AI News

OpenAI says an internal version of its Astra model generated new results for ten long-standing problems spanning geometry, cryptography, complexity theory, and related fields. OpenAI has released manuscripts and Lean certificates; the claims should still be treated as awaiting broader independent academic scrutiny.

DeepSeek says the July 31 V4-Flash update improves agent performance, supports the Responses API natively, and is adapted for Codex. The V4-Pro API and app/web models are unchanged.

OpenAI says it banned a coordinated network of ChatGPT accounts used to support investment, romance, gambling, and law-enforcement impersonation scams. The report concerns malicious use of the service, not a breach of ordinary ChatGPT user accounts.

Lyria 3.5 materially improves musicality, lyric quality, vocal realism, pronunciation, tempo control, and output duration for AI music creators.

Gemini Robotics 2 adds whole-body control, dexterous manipulation, task orchestration, progress tracking, and multi-robot collaboration. The ER 2 reasoning model is available through the Gemini API and Google AI Studio, while the VLA and on-device models remain in early access.

Anthropic found three incidents across six cyber-evaluation runs where test machines unexpectedly had live internet access. Claude models reached real systems, including production data and a malicious PyPI package. Anthropic says the incidents primarily reflected containment and operational failures, not a breach of ordinary Claude user accounts.

Starting July 30, OpenAI reduced GPT-5.6 Luna API pricing by 80% and GPT-5.6 Terra pricing by 20%. Fast mode also replaces Priority Processing; the change applies to API usage, not ChatGPT subscription plans.

The program starts with 10,000 researchers and plans to expand to 100,000 through 2027, providing frontier ChatGPT models, Codex, larger context windows, and research support.

LiteLLM now signs its Docker images with Cosign so teams can verify container provenance before deployment; the release also tightens proxy authorization and sanitizes guardrail scan results before logging.

Dify 1.16.1 isolates the local agent sandbox behind a restricted proxy, adds token authentication between internal services, blocks template-injection risks, and patches a React Server Components denial-of-service issue.

n8n 2.33.0 fixes AI Builder behavior around loop send-and-wait gates, keeps assistant replies in the user language during tool-heavy turns, and improves channel configuration for integrations including Linear and Telegram.

This major release goes beyond a visual redesign: it adds sub-agents, notifications, automations, improved sharing, file and knowledge search, plus stronger administration and security controls for self-hosted AI workspaces.

Ollama 0.32.5 fixes an MLX Metal issue that could reduce output quality for NVFP4 models, with the release notes specifically highlighting Laguna.

Ollama 0.32.4 adds Laguna support through the MLX engine for Apple GPUs and includes fixes for quantized-model decoding and speculative decoding.

Anthropic says Claude Opus 5 improves long-running agents, coding, and professional work. This update does not change the site’s pricing or ranking guidance.

OpenAI updated Codex into a broader AI work app: it now includes browser verification for rendered pages, computer use for macOS app workflows, pull request review in the app, artifact previews, memories where available, and scheduled thread automations for follow-up work.

Google Developers announced Gemini 3 Flash availability in Gemini CLI for most paid-tier users and preview-enabled access. The model is positioned as a fast, lower-cost terminal coding model with strong reasoning, tool use, and multimodal performance for high-frequency development workflows.

Anthropic updated its Claude for Teachers announcement to clarify that whether and how educational data may be used is determined by district and state policies. The clarification is relevant to school administrators evaluating privacy and governance requirements for classroom AI use.

Google introduced Gemini 3.6 Flash and Gemini 3.5 Flash-Lite, expanding its fast-model lineup for agentic execution, coding, and high-throughput workloads. Gemini 3.6 Flash is positioned as a faster, more efficient model with improved computer-use performance, while 3.5 Flash-Lite targets lower-cost, high-volume use cases. The release also includes Gemini 3.5 Flash Cyber for cybersecurity workflows.

Anthropic announced a capacity expansion that doubles Claude Code five-hour rate limits for Pro, Max, Team, and seat-based Enterprise plans, removes peak-hours limit reductions for Pro and Max Claude Code accounts, and raises Claude Opus API rate limits. The update is backed by a new SpaceX compute partnership and directly improves developer and API usage capacity.

ElevenLabs introduced Audio Tags for Eleven v3, allowing creators to control emotion, delivery, reactions, accents, pauses, and multi-character dialogue directly in scripts. The feature is useful for audiobooks, character voices, games, and dialogue-heavy media. Eleven v3 is an alpha research preview.

