Model and source freshness¶
Last verified: 2026-05-06 · Drift risk: high (this page exists because models drift fast)
Specific model names, context-window sizes, rate limits, and prices change every few weeks across all major vendors. This page is the single place where the guide is honest about that.
Reading rule¶
When you see any of the following anywhere in this guide, treat it as a snapshot, not a fact:
- A specific model identifier (
gpt-…,claude-…,gemini-…,grok-…, etc.). - A specific context window size in tokens.
- A specific rate limit number.
- A specific dollar/credit cost.
Each of those should have a Last verified: date nearby. If the date is older than ~30 days for a vendor that ships fast (OpenAI, Anthropic, Google, xAI), re-check the official source before relying on it.
Recommended pattern: env-var-driven model selection¶
In any code or config example, prefer reading the model name from an environment variable so readers can update without editing the page:
| Vendor | Recommended env var |
|---|---|
| OpenAI / Codex | OPENAI_MODEL |
| Anthropic / Claude | ANTHROPIC_MODEL |
| Google / Gemini | GEMINI_MODEL |
| xAI / Grok | XAI_MODEL |
| Local (Ollama, llama.cpp) | LOCAL_MODEL |
Python example:
TypeScript example:
When REPLACE_WITH_CURRENT_MODEL appears in an example here, replace it with the current model name from the linked official source — and set the env var locally, don't hand-edit the page.
Where to check current models¶
| Vendor | Canonical "what's current" surface |
|---|---|
| OpenAI | Models docs, Codex CLI docs |
| Anthropic | Claude models docs |
| Gemini models, AI Studio | |
| xAI | xAI overview, models reference |
| Microsoft / GitHub | Copilot plans for what's available, Copilot models when applicable |
| Local | Your model registry of choice (Ollama tags, Hugging Face). |
Drift checklist (run quarterly, or when you bump a recipe)¶
When you edit a page or starter-kit script, walk this checklist before merging:
- Model identifiers in code blocks are either env vars or have a
Last verified:date within 30 days. - Context window claims are still accurate per the vendor's current models doc.
- Rate limit claims are described qualitatively (e.g. "free tier"), or carry a date and a link.
- Cost claims are either qualitative or absent. We do not put dollar figures in the guide.
- Tool / function calling syntax matches the current SDK version pinned in the example.
- Structured-output / response-format syntax matches current docs.
- Auth flows (login, API key location) match the current vendor UI.
- Filesystem paths for desktop apps match current OS conventions.
- Connector / Action UI descriptions match the current product UI.
- The page's
Last verified:line is updated to today. - Any new canonical URL is added to
source-map.md. - If a vendor renamed or removed a feature, log it in
source-audit.md.
What "high-drift" means¶
Per ADR 0002, every page declares a drift risk:
| Risk | Re-check cadence |
|---|---|
| Low | Annually or when a major vendor change is announced. |
| Medium | Quarterly. |
| High | Monthly, plus on any vendor announcement. |
Pages that are automatically high-drift regardless of content:
- Anything about Codex CLI surface (
docs/platforms/codex.md, related recipes). - Anything about Antigravity (
docs/platforms/antigravity.md). - Anything about computer use (Anthropic and OpenAI surfaces).
- Anything about Copilot cloud agent plan availability, organization policy controls, or premium-request accounting.
- Anything that names a specific model.
If you add a new page that fits any of those categories, mark it Drift risk: high and set a calendar reminder for monthly re-check.
How readers should treat stale text¶
If you find a page that names a model, a price, or a rate limit that is wrong:
- Open a Source drift issue.
- Or, file a PR that updates the page and updates
Last verified:and any new sources.
Why we don't just bake current model names into the prose¶
- They go stale within weeks.
- They make the guide look authoritative when it is just a snapshot.
- Env vars work cleanly across Claude, OpenAI, Gemini, Grok, and local examples without per-page rewrites.
- Readers usually know what they want to use; what they need from us is a stable shape, not a model recommendation.