Models & pricing
GuRouter routes across every major model family behind one key and one base URL. Pick a
model by what you are building and how much you want to spend, then send its id in the
model field. Ids are exact, including the vendor prefix.
Showing all 40 models
claude-opus-5.523 tok/s | Other | anthropicopenai | $4.00 | $4.00 | $0.40 |
|---|---|---|---|---|---|
cx/gpt-5.5latest GPT-5.5 for general complex work in Codex3 tok/s | OpenAI | openai | $0.60 | $3.00 | $0.18 |
cx/gpt-5.6-lunacost-sensitive GPT-5.6 Luna for Codex agentic coding7 tok/s | OpenAI | openai | $0.050 | $0.20 | $0.015 |
cx/gpt-5.6-solfrontier GPT-5.6 Sol for complex professional work in Codex18 tok/s | OpenAI | openai | $0.40 | $1.50 | $0.12 |
cx/gpt-5.6-terraGPT-5.6 Terra balancing intelligence and cost for Codex21 tok/s | OpenAI | openai | $0.20 | $1.20 | $0.060 |
cx/gpt-6-astrafrontier GPT-6 Astra for Codex agentic coding13 tok/s | OpenAI | openai | $2.00 | $5.00 | $0.60 |
deepseek/deepseek-v4-flashfast hybrid-attention reasoning.Score 69143 tok/s | DeepSeek | openai | $0.22 | $0.66 | $0.066 |
deepseek/deepseek-v4-flash-fastlow-latency V4 Flash deployment89 tok/s | DeepSeek | openai | $0.28 | $0.56 | $0.042 |
deepseek/deepseek-v4.1-flashV4.1 hybrid-attention reasoning with vision.95 tok/s | DeepSeek | openai | $0.22 | $0.66 | $0.055 |
google/gemini-3.8-flashnewest Gemini Flash, improved core reasoning58 tok/s | openai | $1.50 | $7.50 | $0.15 | |
gpt-6-astra8 tok/s | Other | anthropicopenai | $2.00 | $4.00 | n/a |
gpt-6-luna57 tok/s | Other | anthropicopenai | $0.10 | $0.50 | n/a |
gpt-6-sol30 tok/s | Other | anthropicopenai | $2.00 | $10.00 | $0.20 |
MiniMax-M2.7-highspeedFree 5B on 5h44 tok/sFree | Other | openai | Free | Free | n/a |
MiniMax-M352 tok/sFree | Other | openai | Free | Free | n/a |
minimaxai/minimax-m3frontier coding, agents & native multimodalityScore 5929 tok/s | MiniMax | openai | $0.30 | $1.20 | $0.060 |
MiniMaxAI/MiniMax-M3frontier coding, agents & native multimodalityScore 5931 tok/s | MiniMax | openai | $0.30 | $1.20 | $0.060 |
moonshotai/Kimi-K2.5multimodal frontend codingScore 479 tok/s | Moonshot | openai | $0.60 | $3.00 | $0.10 |
moonshotai/Kimi-K2.6long-horizon coding with visionScore 6210 tok/s | Moonshot | openai | $0.95 | $4.00 | $0.16 |
moonshotai/Kimi-K2.7-Codeimproved long-horizon coding with visionScore 6126 tok/s | Moonshot | openai | $0.95 | $4.00 | $0.19 |
Qwen/Qwen3.6-Max-Previewvibe coding & efficient agent execution49 tok/s | Alibaba | openai | $1.30 | $7.80 | $0.26 |
Qwen/Qwen3.6-Plusagentic coding & reasoningScore 5446 tok/s | Alibaba | openai | $0.50 | $3.00 | $0.10 |
Qwen/Qwen3.7-Flashfast low-cost agentic coding & reasoning80 tok/s | Alibaba | openai | $0.030 | $0.13 | $0.006 |
Qwen/Qwen3.7-Maxfrontier coding & long-horizon agent executionScore 6673 tok/s | Alibaba | openai | $2.50 | $7.50 | $0.50 |
Qwen/Qwen3.7-Plusagentic coding & reasoning at lower costScore 5658 tok/s | Alibaba | openai | $0.40 | $1.60 | $0.080 |
Qwen/Qwen3.8-27Bcompact vision-language coding & agentic workScore 6823 tok/s | Alibaba | openai | $0.40 | $3.00 | $0.040 |
