AI models
Compare text, image, video, decision and robotics models. Read capability and license guides, check API access, and estimate running costs with a calculator.
Models found: 76
Text
Compare models in this category
Choose 2–3 models. Fields come from dated catalog records; different API units stay separate.
0 of 3 selected
Select at least two models to see their available facts side by side.
- TextOpen source
DeepSeek-V4.1-Flash
DeepSeek-V4.1-Flash reads text and images and responds in text, with a 1M-token context window and open weights. Start with hosted deepseek-flash or a compatible client, then compare the whole task’s token and tool bill with the memory, runtime and hardware needed to self-host an exact checkpoint.
- TextOpen source
Qwen3.8-27B
Qwen3.8-27B is a dense open-weight model for text, image and video understanding, with text output and an Apache 2.0 license. It offers a simpler local starting point than the much larger Flash-Next. A quantized download can fit a desktop GPU, while the conversation cache and runtime still need extra space. Choose the format and context together, then measure the time to a correct answer on your own task.
- TextOpen source
Gemma 4 12B it
Gemma 4 12B is Google’s dense instruction-tuned model for text output from text and supported visual or audio inputs. Its approximately 11.95B parameters are available under Apache 2.0. Google’s QAT Q4_0 GGUF offers a smaller local download; the memory used while running also depends on context, cache settings and the runtime.
- TextOpen weights
Qwen3.8-Flash-Next
Qwen3.8-Flash-Next brings a large mixture-of-experts model to local coding and visual tasks. Its 125B main model activates about 6B parameters per token, with another 51B in n-gram embeddings and 4B for multi-token prediction. On a PC, Strata can divide that workload between GPU memory, system RAM, the CPU and SSD lookup. The useful choice is which format completes your tasks within your memory and waiting-time budget.
- TextHosted API
Claude Opus 5.5
Claude Opus 5.5 is being used for long-running coding agents and, more visibly this week, code-drawn animations and product videos. It accepts text and images and returns text; a browser or rendering tool turns the code into moving images. Its Claude API rates are $4 per million input tokens and $20 per million output tokens, twice Sonnet 5.5’s. The choice depends on the accepted result, the revisions needed, and the total task cost.
- TextHosted API
Claude Sonnet 5.5
Claude Sonnet 5.5 is being used for routine coding, bug fixes, documents, and code-drawn animation. Its standard Claude API rates are $2 per million input tokens and $10 per million output tokens, half Opus 5.5’s uncached rates. Early creator reports show that a cheaper model can also build visual work, but the useful measure is the time, tool calls, and revisions needed for an accepted result.
- TextHosted API
GPT-6.1 Sol
GPT-6.1 Sol is OpenAI’s model for complex coding, research and computer use. Its standard short-context API rates are $2 per million input tokens and $10 per million output tokens. Compare it with GPT-6 Astra on task cost, reasoning effort and time to a usable result. The reports below cover repository repairs, visual builds and the tradeoffs users found.
- TextHosted API
Gemini 4 Argon
Gemini 4 Argon is Google’s model for long software-engineering, enterprise and research tasks. Since October 13, developers can use its paid API and subscribers can access it through supported Google products. Introductory API input/output rates are $2/$10 per million tokens through November 30, then $4/$20 from December 1. Independent evaluations show strong task results alongside substantial output-token use, which makes the full task bill worth comparing.
- TextOpen weights
Ling 3 Flash
Ling-3.0-flash is inclusionAI’s open-weight text model with 124B total parameters, 5.1B active per token and a 256K context window. Its low hosted token prices make it worth trying for coding and tool use. Community results point to a practical starting place: a bounded task with an external completion check. Local deployment is a separate choice, with quantized weights around 70GB in several reported setups.
- TextOpen weights
Kolibri 1
Kolibri 1 is Aleph Alpha’s Apache-2.0 text model for German and English. Its mixture-of-experts architecture has 78.1B total parameters and about 3.46B active per token. Consider it for document questions, drafting and extraction on infrastructure you control. The official FP8 weights occupy about 78GB, so the first deployment decision is GPU capacity; the next is how much reasoning time each accepted task needs.
