Four axes that decide
Every LLM decision comes down to four axes: capability, latency, cost, and compliance. Rank them for your workflow before shortlisting models. Most bad choices happen when teams optimize the wrong axis.
This guide is opinionated. The right model in 2026 depends on the shape of the workload, not on which vendor's demo you watched last week.
Chat and drafting
For user-facing chat that emphasises prose quality, Claude Sonnet 4 is the default recommendation. It follows instructions cleanly, writes well, and its refusals are more appealable than the alternatives.
GPT-4o is the close second and wins for teams that value function-calling reliability and the Realtime API for voice.
Gemini 2.5 Pro competes at the top end for teams already on Google Cloud or working with multimodal inputs.
Tool use and agents
For agent workflows with heavy tool orchestration, GPT-4o with strict structured outputs remains the safest choice — schema compliance is effectively guaranteed.
Claude Sonnet is a strong alternative if brief adherence matters more than schema strictness. Its tool use is slightly less strict but stays more consistently on-task across long agent runs.
Avoid open-weights models as the top-level agent controller for now. Use them for sub-tasks; keep the orchestration layer on frontier models.
Long-document analysis
Gemini 2.5 Pro wins any workflow that pushes past 200K tokens. Its recall accuracy at 1M+ tokens is the current state of the art.
Claude Sonnet is the choice at 100-200K tokens where you also want nuanced writing. Its long-context recall is second only to Gemini and its prose quality is higher.
GPT-4o's 128K context is fine for most 'long' workflows but shows more recall degradation than either alternative at the top end.
Cost-first pipelines
Classification, extraction, routing: Gemini Flash-Lite or GPT-4o-mini. Both cost pennies per million tokens.
Reasoning-heavy batch jobs: DeepSeek R1 via API or self-hosted, at ~10x lower price than closed frontier reasoning models. Latency is higher; batch tolerates it.
Rule of thumb: route 80% of traffic to the cheap tier, escalate the 20% that fails a self-critique step to a frontier model. Real deployments save 60-80% at similar overall quality.