Tags: AI model

Anthropic announced Thursday that its latest large‑language model, Claude Opus 4.8, will roll out to customers this week. The company says the new version emphasizes "honesty," flagging uncertainty and avoiding unsupported claims more effectively than its predecessor. Early testers report a four‑fold drop in unnoticed coding flaws. Opus 4.8 also lets users dial the amount of computational effort the model spends on a task, helping manage token limits. A new "dynamic workflows" feature, launched in research preview, enables Claude to orchestrate hundreds of parallel sub‑agents and verify their output before returning results. Lire la suite

OpenAI announced on Tuesday that its latest foundation model, GPT‑5.5 Instant, will replace GPT‑5.3 Instant as the default engine behind ChatGPT. The upgrade targets reduced hallucinations in high‑risk domains such as law, medicine and finance while preserving the low‑latency performance users expect. GPT‑5.5 also scores higher on math and multimodal reasoning benchmarks, and introduces a context‑management feature that lets the model reference past chats, files and Gmail. The changes roll out to Plus and Pro users on the web now, with broader availability slated for mobile and other plans in the coming weeks. Lire la suite

Anthropic disclosed that a small group of unauthorized users gained access to its newly released Claude Mythos model on the day the company announced a limited rollout. According to Bloomberg, the intruders guessed the model’s online location using details leaked from a prior breach at data‑training firm Mercur and insider knowledge from a contractor who had evaluated Anthropic’s models. Anthropic said it is investigating the incident and reviewing its monitoring systems, which were designed to log and track model usage. The breach, described by security researchers as a standard “educated guess” attack rather than a sophisticated exploit, did not appear to target the model’s advertised cybersecurity capabilities. The episode raises questions about the robustness of Anthropic’s security controls for a product it has marketed as a “watershed moment” for defending digital infrastructure. Lire la suite

The National Security Agency has begun using Anthropic’s new Mythos Preview, a general‑purpose language model touted for its strength in computer‑security tasks. Sources familiar with the rollout say the NSA is one of roughly 40 agencies granted access and that usage is expanding within the department. The move comes despite a months‑long feud between the AI firm and the Pentagon, a February order from former President Trump to halt government use of Anthropic services, and ongoing lawsuits over the company’s designation as a supply‑chain risk. Lire la suite

Anthropic has unveiled Claude Opus 4.7, the company’s most capable model offered to the public to date. Marketed as a step up from Opus 4.6, the new system promises stronger performance on software‑engineering tasks, improved image analysis, and more creative output for slides and documents. While Anthropic continues to restrict its flagship Mythos Preview to a handful of partners, Opus 4.7 ships with added cybersecurity safeguards and the same token‑based pricing as its predecessor. Early adopters include Intuit, Shopify, Databricks and other tech firms eager to test the model’s enhanced capabilities. Lire la suite

Treasury Secretary Scott Bessent and Federal Reserve Chair Jerome Powell called senior executives from the nation’s largest banks to a closed‑door meeting this week, urging them to pilot Anthropic’s newly unveiled Mythos model for spotting security flaws. JPMorgan Chase is the first bank granted access, while Goldman Sachs, Citigroup, Bank of America and Morgan Stanley are already testing the system. Anthropic says the model is not a dedicated cybersecurity tool but its ability to uncover vulnerabilities has drawn both interest and skepticism, especially as the company fights a Trump administration lawsuit over a DoD supply‑chain risk designation. Lire la suite

Anthropic announced a limited rollout of Claude Mythos Preview, a cybersecurity‑focused artificial‑intelligence model, to a handful of vetted customers such as Amazon, Apple, Microsoft, Broadcom, Cisco and CrowdStrike. The move follows two recent data leaks that exposed internal documents and source code, prompting the company to tighten distribution while it continues talks with the U.S. government about the model’s use. Anthropic says Mythos can spot vulnerabilities at a scale beyond human analysts but could also be weaponized if it falls into the wrong hands. Lire la suite

Meta announced Muse Spark on Wednesday, the inaugural AI model from its Superintelligence Labs. Marketed as a "ground‑up overhaul" of the company’s artificial‑intelligence work, the proprietary system will draw on public content from Instagram, Facebook and Threads to enhance its answers. While Meta says future Muse models will be open source, Spark marks a clear break from the earlier Llama family. Benchmarks show the model performing on par with or better than rival offerings from OpenAI, Anthropic, Google and xAI, though Meta admits gaps remain in long‑term reasoning and coding tasks. Lire la suite

Anthropic announced Tuesday that its newest frontier AI model, Mythos, will be deployed in a restricted preview for twelve leading tech firms under a new initiative called Project Glasswing. The model, described as the company’s most powerful to date, will scan both proprietary and open‑source software for zero‑day vulnerabilities. Anthropic says Mythos has already identified thousands of critical bugs, many decades old, and will be used for defensive security work while the firm continues discussions with U.S. officials about its broader applications. Lire la suite

OpenAI has expanded its GPT-5.4 family with two new variants—GPT-5.4 mini and GPT-5.4 nano. The mini model is now accessible to Free and Go ChatGPT users via the "Thinking" menu and serves as a fallback for paid users who hit rate limits. It delivers reasoning, multimodal understanding, and tool‑use capabilities that approach the full GPT-5.4 while running more than twice as fast. The nano model is targeted at data‑classification and extraction tasks, offered exclusively through the API at a cost‑effective price of $0.20 per million input tokens. Lire la suite

