Copy page

Get Started

Models

Hyphen serves ten models behind one base URL. They are real provider model names — all MiniMax models, attributed as such. Put the exact ID in the model field of a request; there are no invented aliases, and any other ID returns an error.

Rates are USD per 1M tokens, metered per token. These are the rates — there are no others: credits and subscription allowances both spend at exactly these numbers.

Model	One-line description	Context	Provider	Input / 1M	Output / 1M
minimax-m3	Flagship — newest M-series generation, built for hard reasoning	~205k	MiniMax	$0.30	$1.20
minimax-m2.7	Balanced daily driver — the sensible default for most agent work	~205k	MiniMax	$0.30	$1.20
minimax-m2.5	Fast — cheap, low-latency, high-volume calls	~205k	MiniMax	$0.30	$1.20
minimax-m2.7-highspeed	Latency-tuned minimax-m2.7 — same model, faster serving, 2x price	~205k	MiniMax	$0.60	$2.40
minimax-m2.5-highspeed	Latency-tuned minimax-m2.5 — same model, faster serving, 2x price	~205k	MiniMax	$0.60	$2.40
minimax-m2.1-highspeed	Latency-tuned minimax-m2.1 — same model, faster serving, 2x price	~205k	MiniMax	$0.60	$2.40
minimax-m2.1	Previous-generation midrange — superseded by m2.5, same rate	~205k	MiniMax	$0.30	$1.20
minimax-m2	Oldest M-series generation still served — for legacy configs	~205k	MiniMax	$0.30	$1.20
minimax-m2-her	Role-play / dialogue chat tuning — not a coding model	66k	MiniMax	$0.30	$1.20
minimax-text-01	Long-context text — for very large inputs	4M	MiniMax	$0.20	$1.10

Three groups, in the order of the table:

• Flagship / daily / fast — minimax-m3, minimax-m2.7, minimax-m2.5. Start here; these are the models the guides use.
• Highspeed variants — minimax-m2.7-highspeed, minimax-m2.5-highspeed, minimax-m2.1-highspeed. The same weights as their standard siblings on latency-tuned serving, at exactly twice the per-token rate. Pick one only when time-to-first-token matters more than cost.
• Older generations & specialty — minimax-m2.1 and minimax-m2 are earlier M-series generations kept so existing configs keep working. Note the rate card: they meter at the same rate as minimax-m3, so there is no price reason to pick them — only compatibility. minimax-m2-her is tuned for role-play and dialogue chat (66k context); minimax-text-01 is the long-context text model and the cheapest per token.

All ten are served by the same gateway and billed against the same monthly cap on your key.

How metering works

Per-token, at the rates in the table, everywhere. A request costs input_tokens × input rate + output_tokens × output rate — no request fees, no minimums, no markup layered on top. The same rates apply to every way you pay:

• Credits are 1:1. $1 of prepaid credit buys $1 of usage at the listed rates, with no deposit or platform fees.
• Subscription allowances draw down at the same rates. A plan's monthly allowance is spent per token exactly like credit — there is no separate, worse "subscription rate."

Every response includes a usage field with the exact input and output token counts that were billed, so you can audit any request against this table — multiply it out yourself and it will match what the console shows.

For scale: at a typical coding-agent mix (~3:1 input

), $10 buys roughly 19M tokens on the standard M-series, ~9.5M on a highspeed variant, or ~23.5M on minimax-text-01.

The catalog can change — with notice

These are real model IDs, not permanent aliases. When MiniMax ships a new generation, or Hyphen adds or retires a model, the catalog changes and we announce it in advance so you can update your configs. GET /v1/models always returns the current catalog — treat it as the source of truth.

Staged next: DeepSeek V4 Flash (deepseek-v4-flash) is coming to the catalog. It is not live yet — it will appear in GET /v1/models when it ships.

Fallback during provider incidents

During a provider incident, a request may be served by a smaller model of the same family instead of failing. The gateway never silently swaps in a different provider's model. The fallback ladders are:

• minimax-m3 → minimax-m2.7 → minimax-m2.5
• Any -highspeed model → its standard sibling → the next model down (for example minimax-m2.7-highspeed → minimax-m2.7 → minimax-m2.5)
• minimax-m2.5 → minimax-m2.1 → minimax-m2
• minimax-m2, minimax-m2-her, and minimax-text-01 have no fallback — during an incident those requests fail rather than being served by a different model.

The model field in every response reports the model that actually served the request, so you can always tell.

List what your key can reach

Code
curl https://api.hyphen-solution.com/v1/models \
  -H "Authorization: Bearer $HYPHEN_API_KEY"

Code
{
  "object": "list",
  "data": [
    { "id": "minimax-m2", "object": "model", "owned_by": "minimax" },
    { "id": "minimax-m2-her", "object": "model", "owned_by": "minimax" },
    { "id": "minimax-m2.1", "object": "model", "owned_by": "minimax" },
    { "id": "minimax-m2.1-highspeed", "object": "model", "owned_by": "minimax" },
    { "id": "minimax-m2.5", "object": "model", "owned_by": "minimax" },
    { "id": "minimax-m2.5-highspeed", "object": "model", "owned_by": "minimax" },
    { "id": "minimax-m2.7", "object": "model", "owned_by": "minimax" },
    { "id": "minimax-m2.7-highspeed", "object": "model", "owned_by": "minimax" },
    { "id": "minimax-m3", "object": "model", "owned_by": "minimax" },
    { "id": "minimax-text-01", "object": "model", "owned_by": "minimax" }
  ]
}

Picking a model for a coding agent

Coding agents lean hard on whichever model you set as primary. For Claude Code, Codex CLI, Hermes, and OpenClaw we suggest minimax-m3 (the flagship) as the primary model, with minimax-m2.5 for cheap background calls. Each agent page shows the exact config. minimax-m2-her is a chat-tuned model — don't wire it into a coding agent.

Related

• Quickstart — your first request.
• Connect your coding agent — per-tool setup.
• API Reference — every endpoint, interactively.