Reduzierte KI-Modellpreise: aktuelle Anbieterangebote

Drama 3 is a text-to-speech model from Fish Audio built for directed voice performances. You write the direction in plain language in square brackets before a line, such as [nervous, barely holding it together], and can add sounds and pauses inline or change the delivery of specific words with paired tags like <whisper>.

It supports multi-speaker scenes in one request, switching language mid-sentence across 80+ languages in the same voice, and phoneme tags for pinning pronunciation of names and terms.

Mercury 2.5 is the fastest reasoning LLM, and the latest diffusion LLM (dLLM) from Inception. Instead of generating tokens sequentially, Mercury 2.5 produces and refines multiple tokens in parallel, achieving 1,107 tokens/sec on standard GPUs. It delivers a 10+ point jump in intelligence over Mercury 2, comparable quality to cost-optimized frontier models like GPT-5.6 Luna (Low), Gemini 3.5 Flash-Lite, and Claude Haiku 4.5. Mercury 2.5 supports tunable reasoning levels, parallel tool calls, and schema-aligned JSON output. It's built for production workloads where latency compounds: search agents, voice pipelines, and coding subagents. Read more in the blog post.

Ling 3.1 Flash is a hybrid reasoning mixture-of-experts model from inclusionAI, with 25B active parameters out of 560B total.

Qwen3-235B-A22B-Instruct-2507 is a multilingual, instruction-tuned mixture-of-experts language model based on the Qwen3-235B architecture, with 22B active parameters per forward pass. It is optimized for general-purpose text generation, including instruction following, logical reasoning, math, code, and tool usage. The model supports a native 262K context length and does not implement "thinking mode" (<think> blocks).

Compared to its base variant, this version delivers significant gains in knowledge coverage, long-context reasoning, coding benchmarks, and alignment with open-ended tasks. It is particularly strong on multilingual understanding, math reasoning (e.g., AIME, HMMT), and alignment evaluations like Arena-Hard and WritingBench.

Ling-3.0-flash is a 124B-parameter Mixture-of-Experts (MoE) model, with approximately 5.1B parameters activated per token.

The model is designed with token efficiency and production-scale agentic inference as key priorities, enabling developers to complete more useful work within constrained token, latency, and serving-cost budgets.

LongCat 2.0 is a sparse mixture-of-experts language model from Meituan, with 48B active parameters out of 1.6T total. It is suited for coding, repository-level changes, long-horizon problem solving, and agentic workflows.

DeepSeek V4 Flash Vision Exp is an experimental vision-enabled version of DeepSeek V4 Flash 0731 from DeepSeek, adding image understanding while matching the base model on text capabilities including agents, reasoning, and world knowledge. It is a sparse mixture-of-experts model with 13B active parameters out of 284B total.

It is suited for document and chart understanding, visual question answering, and multimodal agent workflows that interleave text and images.

This model always redirects to the latest model in the Gemini Flash family.

This model always redirects to the latest model in the Gemini Pro family.

Gemini 3.7 Flash is a multimodal model from Google for fast agentic workflows, coding, and complex multi-step reasoning. It is designed for tasks that require responsive performance and reliable multi-step problem solving.

Gemini 3.7 Flash is a multimodal model from Google for fast agentic workflows, coding, and complex multi-step reasoning. It is designed for tasks that require responsive performance and reliable multi-step problem solving.

Gemini 3.8 Flash is Google's most intelligent Flash model with significant gains from 3.7 Flash across software engineering, agentic tasks, and multi-step reasoning.

Gemini 3.8 Flash is Google's most intelligent Flash model with significant gains from 3.7 Flash across software engineering, agentic tasks, and multi-step reasoning.

Gemini 4 Argon is Google's flagship reasoning model, built for complex software engineering, deep mathematical and scientific problem solving, and autonomous multi-step agents. Google trained it to keep context across long, multi-step tasks and to turn long documents mixing text, tables, charts and images into a single grounded output. It leads the Vals Index across finance, legal, coding and tax work, and supports long decode of up to 1M output tokens in a single response.