- TextOpen weights
Reflection Beam
Reflection Beam 1.0 is a text model for coding, reasoning and agents. Its released Apache 2.0 checkpoints contain 501.3B total parameters, with 23.4B active per token, in BF16, FP8 and NVFP4 formats. Choose the hosted Reflection API for a first task or size a deployment around the full checkpoint and your intended context length.
- TextHosted API
Ling 3.1 Flash
Ling 3.1 Flash is inclusionAI’s text model for reasoning, coding and tool-driven work. Start with a bounded code or research task through Vercel AI Gateway, OpenRouter or OpenCode Zen. Vercel’s free offer runs through October 13, 2026: its regular model ID starts billing afterward, while its separate -free ID stops serving. For a trial with no automatic switch to paid inference, choose that -free ID and compare the accepted result with a paid alternative before moving ongoing work.
- TextHosted API
Mistral Large 4
Mistral Large 4, nicknamed “Le Chonk,” is a multimodal model for document questions, structured extraction and tool-driven work. The practical starting point is its public-preview API: launch pricing is $0.68 per million input tokens and $2.09 per million output tokens. Its 1.05 trillion total parameters make self-hosting a separate infrastructure decision, even though each token activates a much smaller part of the model.
- TextHosted API
Claude Haiku 5.5
Claude Haiku 5.5 is Anthropic’s fast hosted model for routing, extraction and coding subagents. Its low starting price makes frequent calls practical, but agent benchmarks show why effort, context length and the number of attempts belong in the budget. It accepts text and images, supports a one-million-token context, and generates up to 128K tokens per standard request.
- TextHosted API
MAI-Code-1.1-Flash
MAI-Code-1.1-Flash is Microsoft’s coding model for repository questions, refactoring and tool-driven development, including image-to-code work. The cloud model is available in GitHub Copilot. Microsoft’s October update also introduces a local build, with a much larger device-memory budget than its lightweight coding category might suggest.
- TextHosted API
GPT-6 Luna
GPT-6 Luna is OpenAI’s efficient general-purpose model for focused, high-volume work: document extraction, summaries, grounded replies and bounded coding steps. Standard API rates start at $0.10 per million input tokens and $0.50 per million output tokens. Longer prompts, cache writes and repeated agent turns change what it costs to finish a task.
- TextOpen weights
Step 5 Preview
Step 5 Preview is StepFun’s reasoning and coding model, available through the API and downloadable weights. Its hosted API accepts text, images and video and returns text, with a 1M-token input context and up to 64k output tokens. International and China accounts have separate token prices.
- TextHosted API
Claude Mythos 5.1
Claude Mythos 5.1 is the verified-access version of Claude Fable 5.1, with the same underlying model and more permissive safeguards for approved cybersecurity and life-sciences work. It accepts text and images and produces text, with a 1M-token context window and up to 128K output tokens. The Claude API lists $10 per million input tokens and $50 per million output tokens; use it through an approved verification grant.
- TextHosted API
Muse Wonder
Muse Wonder is Meta’s hosted model for coding, reasoning and agents that work across many steps. It reads text, images, audio, video and PDFs and returns text. Use muse-wonder-1.0 through Meta Model API at $4 input and $20 output per million tokens, or use it in Muse Code. Muse Wonder Full is a separate subscription variant in Muse Code.
- TextOpen weights
MiMo V2.6
MiMo V2.6 is Xiaomi’s multimodal Pro and Flash family for coding, research and tool workflows. The September 22 release offers hosted APIs and downloadable checkpoints. Flash starts at $0.14 per million uncached input tokens and $0.28 per million output tokens on the overseas API; Pro’s ordinary overseas input/output rates are $0.435/$0.87 per million tokens; Xiaomi recommends Pro for complex long-running work and Flash for frequent calls. Choose the billing route before comparing a complete task.
- TextOpen weights
Youtu-Parsing-Omni
Youtu-Parsing-Omni is Tencent’s downloadable 5B model for turning documents, images and audio/video into structured JSON. It fits workflows that need reading order, tables, locations or timestamps as well as recognized content. Its custom license excludes European Union use. Choose the official task recipe and measure the whole pipeline before renting a GPU; Tencent’s published model materials give no hosted API tariff or verified minimum VRAM requirement.
