Full model list

The platform currently offers 123 callable models: 55 text chat models across 9 vendors plus the platform's own xiaozhanAI, plus 23 image (including the platform's own xiaozhanimage) and 38 video generators, and 6 music / voice / avatar SKUs (last section on this page). Text pricing is in credits per 1M tokens; images bill per image and video per second.

This page is generated from the same pricing registry the gateway bills against, so every id here is callable and every price is the one you are charged.

⚠️ The 123 here and the 124 on the model catalogue count different things — this is not drift: the plaza additionally shows not-yet-launched placeholder cards (badged "coming soon"), while this page lists only ids you can call today. Some SKUs that are not on the plaza (API-only, or temporarily not shown) appear on neither. Live data: GET https://api.dflop.top/api/v1/models/public, or the model catalogue.

Reading the tables#

Video analysis in web chat: in the AI chat on model.dflop.top, a text card that cannot read video can still take an upload (up to 300 MB). The model decides whether to open the cloud sandbox and extracts frames there with ffmpeg. It sees the picture and any text in it, and cannot hear sound. There is no separate charge: pictures bill as the model's tokens, and using the sandbox bills the usual container fee. This is only in web chat, not in the API; supports_video_analysis in /api/v1/models/public says which cards have it.

Cached pricing needs no opt-in: when upstream usage reports cached_tokens, the gateway settles that portion of the input at the cached rate automatically — clients don't (and can't) enable it explicitly. Models without a cached figure bill all input at the Input price.

Only Anthropic Claude, xiaozhanAI and the -bd line charge for cache writes (the xiaozhanAI and -bd write rates are in their own sections). The first time Claude writes a prefix into the cache, those tokens bill at Input × 1.25 (cache_creation_input_tokens); later hits bill at the cached rate above (Input × 0.1). The cache lives about 5 minutes, refreshed on every hit — if no follow-up turn reuses the prefix within that window, the write premium buys nothing. Every other vendor (the other OpenAI cards / xAI / Alibaba / Zhipu / Moonshot / DeepSeek / MiniMax / Tencent / Volcengine) only has a hit discount and does not bill writes separately.

Limited-time promotion: the whole Anthropic Claude line is charged at 40% off (list price × 0.6) and the whole OpenAI GPT text line (except the -bd line) at 70% off (× 0.3), and gemini-4-argon at 50% off (× 0.5), with cached rates discounted equally. The tables show list prices; actual billing always uses the discounted rate, for both API-key calls and in-platform conversations. Discounted models carry an "X% OFF" badge on the model catalogue.

Long-context tier: for grok-4.5 / grok-4.6 / grok-4.7, a prompt of ≥200K tokens bills input, cache-hit and output all at 2× the table price. That is xAI's whole-turn doubling for very long requests, passed straight through (verified to the cent on 2026-08-23). The trigger is the total prompt size of the turn, independent of how much was generated.

gpt-6 / gpt-6-astra / gpt-6-sol / gpt-6-luna / gpt-6.1-sol / gpt-6.1-terra / xiaozhanAI and the seven -bd text cards work the same way: a prompt of ≥272K tokens bills the whole turn at the long-context tier — input and cache-hit at 2× the table price, output at 1.5× — which is OpenAI's official whole-request repricing for very long prompts, passed straight through. No other model has this tier.

Fast tier: the Fast-capable GPT cards (the ones with a "Fast" badge on the model catalogue) accept "service_tier": "priority" in the request body for the faster tier, billed at the Fast price — for gpt-6.1-sol / gpt-6.1-terra / gpt-6-sol / gpt-6-luna that matches OpenAI's official Fast pricing, 2× the table prices (long-context tier and cached rates included); the other Fast-capable GPT cards are 1.5×. Promotions still apply on top. The "Fast N×" badge on a model-catalogue card is that card's multiplier. If the fast lane is temporarily unavailable the turn is served — and billed — at the standard tier instead of failing. xiaozhanAI is not on this list: it always runs on the official Fast tier, and its table price already is the Fast price (see its section). The -bd line has no Fast tier.

Server-side tools bill per call: when a grok model runs web search / tools, the upstream charges 2.022 points per tool call on top of tokens. The count comes from the upstream's usage.num_server_side_tools_used and is written to the server_tool_calls column of every usage-log row, so you can re-derive your bill. Turns that use no tools incur no such charge. xiaozhanAI and the -bd line have their own per-search web-search price, see their sections.

Drawing and the cloud sandbox#

The text models below can draw pictures and can use a cloud sandbox to make Word / PowerPoint / Excel / PDF files, both in AI Chat and over the API. The model decides from what you say: ask for a poster and it draws one, ask for a deck and it makes one, just chat and it answers in text.