Gemini 4 Argon is Google's flagship reasoning model, built for complex software engineering, deep mathematical and scientific problem solving, and autonomous multi-step agents. Google trained it to keep context across long, multi-step tasks and to turn long documents mixing text, tables, charts and images into a single grounded output. It leads the Vals Index across finance, legal, coding and tax work, and supports long decode of up to 1M output tokens in a single response.

HeyGen Video is a general-purpose video generation model from HeyGen. It renders short clips with synthesized audio (dialogue, ambience, and sound effects) in a single call, working from a text prompt, from a first-frame image, or from a set of image, video, and audio references that the prompt can address directly. It is aimed at short-form business video with a lot of motion in frame: brand moments, product and equipment demos, and training or onboarding content.

It is not an avatar product. Unlike HeyGen's Avatar IV and Video Agent, there is no talking-head, script, or lip-sync step; the prompt and references drive the whole scene.

HeyGen Voice is a text-to-speech model from HeyGen. It comes with 10 default voices, or bring your own HeyGen key to use voice clones from your own account. It streams 44.1 kHz mono PCM audio, with support for 39 languages and up to 5,000 characters per request.

HeyGen Voice Turbo is a lower-latency version of HeyGen Voice, HeyGen's text-to-speech model. It has the same 10 default voices, or bring your own HeyGen key to use voice clones from your own account. It streams 44.1 kHz mono PCM audio, with support for 39 languages and up to 5,000 characters per request.

Mercury 3 is a diffusion-based language model from Inception, and the lab's most capable model to date. It is suited for agentic workflows and real-time applications where latency compounds, such as search agents, voice pipelines, and coding subagents. Its distinguishing trait is speed: more than 40x larger than the first Mercury model while generating at the same throughput, because it produces and refines tokens in parallel rather than one at a time.

Mercury Decide is Inception's structured decision model, served as a System One endpoint. Send a state along with typed questions, and it returns a choice, a score, or a yes/no answer, each with a calibrated probability taken directly from the model rather than written out as text, so output tokens are free. It makes up to 14 decisions per second and reports how certain it is, so a decision system can run it on every case and escalate the unsure ones to a human. Mercury Decide uses the same /v1/systemone schema as Jev.

Krea K3 Edit is an image-to-image editing model from Krea. It takes one to eight reference images plus a text prompt and returns a single edited image, following instructions that range from small local changes to complex multi-step edits while keeping the subjects and composition of the references intact. The output follows the shape of the input unless an aspect ratio is requested.

The default fast mode completes most edits in under ten seconds, which suits interactive editing tools and workflows that generate many variations of the same source image.

Drex Max is the largest decision model from Nace.AI, the successor to Drex v2.0. Given a state (text, JSON, or up to four images) and typed questions (yes/no, multiple choice, or ordinal score), it returns a calibrated probability for every option in a single forward pass, with no text generation. It is TypeSafe System One compatible and suits routing, classification, and policy checks that need a scored answer instead of prose.

Solar Decide is Upstage's structured decision model, served as a System One endpoint on Solar Mini 4. Send a state along with typed questions, and it returns a choice, a score, or a yes/no answer, each with a calibrated probability taken directly from the model rather than written out as text. Because it generates no prose, each decision takes a single forward pass and output tokens are free. With a 512K context window, an entire document can serve as the state. Solar Decide uses the same /v1/systemone schema as Jev, bringing Solar Mini 4's strong Korean understanding to routing, classification, and policy checks.

Solar Decide Flash is Upstage's low-latency structured decision model, a faster variant of Solar Decide built on Solar Mini 4 and served through the System One (/v1/systemone) API. Instead of generating text, it reads a state and answers typed questions, returning a choice, a score, or a yes/no answer, each with a probability taken directly from the model. Each decision is a single forward pass, and output tokens are free.

It is tuned for consistently fast response times in routing, classification, and policy checks. It keeps the 512K context window, so a full document can serve as the state, and carries Solar Mini 4's Korean-language strength.