- TextOpen weights
LFM2.5-24B-A2B
LFM2.5-24B-A2B is Liquid AI’s downloadable reasoning model for text and tool agents. It has 24B total parameters and activates about 2B per token. Use the Liquid API to start without managing weights, or serve the model locally when keeping the agent’s data on your hardware matters. The current API tariff is $0.05 per million input tokens, $0.025 for cached input and $0.70 for output.
- TextOpen weights
Muse Spark Open
Muse Spark Open is Meta’s open-weight model for coding, multimodal reasoning and tool agents, released October 15, 2026. It has 816B total parameters, 42B active per token and a 1M-token model context. Hosted inference is the practical first step for most users; self-hosting needs substantial GPU or unified memory, with hardware requirements depending on the quantization and runtime.
- TextOpen weights
Nemotron 3.5
NVIDIA Nemotron 3.5 Super VL is an open-weight vision-language model for coding, tools and document or video understanding. The 120B model activates 12B parameters per token and supports up to 1M context. Bolt is a separate 74B/A9B model with up to 256K context and compact local formats; its 24GB variant name does not describe Super VL’s hardware requirement. The newer text-only Bolt E4B is a separate 8B-total / 4B-effective checkpoint.
- TextOpen weights
MiniCPM-V 4.7
MiniCPM-V 4.7 is OpenBMB’s open-weight vision-language family for images, documents and video understanding, with text responses. Choose the compact 1B model or the 35B-A3B mixture of experts. The much-publicized phone memory figure belongs to an optimized LES deployment, so choose the model artifact and runner before budgeting a device.
- TextRestricted preview
Elorian Glimpse
Glimpse is Elorian’s research-preview visual reasoning model. It examines images and video, revisits visual evidence and returns answers for tasks such as chart analysis, counting and event understanding. The announced hosted input/output rates are $0.60/$3.00 per million tokens. Start through Elorian’s platform and budget the complete reasoning task.
- TextOpen weights
Mellum 2.1 Thinking
Mellum 2.1 Thinking is JetBrains’ open-weight text model for coding agents, tool use and reasoning. Its 12B-parameter mixture of experts activates 2.5B parameters per token. Official GGUF files are available for local runners alongside the BF16 checkpoint. Start with a small code task, then choose a quantization and context that leave room for the runtime.
- TextOpen weights
Underdog Saluki 27B
Underdog Saluki 27B is a compressed Qwen3.8-27B model for local text tasks and tool calling. Its recommended IQ2-mix GGUF download is 7.89GB. That smaller file can make a local trial easier, while context, runtime buffers and optional vision still need extra memory. Start with text, then check the quality and complete task time on your own workload.
- TextOpen weights
MiniMax M3.1 Flash
MiniMax M3.1 Flash is a multimodal coding and agent model with a one-million-token context window. Use the MiniMax API for your application, MiniMax Code for an existing coding workflow, or the published weights with a compatible serving recipe. Thinking stays enabled, with an effort setting that changes how much reasoning the model performs. For a budget, count the full sequence of model calls, tool results and retries needed to finish the task.
- TextHosted API
Mercury 3
Mercury 3 is Inception’s hosted diffusion language model for text applications. It supports a one-million-token context window, tool calling and structured outputs through a chat-completions API. Its launch discount lowers token rates, while the time and cost of an agent task still depend on context, reasoning and external tools.
- TextOpen weights
Flint
Flint is Recursive’s Apache 2.0 language model for coding agents: investigating repositories, fixing bugs, implementing changes and writing tests. It has 39.8B total parameters with about 2.5B active per token. The official rec engine offers a 4-bit Apple Silicon path with at least 48GB unified memory and about 30GB free disk space; the original BF16 checkpoint is a separate, roughly 79.6GB download.
