Skip to Content
Models

Models

GET https://api.devotel.com/v1/models
{ "object": "list", "data": [{ "id": "kimi-k3", "object": "model", "created": 0, "owned_by": "system" }] }

created is always 0 — the catalogue does not track model creation dates. It is present for shape compatibility with clients that expect the field.

GET /v1/models lists what YOUR key can reach, not the whole catalogue. A key can be restricted to specific models, and a few models are available only on request — so two keys can legitimately see different lists. Pass any of the ids below as "model".

auto is in that list. Sending "model": "auto" lets the gateway pick the cheapest model that can actually serve the request, keeps a conversation on the model it started with, and bills you for the model that ran. See Auto (cost optimized).

The catalogue

ModelIDTypeToolsVisionIn / 1MOut / 1M
Auto (cost optimized)autochatYesYes——
Claude Haiku 4.5anthropic-claude-haiku-4.5chatYesYes$1$5
Claude Sonnet 5anthropic-claude-5-sonnetchatYesNo$2$10
Claude Sonnet 5.5anthropic-claude-sonnet-5.5chatYesYes$2$10
Claude Sonnet 4.5anthropic-claude-4.5-sonnetchatYesYes$3$15
Claude Sonnet 4.6anthropic-claude-4.6-sonnetchatYesYes$3$15
Claude Opus 5.5 *anthropic-claude-opus-5.5chatYesYes$4$20
Claude Opus 4.5anthropic-claude-opus-4.5chatYesYes$5$25
Claude Opus 4.6anthropic-claude-opus-4.6chatYesYes$5$25
Claude Opus 4.7anthropic-claude-opus-4.7chatYesYes$5$25
Claude Opus 4.8anthropic-claude-opus-4.8chatYesYes$5$25
Claude Opus 5anthropic-claude-opus-5chatYesYes$5$25
Claude Fable 5 *anthropic-claude-fable-5chatYesYes$10$50
Claude Fable 5.1 *anthropic-claude-fable-5.1chatYesYes$10$50
GPT-5 nanoopenai-gpt-5-nanochatYesYes$0.05$0.4
GPT-OSS 20Bopenai-gpt-oss-20bchatYesNo$0.05$0.45
GPT-OSS 120Bopenai-gpt-oss-120bchatYesNo$0.06$0.39
GPT-6 Lunaopenai-gpt-6-lunachatYesYes$0.1$0.5
GPT-4o miniopenai-gpt-4o-minichatYesYes$0.15$0.6
GPT-5.4 nanoopenai-gpt-5.4-nanochatYesYes$0.2$1.25
GPT-5.6 Lunaopenai-gpt-5.6-lunachatYesYes$0.2$1.2
GPT-5 miniopenai-gpt-5-minichatYesYes$0.25$2
GPT-5.4 miniopenai-gpt-5.4-minichatYesYes$0.75$4.5
o3-miniopenai-o3-minireasoningYesNo$1.1$4.4
GPT-5openai-gpt-5chatYesYes$1.25$10
GPT-5.2openai-gpt-5.2chatYesYes$1.75$14
GPT-5.3 Codexopenai-gpt-5.3-codexcodingYesYes$1.75$14
GPT-4.1openai-gpt-4.1chatYesYes$2$8
GPT-5.6 Terraopenai-gpt-5.6-terrachatYesYes$2$12
GPT-6 Solopenai-gpt-6-solchatYesYes$2$10
o3openai-o3reasoningYesYes$2$8
GPT-4oopenai-gpt-4ochatYesYes$2.5$10
GPT-5.4openai-gpt-5.4chatYesYes$2.5$15
GPT-5.5openai-gpt-5.5chatYesYes$5$30
GPT-5.6 Solopenai-gpt-5.6-solchatYesYes$5$30
GPT-6 Astra *openai-gpt-6-astrachatYesYes$10$50
o1openai-o1reasoningYesYes$15$60
GPT-5.2 Pro *openai-gpt-5.2-prochatYesYes$21$168
GPT-5.4 Pro *openai-gpt-5.4-prochatYesYes$30$180
DeepSeek 4 Flashdeepseek-4-flashchatYesNo$0.07$0.17
DeepSeek V4 Flash (0731)deepseek-v4-flash-0731chatYesNo$0.08$0.25
DeepSeek 3.2deepseek-3.2chatYesNo$0.25$0.8
DeepSeek V4 Pro (0813)deepseek-v4-pro-0813chatYesNo$1.32$3.96
DeepSeek V4 Prodeepseek-v4-prochatYesNo$1.74$3.48
Kimi K2.6kimi-k2.6chatYesYes$0.95$4
Kimi K3 (DigitalOcean)kimi-k3reasoningYesYes$3$15
Qwen3.5 397B A17Bqwen3.5-397b-a17bchatYesYes$0.55$3.5
Qwen3.8 Maxqwen3.8-maxchatYesYes$2$6
GLM-5.3 Flashglm-5.3-flashchatYesYes$0.15$0.5
GLM-5.2glm-5.2chatYesNo$1.4$4.4
GLM-5.3glm-5.3chatYesNo$1.4$4.4
Gemini 3.1 Flash-Litegemini-3.1-flash-litechatYesYes$0.25$1.5
Gemini 3.7 Flashgemini-3.7-flashchatYesYes$0.75$3.75
Gemini 3.8 Flashgemini-3.8-flashchatYesYes$0.75$3.75
Gemini 3.1 Progemini-3.1-pro-previewreasoningYesYes$2$12
Gemma 4 31B ITgemma-4-31B-itchatYesYes$0.18$0.5
Llama 4 Maverickllama-4-maverickchatYesYes$0.2$0.7
Mistral 3 14Bmistral-3-14BchatYesNo$0.2$0.2
Nemotron Nano 12B v2 VLnemotron-nano-12b-v2-vlchatYesNo$0.2$0.6
MiniMax M2.5minimax-m2.5chatYesNo$0.3$1.2
MiMo v2.5 Promimo-v2.5-prochatYesNo$0.4$1.5
Nemotron 3 Nano Omninemotron-3-nano-omnichatYesYes$0.5$0.9
Nemotron 3 Ultra 550Bnemotron-3-ultra-550bchatYesNo$0.9$1.7