Anthropic has introduced Sonnet 4.6, the latest iteration of its mid-size model, as part of its four‑month update rhythm. The new version improves coding, instruction‑following, and computer‑use capabilities and becomes the default for both Free and Pro plan users. A beta rollout offers a one‑million‑token context window—twice the size of the previous maximum—enabling handling of entire codebases, lengthy contracts, or dozens of research papers in a single request. The launch follows the Opus 4.6 release and is accompanied by strong benchmark results, including a 60.4% score on ARC‑AGI‑2, positioning Sonnet 4.6 above most comparable models. Lire la suite

ByteDance announced Seedance 2.0, a next‑generation AI model that can create short video clips from combined text, image, audio, and video prompts. The system supports up to nine images, three video clips, and three audio clips per request and can produce 15‑second videos that respect camera movement, visual effects, and physical laws. Demonstrations include synchronized figure‑skating routines, anime‑style scenes, and celebrity‑lookalike cinematic fights. Seedance 2.0 is currently available through ByteDance’s Dreamina AI platform and the Doubao assistant, with no clear plan for TikTok integration. Lire la suite

Anthropic released Opus 4.6, adding "agent teams" that let multiple agents work on segmented tasks in parallel. The model now supports a 1 million‑token context window and integrates Claude directly into PowerPoint, expanding its usefulness beyond software developers to a wider range of knowledge workers. Lire la suite

Google unveiled Gemini 3 Flash, a faster and cheaper AI model built on the recent Gemini 3 architecture. The company is making Flash the default model in its Gemini app and AI‑enabled search, while still offering the Pro version for more demanding tasks. Gemini 3 Flash delivers notable performance gains on benchmark tests, supports multimodal inputs such as video, sketches, and audio, and is available through Vertex AI, Gemini Enterprise, and an API preview. Early adopters like JetBrains and Figma are already integrating the model, and Google highlights its suitability for bulk, work‑horse workloads. Lire la suite

OpenAI introduced ChatGPT-5.2, highlighting a suite of new capabilities that aim to improve the model's adherence to user‑specified constraints, offer richer perspective shifts, provide more personalized recommendations, anticipate potential failures, and help users clarify vague intuitions. The rollout includes claims of better rule‑following, nuanced viewpoint generation, targeted question‑driven recommendations, risk‑aware planning, and structured hunch analysis, positioning the model as a more thoughtful and practical AI assistant. Lire la suite

Anthropic has released Claude Opus 4.5, its latest AI model, describing it as “meaningfully better” than prior versions. The upgrade targets faster, more accurate performance on real‑world tasks such as email drafting, document creation, slide‑deck generation, and coding challenges. It also aims to improve reliability for both individual users and enterprise workflows while keeping costs stable. Enhanced handling of longer contexts, denser prompts, and multi‑step workflows are highlighted, along with better visual output capabilities and stronger integration with external tools. The company acknowledges that the model still has blind spots but positions the release as a tangible step forward for everyday productivity. Lire la suite

Anthropic has introduced Claude Haiku 4.5, a new AI model that prioritizes speed and cost efficiency while delivering performance close to its larger sibling, Claude Sonnet. Marketed as a sub‑agent that can handle small, targeted tasks under the direction of larger models, Haiku 4.5 becomes the default option for all Claude free‑tier users. The model promises double the latency speed of previous small models, lower sycophancy, and tighter integration with Anthropic’s tool ecosystem, offering a faster, cheaper entry point for developers and everyday users alike. Lire la suite

DeepSeek announced a new experimental AI model featuring Sparse Attention technology that dramatically lowers inference costs for long‑context tasks. The model, released on Hugging Face and accompanied by a research paper on GitHub, uses a lightning indexer and fine‑grained token selection to focus computational resources on the most relevant excerpts. Preliminary tests suggest API call prices can be cut by as much as 50 percent in long‑context scenarios. The open‑weight release invites third‑party validation and positions DeepSeek as a notable player in the ongoing effort to make transformer‑based AI more cost‑effective. Lire la suite

Anthropic has made its new Claude 4.5 model available through the API and its web interface, keeping pricing identical to Claude Sonnet 4 at $3 per million input tokens and $15 per million output tokens. The model can maintain focus for 30 hours on multistep tasks and now includes built‑in code execution, file creation, and the ability to generate spreadsheets, slides, and documents directly in chat. A five‑day research preview called “Imagine with Claude” showcases real‑time software generation. Claude Code receives checkpoints, a refreshed terminal, and a native VS Code extension, while the API adds context‑editing and memory tools. Anthropic also highlights reductions in sycophancy, deception, power‑seeking, and delusional prompting. Lire la suite

DeepSeek announced its experimental model V3.2‑exp, featuring a new Sparse Attention mechanism that dramatically lowers inference expenses for long‑context tasks. The architecture employs a lightning indexer to prioritize excerpts and a fine‑grained token selector to feed a limited attention window, allowing the model to process extensive context with reduced server load. Preliminary tests suggest API calls in long‑context scenarios could cost up to half as much as before. The model is open‑weight and freely available on Hugging Face, inviting independent verification and broader adoption. Lire la suite