- TextOpen weights
Raon-1 mini preview
Raon-1 mini preview is KRAFTON AI’s Apache 2.0 text reasoning and tool-calling model, with Korean conversation among its stated focuses. The card calls it 30B-A4B; its detailed table lists 28.8B total and 4.3B activated parameters. A documented NVFP4 and DFlash2 recipe targets one RTX 5090 32GB at a 32,768-token context. Choose that exact configuration before budgeting hardware.
- TextHosted API
Kimi K3.1 Pro
Kimi K3.1 Pro is Moonshot AI’s 3T-class model for coding and longer tool-driven work, with a 1,048,576-token context window. Pro and Pro Highspeed are available through Kimi applications and the API. Highspeed doubles the fresh-input and output rates; both use the same cache-read rate. Full weights are promised by November 4, 2026, so today’s deployment choice is hosted access.
- TextOpen weights
Nemotron 3.5 Bolt E4B
NVIDIA Nemotron 3.5 Bolt E4B is a downloadable text model for conversational interfaces and tool use. NVIDIA lists 8B total parameters, 4B effective and up to 64K context, with a Q4_K_M GGUF release. The “4GB” label does not include the full download or runtime memory: the main GGUF alone is about 6.18 GB, before cache, buffers and the optional drafter.
- TextOpen weights
Ornith-1.5-122B
Ornith-1.5-122B is an MIT-licensed reasoning and text-generation model with a Qwen3.5 MoE architecture. It supports tool calls and a native 262,144-token context. Start by choosing a checkpoint format and a serving runtime: the public BF16 weights occupy about 250GB, before runtime memory and request state.
- TextOpen weights
Ling-3.0-flash-Sante
Ling-3.0-flash-Sante is a text-generation LLM built on Ling-3.0-flash, specialized for healthcare research. Its hybrid MoE architecture has 124B total parameters and activates 5.1B per token. The weights are available under MIT.
- TextHosted API
GPT-6.1 Terra
GPT-6.1 Terra is OpenAI’s model for visual reasoning, computer use and everyday writing. It reads text and images and produces text or tool calls. Use it in ChatGPT Work or Codex, or connect the API to an application and an execution environment. Standard API rates start at $1 per million input tokens and $5 per million output tokens; a completed task can include many model turns, screenshots and retries.
Image
Compare models in this category
Choose 2–3 models. Fields come from dated catalog records; different API units stay separate.
0 of 3 selected
Select at least two models to see their available facts side by side.
- ImageOpen weights
Qwen-Image-2.1
Qwen-Image-2.1 generates and edits images with one set of downloadable weights, including transparent PNGs and edits with several reference images. It is worth evaluating for local creative workflows if you can manage GPU memory and setup. The weights use the Qwen Research License; commercial use needs a separate license.
- ImageOpen source
Ming-Image-0.1-Design
Ming-Image-0.1-Design is inclusionAI’s open-weight image model for posters, interfaces, dashboards and infographics. Its MIT-licensed release produces complete compositions and can save transparent RGBA images. The 6B label describes the image transformer; deployment also loads an encoder, connector and VAE. Choose the complete pipeline and a measured workload when budgeting GPU time.
- ImageHosted API
Nano Banana 2.1
Nano Banana 2.1 generates and edits images through Google’s Gemini API. Standard image output costs $0.0336 at 1K, $0.0504 at 2K and $0.1134 at 4K, plus input, text/thinking and any billable search. Use it for product variations, text-heavy designs and edits that keep a subject consistent across turns. The stable API model is gemini-nano-banana-2.1.
- ImageHosted API
Seedream 5 Flash
Seedream 5 Flash is ByteDance’s hosted image-generation and editing model for quick visual iterations. It supports 1K, 1.5K and 2K images, up to ten reference images and layered output. BytePlus lists $0.018 per output image; Volcengine lists ¥0.12. Choose the model ID and billing route for your account before estimating a batch.
- ImageOpen weights
FLUX 3 Image Klein
FLUX 3 Image Klein is Black Forest Labs’ open-weight image generator and editor, released October 15, 2026. Use its hosted API for a quick start at $0.009 per 1K image, or choose a local ComfyUI or Diffusers workflow. The 5.6B figure counts the image transformer; the Qwen3-VL-4B text encoder and VAE are additional components.
