New AI Models

GPT-4 custom model

GPT-4 custom model is OpenAI's Custom Models program, expanded on April 14, 2024, in which a dedicated group of OpenAI researchers partners with organizations to build GPT-4 based models tailored to a specific domain, beyond what the self-serve fine-tuning API offers. It has two tiers: Assisted Fine-Tuning applies additional hyperparameters and parameter-efficient fine-tuning (PEFT) methods at a larger scale, with support for training data pipelines, evaluation systems, and bespoke configurations; Custom-Trained Model builds a purpose-built model from scratch by modifying key steps of the training process, using novel mid-training and post-training techniques to teach the model proprietary knowledge from millions of examples or billions of tokens. Official figures include SK Telecom's GPT-4 fine-tune for Korean telecom customer service, with a 35% increase in conversation summarization quality, a 33% increase in intent recognition accuracy, and a satisfaction score up from 3.6 to 4.5 out of 5 versus GPT-4, and Harvey's custom-trained legal case-law model built on data equivalent to 10 billion tokens, with an 83% increase in factual responses and attorneys preferring its outputs 97% of the time over GPT-4. The program fits organizations with large proprietary datasets and highly specific use cases.

Voice Engine

Voice Engine is OpenAI's voice cloning model: it takes text input plus a single 15-second audio sample of a real speaker and generates natural-sounding, emotive speech that closely resembles that speaker. A small model first developed in late 2022, it powers the preset voices in OpenAI's text-to-speech API as well as ChatGPT Voice and Read Aloud, and it preserves a speaker's native accent when generating translated speech in other languages. Announced on March 29, 2024, it was offered only as a small-scale preview to trusted partners, with early uses including reading assistance for children and non-readers, video and podcast translation, communication support for non-verbal people, and restoring speech for patients who lost their voice. Usage policies require explicit speaker consent and clear AI disclosure, and OpenAI applies safeguards such as audio watermarking and proactive monitoring, while a broader public release remains undecided.

Claude 3 Haiku

Claude 3 Haiku is Anthropic's fastest and most affordable model in the Claude 3 family, released in March 2024 with a 200K token context window and vision capabilities. Anthropic states it runs about three times faster than peers for the vast majority of workloads, processing roughly 21K tokens per second on prompts under 32K tokens, at $0.25 per million input tokens and $1.25 per million output tokens, about half the cost of other models in its performance tier. It is aimed at customer support, responsive chat, many small tasks in tandem, and analysis of large document volumes and images.

Eleven English v2

Eleven English v2 (model id eleven_english_v2) is the English-only text to speech model of ElevenLabs' v2 generation, released alongside Eleven Multilingual v2. The official changelog calls it our advanced Eleven English v2 model, and it powered the platform's Speech to Speech voice changer at launch in November 2023, letting users merge the content and style of an uploaded audio clip with a selected voice while staying accessible to all platform users. It fits English narration and voiceover scenarios such as audiobooks, articles and character audio, where strong English rendering matters more than multilingual coverage. In February 2024 Multilingual Speech to Speech replaced it for voice conversion, and the model is no longer listed in ElevenLabs' current model lineup, a slot later filled by models such as Flash v2 and Multilingual v2.

Eleven Flash v2

Eleven Flash v2 (model id eleven_flash_v2) is ElevenLabs' fastest text-to-speech (TTS) model, built for low-latency, conversational voice agents. It generates speech in about 75ms plus application and network latency, trading slightly lower quality and emotional depth than the Turbo models for significantly lower latency. In blind tests by ElevenLabs' human labelers it consistently outscored comparable ultra-low-latency models. Flash v2 supports English only (the multilingual Flash v2.5 covers 32 languages), costs 1 credit per 2 characters, and is available via the API and the Conversational AI platform for real-time use cases such as live voice agents and interactive applications.

Eleven Turbo v2

Eleven Turbo v2 (eleven_turbo_v2) is a speed-optimized text-to-speech (TTS) model from ElevenLabs. At launch the company billed it as its fastest model, generating speech at roughly 400ms latency, over twice as fast as its V1 models, with quality on par with Multilingual V2. It also added mulaw 8kHz output for VoIP, making it a fit for real-time voice applications where response time matters. Current official documentation lists it as English-only and deprecated, with the Flash models recommended as replacements for low-latency use cases.