Qwen/Qwen3.8-Maxautonomous long-horizon coding & professional workScore 7225 tok/s | Alibaba | openai | $2.00 | $6.00 | $0.25 |
Qwen/Qwen3.8-Omni-FlashQwen3.8-Omni-Flash: Omni Senses. Agentic Delivery.23 tok/s | Other | openai | $0.56 | $3.73 | $0.056 |
stealth/space-bunny-alpha52 tok/sFree | Other | openai | Free | Free | n/a |
swe-2-highSWE-2 with high reasoning effort for autonomous software engineering41 tok/sFree | Other | openai | Free | Free | n/a |
swe-2-maxSWE-2 at maximum reasoning effort for the hardest engineering tasks26 tok/s | Other | openai | $75.00 | $75.00 | n/a |
tencent/hy3-paidsparse-MoE reasoning & agentic tool use.
Score 5922 tok/s | Tencent | openai | $0.14 | $0.58 | $0.040 |
xai/grok-4.7Grok 4.7 is SpaceXAI's flagship model for coding, agentic tasks, and knowledge work, succeeding Grok 4.6.
save 65% for the next 7 days42 tok/s | Other | openai | $2.50 | $10.00 | $0.50 |
xiaomi/mimo-v2.5efficient long-context agentic codingScore 574 tok/s | Xiaomi | openai | $0.14 | $0.28 | $0.003 |
xiaomi/mimo-v2.5-prohigh-capability long-context agentic codingScore 607 tok/s | Xiaomi | openai | $0.43 | $0.87 | $0.004 |
xiaomi/mimo-v2.6-flashfast low-cost MiMo with long-context agentic coding10 tok/s | Xiaomi | openai | $0.14 | $0.28 | $0.028 |
z-ai/glm-5.3-flashxGLM-5.3-FlashX is the high-speed variant of Z.ai's GLM-5.3-Flash, a native multimodal model delivering inference speeds of up to 200 tokens/s56 tok/s | Other | openai | $0.60 | $3.20 | $0.060 |
zai-org/GLM-5multi-mode thinking & long-range planning16 tok/s | Z.AI | openai | $1.00 | $3.20 | $0.20 |
zai-org/GLM-5.1long-horizon autonomous coding agentScore 562 tok/s | Z.AI | openai | $1.40 | $4.40 | $0.26 |
zai-org/GLM-5.2powerful coding with 1M context and long-horizon tasksScore 6917 tok/s | Z.AI | openai | $1.40 | $4.40 | $0.26 |
No model matches those filters. Clear the search or pick a different vendor.
Rates are per 1M tokens, in USD, and are the effective rate for the plan you selected above. Scores and speeds are as published by the gateway; a dash means the vendor has not published one. Last synced— see how this is calculated.
Choosing a model
Section titled “Choosing a model”Most people should not read the whole table. Three questions decide it.
Do you need a specific model? Some tasks have a known best answer. If a model family or an exact id is required, check the table above — availability also depends on your plan.
What is the task?
- Agent loops and long sessions. Most input in an agent run is re-read context, so cache-read rate matters more than headline input price. A cheap model with a good cache rate often beats an expensive one over a long session.
- Bulk work where output dominates, such as batch generation or classification of a large corpus. Output is typically 3–4× the input rate, so compare output price first.
- One-off or short requests. A short request never amortises cache, so the raw input rate is the honest comparison.
How fast? Speed is published as tokens per second. Above roughly 50 tok/s the model feels immediate; below roughly 15 tok/s, long answers visibly stall. A slower model is often fine for background work and painful for anything interactive.
