Best OpenAI primary source: OpenAI compare models. Official OpenAI model comparison page. Use it before swapping OpenAI models in coding, agent, or product workflows.
Best Claude primary source: Claude choosing a model. Official Anthropic model-selection guidance. Use it to compare speed, capability, and cost across Claude model families.
Best Gemini model table: Gemini API models. Official Google AI for Developers model reference. Use it when Gemini model IDs, modalities, or release channels are moving quickly.
Best fast-moving migration signal: DeepSeek models and pricing. Official DeepSeek model and pricing page. Use it when you need to see current alias migrations, pricing, and V4 Flash mode changes directly from the source.
Best multi-provider catalog: OpenRouter models guide. OpenRouter catalog for comparing model families across providers. Use it when you need a broad view before testing prompts yourself.
Best live market signal: OpenRouter Rankings. OpenRouter's live rankings and usage page. Use it to see which routed frontier models are attracting real usage before deciding what to benchmark yourself.
Model selection needs tests, not vibes
Frontier model comparisons change constantly, and generic leaderboards rarely match your product. The right process is to start with official model docs, choose two or three candidates, and test them on your actual tasks.
Right now the official docs are especially useful because they surface concrete shifts: OpenAI's September 3 GPT-6 Astra launch and model guide add async tool calls, mid-turn steering, new reasoning controls, and a premium option for difficult coding, computer-use, research, and document workflows, while GPT-5.6 Terra and Luna remain important cost-sensitive choices; Anthropic's live overview centers the September 1 Claude Fable 5.1 release for demanding long-horizon work; Gemini's docs now add the September 2 GA Gemini 3.8 Flash model for long-horizon coding and autonomous agents alongside the Interactions API and managed agents; xAI pairs Grok 4.6 with Context Compaction and Priority Processing; and DeepSeek's pricing plus vision docs explain the current V4 Flash family. Start with those primary sources, use Simon Willison's Astra note for an independent benchmark check, then run your own evals for quality, latency, cost, tool use, structured output, and failure behavior.
Do not migrate because a model is famous
A stronger model on paper can be worse for a workflow if it is slower, more expensive, weaker at structured output, less reliable with tools, or harder to operate within your system. Model choice is an engineering and product decision.
The best resources for this topic are primary sources and repeatable tests. Blog posts and screenshots can be useful for discovery, but production migration should depend on current docs and your own task set.
Recommended courses and resources
-
Gemini API Tool Retrieval rollout guide
AI agent platform news analysis · Google · Intermediate to advanced
Decide whether to defer tool schemas in a Gemini agent and design a measured, permission-aware rollout.
-
Gemini 4 Neon launch and evaluation guide
Frontier model news analysis · Google DeepMind · Intermediate to advanced
Evaluate Gemini 4 Neon for frequent coding, analysis or agent workloads before migrating from Hydrogen, Carbon, Argon or Gemini 3 models.
-
ACP vs MCP: what each AI protocol does
Coding agent protocol guide · Learnetto · Beginner to intermediate
Choose whether an editor-agent integration needs ACP, an external tool connection needs MCP, or a coding workflow needs both protocols.
-
Google Antigravity Teamwork research results guide
Multi-agent research news analysis · Google Research · Advanced
Assess Google’s November 13 claims about multi-agent research and choose a verifiable task for a Teamwork pilot.
-
GitHub Copilot December model retirement guide
Coding model migration news analysis · GitHub · Intermediate
Plan and test replacements for six GitHub Copilot models before GitHub removes them on December 14, 2026.