1.0

Ideogram 1.0 is the February 2024 release of Ideogram's text-to-image model, trained from scratch and the company's most advanced version at launch. Its headline strength is reliable text rendering: systematic evaluation by the team showed rendered-text error rates cut by almost 2x compared to existing models. It also delivers sharper photorealism, stronger adherence to long and complex prompts, and support for multiple aspect ratios and styles. A new Magic Prompt assistant automatically enhances, extends, and translates prompts. In an evaluation modeled on the DALL·E 3 protocol, human raters preferred Ideogram 1.0 over DALL·E 3 and Midjourney V6 on prompt alignment, image coherence, overall preference, and text rendering quality. Typical uses include posters, logos, memes, T-shirt designs, birthday cards, and other graphics that combine imagery with readable text.

Sora

Sora is OpenAI's video generation model, announced on February 15, 2024 as the largest model in a family of text-conditional diffusion models trained jointly on videos and images. It takes the world simulator route: a video compression network first maps raw video into a lower-dimensional latent space, then a diffusion transformer operates on spacetime patches, so training covers variable durations, resolutions and aspect ratios. Sora generates up to a minute of high definition video in widescreen 1920x1080p, vertical 1080x1920 and everything in between, and it also accepts image and video prompts for editing tasks such as animating a static image, extending a clip forward or backward in time, and zero-shot style or environment changes between videos. OpenAI reports emerging simulation capabilities, including 3D consistency, object permanence and simulating digital worlds such as Minecraft, while noting that the physics of some basic interactions are still modeled imperfectly. It fits video creation, fast content prototyping and world simulation research; no official benchmark numbers have been published.

Claude 2.1

Claude 2.1 is Anthropic's upgraded model released on November 21, 2023, served through the API and claude.ai. It doubles the context window to 200,000 tokens, roughly 500 pages, supporting long document summarization, question answering, trend forecasting and multi-document comparison, with the full 200K context reserved for Claude Pro users. Anthropic reports a 2x decrease in false statements versus Claude 2.0, a 30% reduction in incorrect answers on comprehension and summarization tasks, and a 3 to 4x lower rate of wrongly concluding that a document supports a claim. The release also adds system prompts and beta tool use for calling developer defined functions, APIs and private knowledge bases, fitting enterprise document analysis, retrieval workflows and tool integrations.

DALL·E 3

DALL·E 3 is OpenAI's third-generation text-to-image model, rolled out to ChatGPT Plus and Enterprise users and made available to developers at DevDay on November 6, 2023 through the Images API under the model id dall-e-3. Built natively on ChatGPT, it uses ChatGPT as a brainstorm partner that expands a simple sentence into a detailed prompt automatically, and OpenAI states it understands significantly more nuance and detail than previous systems, generating images that adhere more closely to the text provided, an improvement evident even over DALL·E 2 on identical prompts. The API includes built-in moderation against misuse and offers different format and quality options with prices starting at $0.04 per image; companies such as Snap, Coca-Cola, and Shutterstock have used it to generate images for customers and campaigns. It fits conversational image creation in ChatGPT, marketing visuals, and programmatic image generation inside applications.

GPT-4 Turbo

GPT-4 Turbo is OpenAI's upgraded GPT-4 model, announced at DevDay on November 6, 2023. It supports a 128K context window that fits the equivalent of over 300 pages of text in a single prompt, with knowledge of world events up to April 2023. OpenAI priced its input tokens 3x cheaper and output tokens 2x cheaper than GPT-4. The model improves instruction following and function calling, can call multiple functions in a single message, adds JSON mode via the response_format parameter, and supports reproducible outputs through the seed parameter. It can also accept images as inputs in the Chat Completions API. Released to paying developers as gpt-4-1106-preview, it suits long-document analysis, structured output generation, tool use, and vision-based applications.

GPT-4 Turbo with Vision

GPT-4 Turbo with Vision is the image-understanding version of GPT-4 Turbo announced at OpenAI DevDay on November 6, 2023, exposed in the API as gpt-4-vision-preview. It builds on the Turbo upgrades, including a 128K token context window, improved instruction following, JSON mode, and parallel function calling, and adds the ability to accept images as input in the Chat Completions API for use cases such as generating captions, analyzing real-world images in detail, and reading documents with figures. Official pricing lists a 1080x1080 pixel image at $0.00765, alongside input tokens 3x and output tokens 2x cheaper than GPT-4. OpenAI stated it planned to roll vision support into the main GPT-4 Turbo model as part of its stable release. It fits multimodal applications such as visual question answering, document parsing, and accessibility tools like BeMyEyes.

0.2

Ideogram 0.2 is a text-to-image model version from Ideogram, released on November 2, 2023 as the second iteration of its early model line. Per the official announcement, it significantly improves image quality over 0.1, renders more accurate and natural text in generated images, and offers realistic details at a faster generation speed. It became the default model on ideogram.ai at launch, while v0.1 remained available and users could go back and forth between the two with remix. It fits design work that needs legible typography inside images, such as posters, logos, and signage.