• Billing: tokens bill at the model's prices as usual. Cards that draw through the platform also pay, per image, the current price of the image card GPT-Image 2.5 Flare. The cloud sandbox costs 194.112 points per conversation per 20 minutes, charged only when a command actually ran in it; more files within the same 20 minutes cost nothing extra.
• AI Chat: finished files stay in the conversation as file cards you can preview and download. Images and files uploaded in the conversation, and earlier pictures and files, are placed in the sandbox so the model can read and change them (up to 12, 20 MB each). Turns on the Fast tier, and turns when the balance cannot cover one container fee, do not use the sandbox; knowledge-base questions and model comparison have neither feature.
• API: as with OpenAI, the tools are used only when the request declares them: /v1/responses takes image_generation, shell or code_interpreter; /v1/chat/completions takes image_generation only (with stream: true; the picture comes as ![](link) in the answer). Without the tools it is the plain model. See Cloud sandbox and Image editing. A turn whose balance cannot cover one container fee does not use the sandbox; the model answers in text.

Anthropic (14)#

OpenAI (17)#

• The -bd line (gpt-5.6-luna-bd / gpt-5.6-sol-bd / gpt-5.6-terra-bd / gpt-6-astra-bd / gpt-6-luna-bd / gpt-6-sol-bd / gpt-6.1-sol-bd): a separate route (the bd line) billed at OpenAI's official Standard list price — no limited-time promotion and no Fast tier, so the table price is what you are charged; a prompt of ≥272K tokens bills the whole turn at the long-context tier (see "Long-context tier" at the top of this page).
• On the -bd line, cache writes bill at Input × 1.25 (cache_write_tokens in usage) and cache hits at the cached rate; web search (web_search tool) bills 4.044 points per search, counted from the upstream's reported tool_usage.web_search.num_requests and written to the server_tool_calls column of the usage log.
• The -bd line serves OpenAI Chat and OpenAI Responses. Reasoning effort on Responses (reasoning.effort) accepts low / medium / high / xhigh / max; on Chat (reasoning_effort) the highest is xhigh (Chat requests that carry function tools follow the Responses rules and accept max). gpt-5.6-*-bd, gpt-6-sol-bd and gpt-6-luna-bd also accept none. Function tools, image input, file input (PDF) and web search (with citations) are supported; in-conversation image generation (the image_generation tool) is not. /v1/responses/compact is not supported.

xiaozhanAI (1)#

The platform's own-brand chat model, served straight from the official OpenAI API (no account pool or relay). Every call runs on OpenAI's official Fast tier (about 1.7× the standard output speed); the table shows the Fast price, and a service_tier in the request does not change the tier or the bill. Not part of the limited-time promotion.

• Web search (web_search tool) bills 8.088 points per search, counted from OpenAI's reported tool_usage.web_search.num_requests (a search counts once; opening a page is not charged separately) and written to the server_tool_calls column of the usage log.
• Image generation (image_generation tool) bills 170.6568 points per image, independent of tokens.
• Cache writes bill at 151.65 points / 1M (cache_write_tokens in usage); cache hits bill at the cached rate.
• Cloud sandbox (produces Word / Excel / PPT / PDF files): /v1/responses with the shell or code_interpreter tool; download files via /v1/containers/{id}/files/{file_id}/content. Container fee: 194.112 points per container per 20 minutes. Usage and limits: Cloud sandbox.
• Protocols: OpenAI Chat and OpenAI Responses (the cloud sandbox is Responses-only). background: true is not supported (use stream: true for long requests).

Google (6)#

• gemini-4-argon: the Gemini 4 flagship reasoning model, with thinking levels low / medium (default) / high (Gemini native thinkingConfig.thinkingLevel). Supports function calling, image input and responseSchema structured output; no web search (googleSearch).
• gemini-3.8-flash-nx / gemini-4-argon-nx: a separate relay line (nx) serving models under the same names as gemini-3.8-flash / gemini-4-argon, billed at Google's list price per token (no discount). Supports function calling, image input and responseSchema structured output; no web search (googleSearch); gemini-4-argon-nx takes the same thinking levels as gemini-4-argon. Cache hits are not reported, so input is billed in full. This line has limited capacity: under high concurrency responses get noticeably slower or fail.
• gemini-3.1-pro-preview-1020: a separate relay line (1020), billed at Google's list price per token. Supports Google Search grounding (googleSearch), responseSchema structured output, function calling and image input; it does not report cache hits, so input bills at the full rate.

xAI (3)#

Zhipu GLM (4)#

• The -1022 line (glm-5.3-1022 / glm-5.3-flash-1022): a separate route billed at Zhipu's official price (glm-5.3-flash-1022 uses the standard price, not the limited-time discount). Both serve OpenAI Chat, Responses and Anthropic Messages; thinking is always on and cannot be turned off (thinking.type=disabled returns 400), and reasoning_effort accepts low / high / max. glm-5.3-1022 is text-only; glm-5.3-flash-1022 accepts images. /v1/responses/compact is not supported.