- ImageOpen weights
MOSS-Image
MOSS-Image offers Base and SFT checkpoints for text rendering and controllable layouts. Try them through MOSI’s limited-time zero-credit API, or budget the complete local pipeline and every attempt needed to produce an accepted image.
- ImageOpen weights
Iris-3B and Turbo
Iris-3B generates images directly in pixel space with a 3B diffusion transformer and a separate frozen 4B Qwen3-VL text encoder. Choose Base for adjustable sampling, or the official four-step Turbo checkpoint for a fixed fast sampling path. Both are downloadable under Apache 2.0; plan memory for the complete pipeline rather than just the image-model file.
- ImageOpen weights
Krea 2 RAW / Turbo
Krea 2 is a text-to-image family built around a 12B diffusion transformer. RAW is the undistilled checkpoint for adaptation and LoRA training; Turbo is the eight-step option for generation. Both have downloadable weights under the Krea 2 Community License, which limits commercial use by company revenue. Choose the checkpoint, complete runtime and access route before estimating an image budget.
- ImageHosted API
Krea 3
Krea 3 is a hosted image editing model for changing product scenes, materials and campaign variants. It accepts up to six reference images per edit and supports native 2K output. Start with an existing image and a specific change.
- ImageOpen weights
Fibo 2.0
Fibo 2.0 is Bria’s image generation and editing family. Dev handles both tasks with adjustable sampling; Edit Turbo is a four-step editing checkpoint. Use the hosted services for a first trial, or download the weights for a local non-commercial workflow. Commercial self-hosting requires an agreement with Bria.
- ImageHosted API
Ideogram Vector 1.0
Ideogram Vector 1.0 creates native SVG artwork from text and converts existing images into vector shapes. It is useful for logos, icons and illustrations you need to resize or recolor in a vector editor. Access is hosted through Ideogram and API providers; there are no downloadable model weights for a local GPU workflow.
Video
Compare models in this category
Choose 2–3 models. Fields come from dated catalog records; different API units stay separate.
0 of 3 selected
Select at least two models to see their available facts side by side.
- VideoOpen weights
MiniMax H3
MiniMax H3 generates 4–15-second videos with stereo audio from text and reference materials. Compare hosted API and local H3-Base costs before choosing a workflow.
- VideoHosted API
Kling 4.0
Kling 4.0 is available as kling_4 in Runway’s developer model catalog for 720p or 1080p video lasting 3–30 seconds. Runway API billing is separate from Kling AI subscriptions. The earlier Flash preview and its creator tests below describe a different version and access route.
- VideoOpen weights
Kandinsky 6.0 Video
Kandinsky 6.0 generates five-second videos with synchronized sound from a prompt or reference image. Its downloadable Lite 3B and Pro 29B models suit creators who want to run their own video workflow. Start with a distilled checkpoint and a device preset; the GPU, offloading and number of attempts determine the practical cost.
- VideoOpen weights
LTX 2.5
LTX 2.5 is Lightricks’ current recommended audio-video model, available through an API and downloadable weights. Use the API for a first production test, or the Python and ComfyUI workflows for local control. Fast and Pro offer different prices; LTX 2.3 remains useful for cheaper generation and its hosted editing endpoints.
- VideoOpen weights
EdenLive
EdenLive renders a reference video into a new visual style using a target first frame and a text prompt. It offers downloadable 5B and 14B models, including four-step streaming variants. For interactive scenes, an engine or application maintains the world state while EdenLive renders the frames. The published RTX 5090 path uses a specific quantized LightX2V configuration.
Structured decisions
Compare models in this category
Choose 2–3 models. Fields come from dated catalog records; different API units stay separate.
0 of 3 selected
Select at least two models to see their available facts side by side.
- Structured decisionsHosted API
Jev 1.13
Jev 1.13 is a TypeSafe AI decision model for turning text into categories, ratings and yes/no probabilities. It fits workflows such as ticket routing, content review and relevance filtering. You use it through a hosted API; its fixed model ID is jev-1.13.0.