0.1

Ideogram 0.1 is the first text-to-image model of the Ideogram platform, introduced in public beta when the company launched in August 2023. Ideogram was founded by former Google Brain researchers behind Denoising Diffusion Models and the Imagen text-to-image system, and the model generates images from text prompts with a notable strength in rendering legible text inside images, which makes it practical for posters, logos, and other design work. Early users could start generating images for free at launch.

GPT-3.5 Turbo

GPT-3.5 Turbo is OpenAI's low-cost, fast chat model for general-purpose text work, the series behind the free tier of ChatGPT. On August 22, 2023, OpenAI opened supervised fine-tuning for it: a fine-tuned version can match or even outperform base GPT-4-level capabilities on certain narrow tasks, follow instructions more reliably, keep output formats consistent such as valid JSON, and adopt a custom brand tone. Fine-tuning also shortens prompts, with early testers cutting prompt size by up to 90%, and handles 4k tokens of training data, double the previous fine-tuned models. Pricing was announced as $0.008 per 1K tokens for training, with usage at $0.012 input and $0.016 output per 1K tokens, and fine-tuning support for function calling and the 16k context version to follow. Training data is screened through OpenAI's Moderation API and is not used by OpenAI to train other models.

Claude Instant 1.2

Claude Instant 1.2, released on August 9, 2023, is Anthropic's faster, lower-priced yet still very capable model, incorporating strengths from Claude 2. Compared with Claude Instant 1.1, it improves on math, coding, and reasoning, scoring 58.7% on the Codex evaluation (up from 52.8%) and 86.7% on GSM8K (up from 80.9%). It produces longer, more structured responses, follows formatting instructions more reliably, and improves on quote extraction, multilingual capabilities, and question answering. In automated red-teaming evaluations it hallucinates less and is more resistant to jailbreaks. Intended for casual dialogue, text analysis, summarization, and document comprehension, it is available through the Anthropic API.

Eleven Multilingual v2

Eleven Multilingual v2 is ElevenLabs' foundational multilingual TTS model, released in August 2023 when the platform came out of beta. It supports nearly 30 languages, automatically detects the input language, and generates emotionally rich speech while keeping each speaker's unique voice characteristics, including their original accent, consistent across all languages. It integrates with Professional Voice Cloning, and the official docs describe it as the most lifelike model with rich emotional expression and the most stable option for long-form generation. Typical uses include audiobooks, gaming and animation character voiceovers, dubbing, corporate video, e-learning, accessibility, and other multilingual content.

Claude 2

Claude 2 is Anthropic's second-generation conversational model released on July 11, 2023. It accepts up to 100K tokens of context, enough for hundreds of pages of documentation or a full book, and can produce longer outputs than its predecessor. Per the official announcement it scores 71.2% on Codex HumanEval for Python coding, 88.0% on GSM8k grade-school math, and 76.5% on the multiple-choice bar exam, all improved over Claude 1.3, with GRE reading and writing above the 90th percentile. It powers the public claude.ai chat interface and is offered through the Anthropic API at the same price as Claude 1.3, fitting coding assistance, long-document analysis, math reasoning, and long-form writing workloads.

Eleven Multilingual v1

Eleven Multilingual v1 is ElevenLabs' first multilingual text-to-speech (TTS) model, building on its English-only predecessor (Eleven Monolingual v1) and adding seven new languages: French, German, Hindi, Italian, Polish, Portuguese, and Spanish. It keeps the context-aware delivery of the original model, conveying intent and emotion realistically, and can detect the language of the input text to generate speech in multiple languages from a single prompt, though single-language prompts are recommended for best results. It works with Instant Voice Cloning and Voice Design, so a cloned voice keeps its original traits, including accent, across languages. Numbers, acronyms, and foreign words sometimes default to English and are best spelled out in the target language. The model has a 10,000 character limit per request and was offered across all subscription plans at launch, fitting storytelling, audiobook, and multilingual voiceover use cases.

Claude

Claude is Anthropic's first publicly available AI assistant, announced on March 14, 2023. Built on Anthropic's research into helpful, honest, and harmless AI systems and its constitutional AI approach, it handles conversational and text processing tasks including summarization, search, creative and collaborative writing, Q&A, and coding. It shipped in two versions: the state-of-the-art high-performance Claude and a lighter, faster, less expensive Claude Instant, accessible through a chat interface and API in the developer console. Early customers reported fewer harmful outputs, easier conversation, and more steerable responses, including direction on personality, tone, and behavior.