DeepSeek (3)#

• Every DeepSeek card follows DeepSeek's official peak / off-peak prices: the table shows the peak price. Peak is Beijing time Mon–Fri (excluding Chinese public holidays) 9:00–12:00 and 14:00–18:00; everything else (including weekends and public holidays all day) is off-peak, when input, cache hits and output are all half price (e.g. deepseek-v4-flash and deepseek-v4.1-flash off-peak 60 (cached 1.2) / 240, deepseek-v4-pro off-peak 270 (cached 9) / 810). Decided by the time the request reaches the gateway.

Moonshot Kimi (5)#

⚠️ kimi-k3 / kimi-k3-1013 parameter limits (fixed by the upstream; any other value returns 400, same on /v1/chat/completions and /v1/messages): temperature must be 1, top_p must be 0.95, presence_penalty / frequency_penalty must be 0, n must be 1, stop takes at most 5 entries, top_logprobs is not supported; with thinking enabled, tool_choice cannot name a specific tool. Simply leave these parameters out. The model thinks before answering, so don't set max_tokens too low (the reply comes back empty); on the Anthropic protocol (/v1/messages) only streamed responses report cache hits and get the cached-input price.

MiniMax (1)#

Alibaba Qwen (2)#

• qwen3.8-max-cjx: a separate route (cjx) billed at the vendor's official list price; it serves OpenAI Chat and Responses. On Responses, previous_response_id and /v1/responses/compact are not supported (send the full history each turn), and the web_search tool does not actually search. Thinking tokens are not capped by max_tokens and are billed as the upstream reports them.

Image generation models (23)#

Billed per image through POST /v1/images/generations. See Image / video APIs.

xiaozhanimage#

The platform's own image model, served directly by the official OpenAI API (no pool or relay).

• A flat 30 credits per image, whatever the size or number of reference images; quality is fixed to medium by the platform (a value in the request is ignored).
• size is honoured (pixel width×height, e.g. 1024x1536): both sides must be multiples of 16 and the long side ≤ 3840, up to 3840x2160; auto also works.
• Reference-image edits go through POST /v1/images/edits (or send image on a generations request), at most 9 reference images per request; n returns several images, billed per image actually returned.
• stream / partial_images are not supported; response_format defaults to a URL, b64_json returns base64.

ByteDance Seedream#

Google#

OpenAI#

• The three -bd image cards (the bd line) bill per token, not per image, at OpenAI's official list rates: text input 2022 / image input 3235.2 / image output 12132 points per 1M tokens (reference images count as image input). size is honoured (both sides must be multiples of 16), and so is quality (low / medium / high / auto); reference-image edits go through POST /v1/images/edits.
• gpt-image-2.6-flare / gpt-image-2.6-sunburst (GPT Image 2.6, served straight from OpenAI's official API) also bill per token, at OpenAI's official list rates with no discount: text input 2022 / image input 3235.2 / image output 12132 points per 1M tokens. size is honoured (both sides multiples of 16, long side ≤ 3840), and so is quality (low / medium / high / xhigh / max / auto); reference-image edits go through POST /v1/images/edits (up to 9 references). They can make animated GIFs: output_format: "gif", with optional frame_count (2–16, default 9) and frame_duration_ms (20–2000 ms, default 200); size is then the size of each frame. A GIF can take one reference image (to animate it), but no mask. See Image / video APIs, section "Animated GIFs".

xAI Grok#

Kuaishou Kling#

Alibaba Qwen#

gpt-image-2, gpt-image-2.5-flare and gpt-image-2.5-sunburst all accept POST /v1/images/edits for reference-image edits (sketch-to-image works the same way). The upstream ignores size — put the aspect ratio in the prompt; latency is 30–215s.

Since 2026-07-08 the three nano entries are the only way in: older ids (gemini-2.5-flash-image, gemini-3-pro-image-preview, gemini-3.1-flash-image(-preview), tvod-nano-*) keep working as aliases onto the canonical entries above.

Video generation models (38)#

Billed per second as an async task through POST /v1/videos/generations. Where tiers are listed (resolution, or operation for the subtitle SKU) you are billed at the tier you request. See Image / video APIs.

ByteDance Seedance#

Alibaba Wan#

Both Wan 3.0 cards share one capability set: 2-30s, 480P/720P/1080P, 30fps, aspect 16:9 / 4:3 / 1:1 / 3:4 / 9:16 / adaptive; text-to-video, image-to-video (first frame, first+last frame), and image / video / audio references. Prime is the accelerated tier (same quality, markedly faster) at 1.5x the standard rate. The default delivery is 1080P — pass resolution explicitly for a cheaper tier.

