outputs in a hurry. If I care about stable instruction-following and fast iteration in practical workflows, o3 is an easy pick. It’s particularly good for orchestration: fetch, summarize, call a tool, check a result, move on.
- Strengths: Reliable, efficient, pragmatic. Strong for tool use and productivity pipelines.
- Limits: Creative writing and front-end coding polish aren’t its best domains.
- Where it fits: Research assistants, knowledge workers, and agents that need consistent behavior more than flair.
GLM 4.5 — The Apprentice
GLM 4.5 is the budget-friendly generalist that still handles the basics very well. If your workload is high-volume coding, writing, and routine tool tasks, this is a strong way to keep costs down without tanking quality. It’s not made for heavy reasoning chains. It also doesn’t do vision.
- Strengths: Cheap, practical, dependable for standard tasks.
- Limits: No vision. Not ideal for multi-hop reasoning or deep planning.
- Where it fits: Scaled content ops, code maintenance, simple agent workloads where cost per run matters.
Kimi K2 — The Other Apprentice
Kimi K2’s edge is raw speed via Groq and a low price. If I’m building agents that need to start responding almost instantly, it’s compelling. The tradeoffs are real: higher hallucination rates, poor long-context performance, and no vision. It’s fine for code scaffolding and quick-turn automations where factual accuracy is easy to check.
- Strengths: Extremely fast. Very cheap. Good for agents and coding stubs.
- Limits: Hallucinates more, struggles with long context, no vision.
- Where it fits: Latency-sensitive agents, prototyping at scale, workloads with automated verification.
Which Card For Which Job
Use cases mapped to the five models:
- Agentic automation with real tools: Claude Sonnet 4 or o3
- Long-context creative content or design briefs: Gemini 2.5 Pro
- High-volume, budget-sensitive coding or content tasks: GLM 4.5
- Latency-critical agents, cheap experimentation: Kimi K2
Decision widget: pick by the constraint
Practical Notes for Teams
- Don’t overpay for token quality when the work is tool-bound. You save more by reducing retries, shortening chains, and eliminating brittle steps than by shaving a few cents on tokens. See: Cheap AI Tokens, Expensive Tasks.
- Use long context only when it changes the result. If your prompt doesn’t actually need 100K tokens of memory, you’re paying for narrative comfort, not output quality.
- Put cheap models behind verification layers. For GLM 4.5 or Kimi K2, add automated checks: type checks, test suites, schema validation, or secondary fact checks.
- Split by role, not by vendor loyalty. It’s normal to run Artisan for planning/tooling and Sorcerer for long-context narrative in the same pipeline.
Where Vision Fits
Only a subset of these models support vision well right now. GLM 4.5 and Kimi K2 do not. Claude Sonnet 4 has vision support in public preview. If your workflow needs document parsing, UI inspection, or diagram reading, check current API support and consider mixing a vision-capable model just for that segment of the flow.
Example Playbooks
1) Software agent, real tools, CI checks
Model: Claude Sonnet 4 or o3. The agent plans tasks, writes code, runs tests, and opens a PR. If you want to pinch pennies, you can swap to GLM 4.5 for code generation steps, but keep Artisan/Warrior for planning and tool control to avoid costly retries.
2) Long-context creative and product narrative
Model: Gemini 2.5 Pro. Feed it specs, research, interview notes, and it can produce a unified narrative or creative direction. If you later need to ground it in tools or data, hand off to Artisan/Warrior.
3) High-volume doc updates and code maintenance
Model: GLM 4.5. Focus on batch throughput and add schema or test validation. Route edge cases back to Artisan/Warrior.
4) Real-time assistant for operations dashboards
Model: Kimi K2. Prioritize low latency. Keep a verification or escalation path for anything factual or high impact.
Price vs. Value
Price per token is only a proxy. For agent workflows, the real value is fewer failures and faster convergence. That’s why the Artisan and Warrior often win the ROI battle even though they cost more on paper. For scaled content or code maintenance, the Apprentices make sense because the work is easier to validate and the savings compound over volume.
Reality Check on Reliability
Gemini 2.5 Pro is great when context length wins. But if you find it drifting on instructions or persona, swap the controlling steps to o3 or Sonnet 4 and keep Gemini for the creative drafting stage. Kimi K2’s speed is a real advantage, but don’t ship it without guardrails. GLM 4.5 is steady for the price, but don’t force it into heavy reasoning. Right tool, right step.
Final Take
If you pick by task, not hype, these five models will cover almost everything you need:
- Artisan and Warrior for tools and automation where reliability dominates.
- Sorcerer for long-context creative work and extended memory.
- Apprentices when speed and cost are the priority and you add verification.
If you need a deeper dive on why agentic design changes the cost model, start here: Cheap AI Tokens, Expensive Tasks. Also relevant for context and model strategy: ChatGPT Agent Crushes Every Research Tool and Why OpenAI’s AI Rollouts Are Frustratingly Slow — And Why They Might Be Worth the Wait.