63 models. Prices are USD per million tokens; reasoning tokens are billed as output. Tools and vision are measured against the live provider, not taken from a datasheet. * Restricted: available on request — ask your administrator to grant it to a key.

Several of these reason before answering, even with no reasoning_effort field sent at all — a small max_tokens can get cut off mid-thought and come back with no visible answer. See Reasoning models and max_tokens for which models, and the two fixes.

Model access is a per-key setting. By default a key may use every generally available model; you or your administrator can narrow it to a specific list when the key is created, or later from the admin panel. Asking for a model your key cannot use returns the same 404 model_not_found as asking for one that does not exist — see Rate Limits & Errors.

kimi-k3

Moonshot AI’s Kimi K3, served through DigitalOcean Serverless Inference. It is the default and the most thoroughly audited model here, so the detail below is about it specifically; the table above covers the rest.

The same weights are also available directly from Moonshot, as kimi-k3-ms, added after DigitalOcean’s kimi-k3 capacity proved unreliable during a service incident and DO was not able to raise this account’s limits. It is a separate model id on purpose — a different provider means a different prompt cache, a different invoice, and independent verification — so a session already running on kimi-k3 is never silently moved there. auto will fail over to it automatically if kimi-k3 hard-fails; you can also select kimi-k3-ms directly. reasoning_effort works exactly as it does on kimi-k3, including none — measured live, not assumed. Moonshot bills a cache-WRITE cost DigitalOcean does not; see Pricing below.

CapabilityStatus
Tool calling / function callingYes
Image input (image_url parts)Yes
Reasoning (reasoning_content)Yes
Streaming (SSE)Yes
Context window1,048,576 tokens

Reasoning

Kimi K3 reasons before it answers. Every response carries reasoning_content next to content, and those thinking tokens are billed as output tokens.

"message": { "content": "Standing in between—\nI speak your words, not my own,\nfaceless messenger.", "reasoning_content": "The user wants a haiku about proxies. Proxies could refer to..." }

Measured on this endpoint: that 72-character haiku cost 799 completion tokens, 782 of them reasoning. A five-word greeting cost 158 completion tokens, 141 of them reasoning. Plan for reasoning to dominate output cost on short answers, and read Chat Completions before setting max_tokens.

Tool calling

Verified end to end: the model returns finish_reason: "tool_calls" with a tool_calls array, accepts role: "tool" results, and then answers. Complete loops in Python and TypeScript are in Tool Calling & Agents.