⚠️ On rounds that carry a reference video, the input video's seconds are billed too. This is the upstream's own basis: billed duration = input duration + output duration, and the rate above applies to both legs (upstream's price table heads that column "input and output unit price"). Reference videos are [1,15]s each and 15s total; we hold against that ceiling at submit time and refund the difference once upstream reports the real input duration. Reconcile against the top-level input_video_duration_sec in the response (duration_sec still means the delivered clip only, excluding the input). Text-to-video and image-to-video rounds are unaffected (image input is free upstream, and we don't charge for it either). See Media APIs - how usage relates to billing.

Google#

xAI Grok#

Kuaishou Kling#

Kling 4.0 / 4.0 Omni: integer durations of 3-30s, aspect 16:9 / 9:16 / 1:1 / 21:9, native audio in every clip. The official price has not been published yet; these cards bill at Kling 3.0's official with-audio rates until it is, then follow the official price.

MiniMax Hailuo#

The -tx cards run on the same line with the same parameters and capabilities; only the model version differs. minimax-h3-tx is the default version (the former id minimax-h3-1014 still works). Fast (提速版) has faster inference, bills at the official MiniMax-H3-Highspeed price and also offers a 480p tier; Full (满血版) is full precision and bills at the official MiniMax-H3 price. "IR" (带IR) adds an upstream prompt-rewriting step at no extra charge. All five cards bill the same way (reference-video seconds are billed, up to 5 reference images free).

Shengshu Vidu#

Happyhorse#

Digital human / editing#

Music / voice / avatar creation (6)#

These have their own endpoints and billing units. Full request/response shapes: Image / video / music APIs and Digital human / editing APIs.

Cross-vendor protocol support#

Different models support different client SDKs, and the compatibility matrix is authoritative on what is actually reachable. The catalog's supported_protocols field reflects only the native / primary path; many models are additionally reachable on other protocols through gateway translation (the response then carries an X-Protocol-Translation header). Native paths, counted from the registry that generated this page:

• OpenAI Chat (/v1/chat/completions) — all 56 text models
• Anthropic Messages (/v1/messages) — 18 models: claude-fable-5, claude-fable-5-1, claude-haiku-4-5-20251001, claude-haiku-5-5, claude-opus-4-5-20251101, claude-opus-4-6, claude-opus-4-7, claude-opus-4-8, claude-opus-5, claude-opus-5-5, claude-sonnet-4-5-20250929, claude-sonnet-4-6, claude-sonnet-5, claude-sonnet-5-5, glm-5.3-1022, glm-5.3-flash-1022, kimi-k3, kimi-k3-1013
• OpenAI Responses (/v1/responses) — 30 models: claude-haiku-4-5-20251001, claude-opus-4-6, claude-opus-4-7, claude-opus-4-8, claude-sonnet-4-6, claude-sonnet-5, codex-auto-review, glm-5.3-1022, glm-5.3-flash-1022, gpt-5.6-luna, gpt-5.6-luna-bd, gpt-5.6-sol, gpt-5.6-sol-bd, gpt-5.6-terra, gpt-5.6-terra-bd, gpt-6, gpt-6-astra, gpt-6-astra-bd, gpt-6-luna, gpt-6-luna-bd, gpt-6-sol, gpt-6-sol-bd, gpt-6.1-sol, gpt-6.1-sol-bd, gpt-6.1-terra, grok-4.5, grok-4.6, grok-4.7, qwen3.8-max-cjx, xiaozhanAI
• Gemini Native (/v1beta/...) — 6 models: gemini-3.1-pro-low, gemini-3.1-pro-preview-1020, gemini-3.8-flash, gemini-3.8-flash-nx, gemini-4-argon, gemini-4-argon-nx

For a model × protocol combination with no reachable channel, the gateway returns 503 no_channel_available (the model exists but has no channel on that protocol — not a 404); switch to a supported protocol.

Legacy ids (aliases)#

135 older ids still resolve onto the canonical entries above, so integrations pinned to them keep working. An id that is neither in the tables above nor an alias returns model_not_found — the catalog endpoint is the source of truth.

Live data#

The model list and prices track upstream changes. For live data see https://model.dflop.top/models or the public catalog endpoint (no auth required):

curl https://api.dflop.top/api/v1/models/public | jq .

The response looks like { "models": [...] }. Note that this endpoint returns every registered entry — including placeholder SKUs that aren't live yet and non-chat SKUs — so filter on the three fields below when consuming it from a script rather than looping over the whole list:

To take only the models you can send straight to /v1/chat/completions:

curl -s https://api.dflop.top/api/v1/models/public \
  | jq -r '.models[] | select(.callable and (.supported_protocols | index("openai_chat"))) | .id'