- Structured decisionsOpen source
Laya
Laya is an Apache-2.0 open-weight decision-model family from Convai Innovations. It turns text or structured state into choices, scores and yes/no probabilities in a forward pass, instead of generating a chat response. Use it for repeated judgments such as ticket routing or document triage when you want to run inference in your own environment. This guide covers checkpoint selection, local and hosted trials, self-hosting costs and the tradeoffs against TypeSafe Jev.
- Structured decisionsOpen source
CLM-8B
CLM-8B scores candidate actions against a state. Its open-weight reference release, Contrastive-LM/CLM-v0.1-8B, combines a frozen Qwen3-8B encoder with two projection heads under Apache-2.0. It is useful when your application can describe the choices in advance: routing a ticket, selecting a tool, or ranking proposed answers. Reusing candidate embeddings can reduce repeated work; the cost depends on the state, candidate set, cache misses and serving configuration.
- Structured decisionsOpen source
Cloudflare Clef
Clef is Cloudflare Workers AI’s 27B multimodal decision model, with Clef-flash as a 9B low-latency variant. Instead of generating free-form text, both take a state plus typed questions and return probabilities for the allowed answers in one forward pass. Cloudflare hosts them on Workers AI and publishes both sets of weights under Apache 2.0. Clef-omni adds audio and video inputs on a 30B mixture-of-experts backbone at $0.15 per million input tokens. This guide separates hosted token cost from self-hosting cost, compares the two variants with Jev, and shows what to measure before placing a decision model in an agent hot path.
- Structured decisionsHosted API
Liquid d1
Liquid d1 turns text and images into typed decisions with probabilities for routing, classification and rubric scoring. Its paid native d1 route costs $0.02 per million input tokens; d1-ultra costs $0.042. Both have free output and bill each question’s text and supplied images. Choose a hosted route for a first task, or the separate d1-3B and experimental d1-omni-600M checkpoints for local inference.
- Structured decisionsHosted API
OpenAI Decisions API
OpenAI Decisions API turns text, images and JSON into a probability, a category or a rubric score. Use it when your application knows the possible answers: routing a ticket, sorting a queue or flagging a photo for review. It launched in public beta on October 6, 2026 and became generally available on October 22. It uses gpt-6-luna through /v1/decisions, with input-only billing.
- Structured decisionsOpen weights
Perplexity Decider v1.1
Perplexity Decider turns text and images into classification, routing and scoring probabilities. The current native API offers v1.2 at $0.02 per million input tokens and Flash v1.2 at $0.01, with free output. Related questions can share state processing within a billing group. The historical v1.1 Apache 2.0 checkpoint remains a separate self-hosting option.
- Structured decisionsHosted API
Microsoft-Decision-1
Microsoft-Decision-1 scores a fixed set of answers for classification, routing and workflow control. Microsoft post-trained Qwen3.5-9B for single-pass decision scoring. It is available in Microsoft Foundry and listed on OpenRouter through Azure at $0.042 per million input tokens, with free output. Define the allowed actions, then use probabilities to decide whether to act or request review.
- Structured decisionsOpen weights
Decision 3.0
Decision 3.0 is vLLM Semantic Router’s family of open-weight models for bounded decisions. Supply text or JSON, optionally images or video, and define the allowed answers. The model returns probabilities for those answers, so an application can route a request, score a case or ask for review.
- Structured decisionsHosted API
SubQ Snap
SubQ Snap is Subquadratic’s hosted decision model for choosing among answers your application defines. It reads text and images, returns typed answers with probabilities, and supports a 384K-token API window. The public preview costs $0.05 per million input tokens, with free output. Capacity is limited, so new users can be waitlisted.
Embedding / retrieval
Compare models in this category
Choose 2–3 models. Fields come from dated catalog records; different API units stay separate.
0 of 3 selected
Select at least two models to see their available facts side by side.
- Embedding / retrievalOpen weights
EmbeddingGemma 2
EmbeddingGemma 2 turns text, code, images, video and audio into search vectors in a shared space. Use it to retrieve evidence for a RAG answer or find related material across formats. The full 740M-parameter model is downloadable under Apache 2.0; text-only loading uses 270M parameters. A separate generative model writes the answer from the retrieved evidence.