Image input

Send images as image_url content parts, either a data: URI or a URL the provider can fetch.

Verified against this endpoint with generated test images: the model named the shapes and their colours, read the embedded text, and counted objects correctly across three different images — for example “a red circle on the left and a blue triangle on the right, on a white background. Below them, in bold black text, it says: DEVOTEL 42”, and “three purple circles arranged in a row near the top, with a large black number 7 below them”.

{ "role": "user", "content": [ { "type": "text", "text": "What does this diagram show?" }, { "type": "image_url", "image_url": { "url": "data:image/jpeg;base64,/9j/4AAQ..." } } ] }

A 900×640 image cost about 890 prompt tokens in these tests, so images are billed like a page of text rather than a rounding error.

Images count towards the request body, and /v1 bodies are capped at 20 MB — a base64 data URL is about 4/3 the size of the file, so budget roughly 15 MB of actual image bytes across the whole request, or pass a URL instead. Downscale before sending anyway: the model does not need more than about 2000 px on the longest side, and every pixel is billed as prompt tokens.

Context window

The provider’s /v1/models metadata publishes no context length for kimi-k3 at all — unlike kimi-k2.5/kimi-k2.6, which publish 262,144. Rather than leave K3 unstated indefinitely, we measured it: a live probe on 2026-09-08 sent real requests at increasing sizes (max_tokens: 1, filler content, cost tracked from the provider’s own prompt_tokens echo) until the provider rejected one, then read the limit out of its own error text.

ModelContext windowHow we know
kimi-k31,048,576 tokensMeasured 2026-09-08 — provider’s /v1/models publishes nothing for this one
kimi-k2.5 / kimi-k2.6262,144 tokensPublished by the provider’s /v1/models, confirmed by the same 2026-09-08 probe

The number is the combined input + output budget, not input alone — the provider’s own message below counted our one requested output token as part of the total. Exceeding it gets you exactly this, verbatim, as a 400 invalid_request_error (the gateway passes a provider 400 straight through rather than masking it as a 502 — see Rate Limits & Errors):

{ "error": { "message": "This model's maximum context length is 1048576 tokens. However, you requested 1 output tokens and your prompt contains 1048708 input tokens, for a total of 1048709 tokens. Please reduce the length of the input prompt or the number of requested output tokens. (parameter=input_tokens, value=1048708)", "type": "invalid_request_error", "code": null, "param": null } }

No error observed at any size named a separate limit on output tokens alone — only the combined total above.

Pricing

Billed per token, at the rate for the direction of travel. Reasoning tokens are output tokens.

ModelInputCache readOutput
kimi-k3$3.00 / 1M tokens$0.30 / 1M tokens$15.00 / 1M tokens

Cached input

When a request repeats a prompt prefix the provider has already seen, those tokens are served from its cache and billed at the cache read rate — an order of magnitude below the input rate. Two things follow:

  • usage.prompt_tokens includes cached tokens; it is the total prompt size, not the billable part. Full-price input is prompt_tokens - cached_tokens.
  • The cached count arrives as usage.prompt_tokens_details.cached_tokens.

Nothing needs enabling — caching is the provider’s, and a stable prefix (a long system prompt, a fixed set of tool definitions, a document you ask several questions about) is what earns the discount. Putting the variable part of a prompt last is what makes a prefix reusable.

The usage dashboard  shows your daily token counts, the cached share, and an estimated cost computed from exactly these numbers.

kimi-k3-ms (Moonshot direct) has a cache-write cost DigitalOcean does not charge at all. The first time a prompt prefix is seen it is billed as a cache WRITE rather than plain input — at the default 5-minute TTL this happens to be the same rate as input, but at Moonshot’s 1-hour TTL option it is double. The dashboard’s cost estimate for this model does not include the write premium yet, so treat it as a lower bound rather than an exact figure until that formula is proven against a real invoice.

Worked example

A request with 200 input tokens and 115 output tokens (reasoning included):

input:  200 / 1,000,000 × $3 = $0.00060
output: 115 / 1,000,000 × $15 = $0.00172
total                            ≈ $0.00232

Output is 5× the input rate, and reasoning lands on the output side — so on this model the cost of a request is driven by how hard it thinks, not by how long your prompt is.

Choosing a model

There is one. When more arrive they appear in GET /v1/models and on this page; existing model ids keep working.