1 · Which model should I pick?
There is no single “best model” in 2026 — there are best fits. The fastest way to choose is by workload type:
| Workload | Pick | Why |
|---|---|---|
| Hardest reasoning research, math, architecture | Claude Opus 5 / GLM-5.3 | Deepest chain-of-thought quality; Opus 5 carries a full 1M-token context window. |
| Daily coding IDE agents, code review | Claude Sonnet 5 / DeepSeek V4 | Best speed-to-quality balance in tools like Claude Code and Codex; DeepSeek V4 is the open-weights value pick. |
| High-volume / bulk classification, extraction, batch | Haiku 4.5 / DeepSeek V4 Flash / Qwen 3.8 Flash | Fraction of the cost with strong instruction following. |
| Multilingual workflows CJK content, long documents, localisation | Qwen 3.8 / Kimi K3 / GLM | Strongest CJK comprehension and generation; Kimi excels at very long context. |
| Vision / multimodal | Qwen 3-VL Plus | Image understanding at a fraction of flagship vision pricing. |
| Codex / GPT ecosystem | GPT-5.5 / 5.4 | Native fit for OpenAI-aligned agent frameworks. |
Rule of thumb: start with the middle tier (Sonnet 5 / DeepSeek V4) for real work, escalate to Opus-class only when the task proves hard, and drop to Haiku/Flash-class for anything repetitive. Your bill shrinks ~10×, quality loss is often zero.
2 · One gateway, three protocols
Model APIs speak different wire formats. ThunderPeak exposes them all behind one key and one balance:
| Protocol | Endpoint | Used by |
|---|---|---|
| OpenAI Chat Completions | /v1/chat/completions | OpenAI SDK, Cherry Studio, most tools |
| Anthropic Messages | /v1/messages | Claude Code, Anthropic SDK, Cursor |
| Responses | /v1/responses | Codex CLI & extension, newer OpenAI SDKs |
Point an OpenAI-compatible tool at https://thunder-peak.com/v1, or an Anthropic tool at https://thunder-peak.com — the same key works on both. If a third-party client reports a path error, append /v1.
3 · How the bill is calculated
Every request meters five token categories — this is the same model the providers use, so our per-token prices map 1:1 to what you'd expect:
| Category | Meaning | Relative cost |
|---|---|---|
| Input tokens | Your prompt, system prompt, tool definitions | 1× |
| Output tokens | The generated answer | typically 5× input |
| Cache write | First time a prompt prefix is cached | 1.25× input |
| Cache read | Reusing a cached prefix (repeat prompts) | 0.1× input — biggest lever |
| Thinking tokens | Hidden reasoning (Opus/Sonnet/thinking models) | output-priced |
Practical tips: keep your system prompt stable to maximise cache reads (10× cheaper than fresh input); streaming costs the same as non-streaming; and usage history in the dashboard breaks every request down to the token.
4 · Why route through a gateway at all?
5 · Glossary
| Term | Plain meaning |
|---|---|
| API key | Your secret credential; send it as Authorization: Bearer sk-tp-… |
| Base URL | The gateway address you paste into a tool, e.g. https://thunder-peak.com/v1 |
| Context window | Max tokens the model can see per request (e.g. 1M ≈ a whole codebase) |
| SSE / streaming | Token-by-token output; what makes chat and agent UIs feel instant |
| Tool calling | The model asking your app to run functions — the basis of agents |
| Reasoning / thinking | Models that "think" before answering; stronger but pricier per token |
| Rate limit | Requests per minute allowed per key — configurable per key in ThunderPeak |