Which models support which endpoint
Section titled “Which models support which endpoint”Every model is reachable through the OpenAI-compatible route. Four also accept the Anthropic Messages format, which is what Claude Code and other Anthropic-shaped clients send.
| Endpoint | Path | Models |
|---|---|---|
| Chat (all models) | POST /v1/chat/completions |
all 40 |
| Anthropic Messages | POST /v1/messages |
claude-opus-5.5, gpt-6-astra, gpt-6-luna, gpt-6-sol |
| SystemOne (decision) | POST /v1/systemone |
typesafe/jev |
How pricing is calculated
Section titled “How pricing is calculated”The charge is not the headline number. Two multipliers apply, and skipping either gives you a price that does not match your bill.
Step 1 — base rate per 1M tokens. Derived from the model’s published ratio:
base input = 2 × model_ratiobase output = base input × completion_ratiobase cache = base input × cache_ratioStep 2 — your plan’s multiplier. Every model is priced against the group it is enabled in, and each group carries a discount:
| Plan | Multiplier | Example: a $2.50 base model |
|---|---|---|
| Pro | ×0.45 | $1.12 |
| VIP | ×0.50 | $1.25 |
| Subscriber | ×0.60 | $1.50 |
| Default | ×0.70 | $1.75 |
| Vibecode | ×0.70 | $1.75 |
A model can sit in several groups at once. A model available on both Pro and VIP is cheaper for you on Pro, so the table asks which plan you are on rather than guessing.
Worked example
Section titled “Worked example”deepseek/deepseek-v4-flash has a base rate of $0.22 / $0.66 / $0.066 per 1M input,
output and cache-read tokens. On the Pro plan (×0.45):
| Base | On Pro (×0.45) | |
|---|---|---|
| Input | $0.22 | $0.099 |
| Output | $0.66 | $0.297 |
| Cache read | $0.066 | $0.0297 |
An agent turn that sends 10K fresh input, 2K cache-read and writes 1K output costs
10000/1e6 × 0.099 + 2000/1e6 × 0.0297 + 1000/1e6 × 0.297, which is about $0.0013.
The cache-read column is the one to watch in agent loops. On a long session the majority of input tokens are re-read context billed at the cache rate, not the input rate, so the real cost sits much closer to the cache figure than to the headline input price.
Calling a model
Section titled “Calling a model”The id in the table is the id you send, verbatim:
curl https://gurouter.com/v1/chat/completions \ -H "Authorization: Bearer <API_KEY>" \ -H "Content-Type: application/json" \ -d '{ "model": "deepseek/deepseek-v4-flash", "messages": [{ "role": "user", "content": "Say hello in one sentence." }] }'Only the four models marked anthropic above. Used by Claude Code and other
Anthropic-shaped clients.
curl https://gurouter.com/v1/messages \ -H "Authorization: Bearer <API_KEY>" \ -H "Content-Type: application/json" \ -H "anthropic-version: 2023-06-01" \ -d '{ "model": "claude-opus-5.5", "max_tokens": 256, "messages": [{ "role": "user", "content": "Say hello in one sentence." }] }'A decision endpoint, not a chat. It does not stream and takes a different request shape. See the SystemOne reference.
curl https://gurouter.com/v1/systemone \ -H "Authorization: Bearer <API_KEY>" \ -H "Content-Type: application/json" \ -d '{ "model": "typesafe/jev", "state": "…", "questions": { } }'Confirm the exact ids your key can reach with GET /v1/models. It
returns the same list at call time, including each model’s supported_endpoint_types.
Staying current
Section titled “Staying current”The table is a committed snapshot, regenerated from the live pricing API:
bun run sync:modelsThat keeps a docs build from depending on a third-party endpoint, and makes a price change a reviewable diff rather than something that silently changes between one reader and the next. Last synced 26 September 2026.