Solar Mini 4 is Upstage's compact, cost-efficient language model, a 35B-parameter mixture-of-experts with 3B active parameters and a 524K context window. It is built for agentic use cases where response speed and cost matter, with fluent Korean alongside strong English and Japanese, and support for long-context workloads.

When using BYOK with ZDR enforcement, the Console API key must belong to a ZDR-enabled Upstage organization. Contact Upstage to enable ZDR on your org.

Solar Pro 4 is Upstage's cost-efficient large language model, featuring a 524K context window. It is built for long-horizon tasks and agentic workflows, with strong capabilities in office productivity, document-intensive work, and coding.

When using BYOK with ZDR enforcement, the Console API key must belong to a ZDR-enabled Upstage organization. Contact Upstage to enable ZDR on your org.

GLM-5 is Z.ai’s flagship open-source foundation model engineered for complex systems design and long-horizon agent workflows. Built for expert developers, it delivers production-grade performance on large-scale programming tasks, rivaling leading closed-source models. With advanced agentic planning, deep backend reasoning, and iterative self-correction, GLM-5 moves beyond code generation to full-system construction and autonomous execution.

GLM-5.3 is a large-scale reasoning model from Z.ai, built for complex software engineering and long-horizon agent tasks. It supports text input and output with a 1M-token context window, and improves on GLM-5.2 in coding and in the balance between performance and token efficiency.

Reasoning is always on and cannot be disabled. Reasoning efforts low, high, and max are supported; max is the default.

MiniMax-M2.7 is a next-generation large language model designed for autonomous, real-world productivity and continuous improvement. Built to actively participate in its own evolution, M2.7 integrates advanced agentic capabilities through multi-agent collaboration, enabling it to plan, execute, and refine complex tasks across dynamic environments.

Trained for production-grade performance, M2.7 handles workflows such as live debugging, root cause analysis, financial modeling, and full document generation across Word, Excel, and PowerPoint. It delivers strong results on benchmarks including 56.2% on SWE-Pro and 57.0% on Terminal Bench 2, while achieving a 1495 ELO on GDPval-AA, setting a new standard for multi-agent systems operating in real-world digital workflows.

MiMo-V2.5-Pro is Xiaomi’s flagship model, delivering strong performance in general agentic capabilities, complex software engineering, and long-horizon tasks, with top rankings on benchmarks such as ClawEval, GDPVal, and SWE-bench Pro. It can independently and autonomously complete professional tasks that would take human experts days or weeks, involving more than a thousand tool calls. Its context length of up to 1M makes it well suited for integration with a wide range of agent frameworks.

DeepSeek-V3.2 is a large language model designed to harmonize high computational efficiency with strong reasoning and agentic tool-use performance. It introduces DeepSeek Sparse Attention (DSA), a fine-grained sparse attention mechanism that reduces training and inference cost while preserving quality in long-context scenarios. A scalable reinforcement learning post-training framework further improves reasoning, with reported performance in the GPT-5 class, and the model has demonstrated gold-medal results on the 2025 IMO and IOI. V3.2 also uses a large-scale agentic task synthesis pipeline to better integrate reasoning into tool-use settings, boosting compliance and generalization in interactive environments.

Users can control the reasoning behaviour with the reasoning enabled boolean. Learn more in our docs

A 12B parameter model with a 128k token context length built by Mistral in collaboration with NVIDIA.

The model is multilingual, supporting English, French, German, Spanish, Italian, Portuguese, Chinese, Japanese, Korean, Arabic, and Hindi.

It supports function calling and is released under the Apache 2.0 license.

Clef is Cloudflare's open-source 27B multimodal decision model, a fine-tune of Qwen3.8-27B served on Workers AI. It turns a state (text or structured JSON) plus a schema of typed questions into decisions, returning a calibrated probability for every allowed option of every question in a single forward pass instead of generating tokens. Use it for classification, routing, scoring, guardrails, and agentic control flow through the Decisions API. Note: Workers AI currently truncates long text state to roughly the first 2K tokens, so content beyond that is not read; images are counted separately.