- Embedding / retrievalOpen weights
pplx-embed-v2-late
pplx-embed-v2-late retrieves text, images and visual document pages using a vector for each token. Perplexity publishes MIT-licensed 0.6B and 9B checkpoints. The smaller model can query an index built by the larger one, letting you spend more compute while indexing and keep query encoding lighter.
Speech & audio
Compare models in this category
Choose 2–3 models. Fields come from dated catalog records; different API units stay separate.
0 of 3 selected
Select at least two models to see their available facts side by side.
- Speech & audioHosted API
HeyGen Voice
HeyGen Voice turns scripts into speech using a cloned voice. Choose Instant for a short recording or Professional for a longer training set, then use the active voice through its speech API.
- Speech & audioOpen weights
YuE2
YuE2 generates songs from lyrics and a style prompt, using an editable melody-and-chord score before rendering vocals and accompaniment. Its local Python pipeline can create music, cover a recording or render a revised score into 48kHz stereo audio.
- Speech & audioOpen weights
Fish Audio Drama 3
Fish Audio Drama 3 generates speech with directions for emotion, pacing and individual words. It is available in Fish’s web and mobile apps, through the API and as downloadable weights. The API launch offer is $10 per million characters from October 13 to November 12, 2026, against a $50 regular price. Budget the text submitted for every take you generate.
- Speech & audioHosted API
GPT-Live
GPT-Live is OpenAI’s voice model family for conversations where the assistant can listen and speak at the same time. GPT-Live 1 has a developer API billed by session duration. The October 15 update brings GPT-Live 1.1 to ChatGPT, Codex and Dots, with Quip as the default for other users. The API rollout is a separate step.
- Speech & audioHosted API
Voxtral 3
Voxtral 3 is Mistral’s family for transcription, speech synthesis and voice agents. Choose the task before budgeting: speech output uses characters, while transcription and the managed voice agent use minutes.
- Speech & audioOpen weights
Ming-3.0-tts-0.5B
Ming-3.0-tts-0.5B is inclusionAI’s MIT-licensed, dense 0.5B speech model. It supports Chinese, English, German, Spanish, French, Korean and Japanese, with voice cloning and delivery controls. The release bundles multiple speech components; parameter count alone does not size a working rig.
- Speech & audioOpen weights
Phonon
Phonon is Gradium’s family of streaming text-to-speech models for local applications. It turns text into mono 24kHz speech, including the audio decoding needed to produce a waveform. The 31M English model is Apache 2.0; the 40M single-language and multilingual models use CC BY-NC 4.0. Choose the variant and its use rights before comparing device costs.
- Speech & audioOpen weights
Hy-SimulMT
Hy-SimulMT translates incoming Chinese or English speech while the speaker continues talking. Its translation model produces text; SimulInfer connects audio input, the model service and optional text-to-speech. Choose the standard checkpoint for translated text or Interleave for source transcription alongside the translation. Spoken output and voice cloning require a separate TTS service.
Other
Compare models in this category
Choose 2–3 models. Fields come from dated catalog records; different API units stay separate.
0 of 3 selected
Select at least two models to see their available facts side by side.
- OtherOpen weights
Griffin Alpha-S
Griffin Alpha-S is Griffin Labs’ open-weight vision-language-action model for robots. It combines a Qwen3-VL-4B backbone with a 910M-parameter action expert and predicts sequences of robot actions from observations and a task. The release includes a pre-trained base and separate LIBERO and RoboTwin fine-tunes. Start with the checkpoint that matches your robot or simulation, then its data and control setup.
- OtherRestricted preview
Praxis-1
Praxis-1 is Runway’s robot-control model, built from the video pretraining behind its world models. It is being tested with selected robotics partners; public weights are planned for the coming months. Its appeal is adapting a visual prior to physical tasks when robot demonstrations are expensive to collect.
- OtherHosted API
FLUX 3
FLUX 3 is Black Forest Labs’ multimodal model family. Image generates and edits stills; Video creates clips with optional native audio; Action predicts robot controls from observations. Choose the specific product and endpoint for your task: availability, licenses and costs differ across these routes.