CocoboxAI models

The curated set of models Cocobox runs on your behalf — what they're good at and what they cost.

CocoboxAI is the managed-models option. We pick a small, opinionated lineup across the major frontier providers, expose them under stable aliases, and charge a flat credit price per request.

The lineup

Alias	Underlying	Best for	Credits / request
tambor	Amazon Nova Lite	Everyday SQL, autocomplete, the cheapest tier	0.33
trueno	Mistral 7B	Fast classifications, simple Q&A	1
prisma	Amazon Nova Pro	Day-to-day coding, SQL, reviews	3
sonnet	Anthropic Claude Sonnet 4.6	Complex SQL, multi-step reasoning	3
sonnet-5	Anthropic Claude Sonnet 5	Newest Sonnet, faster and cheaper	2
haiku	Anthropic Claude Haiku 4.5	Snappy answers with strong reasoning, MCP tool loops	1
deepseek-r1	DeepSeek R1	Math-heavy / analytical	2
nova-premier	Amazon Nova Premier	Hardest queries, deep reasoning	3
opus	Anthropic Claude Opus 4.6	The most demanding work	7
opus-4-7	Anthropic Claude Opus 4.7	Newest Opus, frontier reasoning	7
opus-4-8	Anthropic Claude Opus 4.8	Latest Opus, frontier reasoning	7
fable-5	Anthropic Claude Fable 5	Highest-tier model, hardest problems	10

Aliases are stable. The underlying model can be upgraded — we’ll announce it in the changelog and quality-test before flipping.

Why aliases

Frontier model names change every few months. By coding to prisma instead of claude-sonnet-4-5-20250929, your snippets, scripts, and integrations don’t break when the provider deprecates a checkpoint.

Picking a model

A simple decision tree:

1. Tight latency budget or burning through credits? → tambor (0.33 credits — cheapest).
2. Day-to-day SQL editing or chat? → haiku (1 credit) for cheaper, or prisma/sonnet (3 credits) for stronger reasoning.
3. Tool-heavy MCP loops? → haiku is best for tool calling (also what Cocobox auto-routes to when tools are attached).
4. Stuck — needs deep thinking on a hard refactor? → opus, opus-4-7, or opus-4-8 (7 credits).
5. The hardest problem you’ve got? → fable-5 (10 credits).

Defaults in the product:

• Editor inline suggestions: tambor.
• “Fix this error”: haiku.
• Chats: sonnet or prisma.
• “Plan a migration”: opus.

You can override defaults in Settings → AI → Defaults.

Limits

• Context window: each alias inherits the underlying model’s context window.
• Concurrent requests: 8 per workspace by default; raised on Team and Enterprise.
• RPM: 120 per workspace. Burstable.

For exact request/response shapes, see API → Chat completions and API → Messages.