Classify-1 is a structured decision model from ScaleDown, served through OpenRouter's Decisions API. It scores text or a document against caller-defined choice, yes/no, and ordered-scale questions and returns calibrated probabilities rather than generated prose, making it suited for routing, moderation, intent detection, and domain tagging inside an application. It is request-compatible with TypeSafe's Jev Decisions API.

Learn more in ScaleDown's docs: https://docs.scaledown.ai/api-reference/decisions-overview

Extract is a description-driven entity extraction model from ScaleDown. Instead of a fixed set of entity types, callers describe the fields they want in plain English through provider.options.scaledown.entities, and the model returns each matching span with a confidence score, character offsets, and surrounding context. It is suited for pulling structured values out of contracts, invoices, emails, and transcripts, and includes built-in OCR so scanned images and PDFs work without a separate step.

Learn more in ScaleDown's docs: https://docs.scaledown.ai/api-reference/extract-overview

Summarize is an abstractive summarization model from ScaleDown. It condenses long text and documents into a fluent, faithful rewrite rather than lifting sentences from the source, and is suited for document review, transcript digests, and RAG pre-processing where retrieved content must be shortened before reaching another model. Format, focus, language, and length are steered through the system prompt, while sampling is fixed to a configuration tuned for faithful output.

Learn more in ScaleDown's docs: https://docs.scaledown.ai/api-reference/summarize-overview

MiMo-V2.6-Pro is the flagship foundation model developed by Xiaomi. Built at a scale of over 1T parameters, it is designed to push the ceiling of capability for the most demanding workloads. The model features a 1M-token context window and native multimodal capabilities. Optimized for agentic workflows, it delivers top-tier performance across coding, visual, general, and research scenarios, excelling at complex, long-horizon tasks with robust generalization across a diverse range of agent harnesses.

Wan 3.0 is a video generation model from Alibaba for text-to-video, image-to-video, and reference-guided video generation. It produces 480p, 720p, or 1080p video with durations from 2 to 30 seconds.

MiniMax-M2 is a compact, high-efficiency large language model optimized for end-to-end coding and agentic workflows. With 10 billion activated parameters (230 billion total), it delivers near-frontier intelligence across general reasoning, tool use, and multi-step task execution while maintaining low latency and deployment efficiency.

The model excels in code generation, multi-file editing, compile-run-fix loops, and test-validated repair, showing strong results on SWE-Bench Verified, Multi-SWE-Bench, and Terminal-Bench. It also performs competitively in agentic evaluations such as BrowseComp and GAIA, effectively handling long-horizon planning, retrieval, and recovery from execution errors.

Benchmarked by Artificial Analysis, MiniMax-M2 ranks among the top open-source models for composite intelligence, spanning mathematics, science, and instruction-following. Its small activation footprint enables fast inference, high concurrency, and improved unit economics, making it well-suited for large-scale agents, developer assistants, and reasoning-driven applications that require responsiveness and cost efficiency.

To avoid degrading this model's performance, MiniMax highly recommends preserving reasoning between turns. Learn more about using reasoning_details to pass back reasoning in our docs.

MiMo-V2.5 is a native omnimodal model by Xiaomi. It delivers Pro-level agentic performance at roughly half the inference cost, while surpassing MiMo-V2-Omni in multimodal perception across image and video understanding tasks. Its 1M context window supports complete documents, extended conversations, and complex task contexts in a single pass, making it ideal for integration with agent frameworks where strong reasoning, rich perception, and cost efficiency all matter.

DeepSeek-V3 is the latest model from the DeepSeek team, building upon the instruction following and coding abilities of the previous versions. Pre-trained on nearly 15 trillion tokens, the reported evaluations reveal that the model outperforms other open-source models and rivals leading closed-source models.

For model details, please visit the DeepSeek-V3 repo for more information, or see the launch announcement.