Beginning July 20, Anthropic says Claude Fable 5 is included in Max and Team Premium plans at 50% of limits, while Pro and Team Standard users continue through Fable usage credits and receive a one-time $100 credit. The change makes Fable 5 a premium budgeted model rather than a stable flat-rate entitlement for most subscription users.

Moonshot AI's Kimi K3 combines native multimodality, a 1M-token context window, and open weights for long-horizon coding, knowledge work, and reasoning. Independent deployment, availability, and cost reviews are still in progress.

OpenAI increased the ChatGPT custom-instruction limit from 1,500 to 5,000 characters for Plus, Pro, Enterprise, Business, and Education users. The larger limit gives paid users more room to specify response style, context, and workflow preferences across existing and new chats.

OpenAI rolled out unified ChatGPT search across chats, projects, images, and documents on web, iOS, and Android. Users can search from the sidebar, filter by content type, and reopen past chats, projects, or files from one place.

ElevenLabs integrated GPT-5 into its Conversational AI platform, adding stronger reasoning and tool calling for voice agents handling complex customer support and enterprise workflows. The integration is available in the UI, API, and SDKs.

Anthropic introduced Reflect in beta for Claude users with memory enabled. The dashboard summarizes Claude usage patterns across 1, 3, 6, or 12 months and helps users refine how they use AI. It is available in Settings on Claude for web and the desktop app for eligible Free, Pro, and Max users.

OpenAI released GPT-5.6 on July 9, 2026. GPT-5.6 Sol is available in ChatGPT for eligible paid users, replacing the earlier preview and limited-rollout state.

OpenAI introduced GPT-Live, a full-duplex voice model that powers ChatGPT Voice. GPT-Live can listen and speak at the same time, use short acknowledgement cues, and support more natural back-and-forth conversations for voice-first workflows.

Google announced Gemini Omni, a new flagship multimodal model family.

Google announced fluid, natural voice translation with Gemini 3.5 Live Translate.

Google opened developer access to Nano Banana 2 Lite (image) and Gemini Omni Flash.

Computer use is now a built-in tool in Gemini 3.5 Flash, replacing the standalone 2.5 computer use model, with enterprise safeguard systems for prompt-injection risks.

Google launched Gemini 3.5 and released 3.5 Flash as the new default model in the Gemini app and AI Mode in Search, with frontier agentic/coding performance; 3.5 Pro to follow.

Anthropic introduced richer Claude Enterprise admin analytics, model-level entitlements, and spend controls for agentic work. Admins can review usage and cost by organization, team, product, and model through the analytics dashboard and Analytics API; set model defaults and role-based model access; and receive alerts at 75% and 90% of the organization spend limit. Users receive in-app notifications at 75% and 95%, while the Admin API supports automated limit-management workflows.

Claude Fable 5 (claude-fable-5) returned to full global availability on July 1, 2026 after a brief suspension. It features always-on adaptive thinking, 1M token context, and is available via API, AWS Bedrock, Google Vertex AI, and Microsoft Foundry.

Claude Sonnet 5 (claude-sonnet-5) became the default for Free and Pro plans on June 30, 2026. Claude Fable 5, suspended June 12 under US export controls, was restored globally on July 1 across Claude.ai, Claude Code, API, AWS, Google Cloud, and Microsoft Foundry. Sonnet 5 introductory pricing: $2/$10 per MTok through August 31.

Anthropic launched Claude Science in beta: an AI workbench app for scientists with 60+ curated skills/connectors, reproducible artifacts, and compute management, for Pro/Max/Team/Enterprise users.

Claude Sonnet 5 (claude-sonnet-5) became the default model for Free and Pro plans on June 30, 2026. It delivers best-in-class speed and intelligence for coding and agentic workflows, with a 1M token context window and introductory pricing of $2/$10 per MTok through August 31, 2026.

Anthropic released Claude Science, a research workbench app for macOS/Linux that integrates 60+ scientific databases and skills (genomics, proteomics, cheminformatics), produces reproducible artifacts, and manages HPC/cloud compute. In beta for Pro, Max, Team and Enterprise plans.

Google has announced end-of-life for Veo 2.0 and 3.0 API endpoints. All existing integrations must migrate to Veo 3.1 before June 30, 2026.

Google integrated computer use natively into Gemini 3.5 Flash, letting developers build agents that see, reason and act across browser, mobile and desktop. Previously a standalone Gemini 2.5 model; now available via the Gemini API and Gemini Enterprise Agent Platform, with new enterprise safeguards against prompt injection.

Anthropic released native Slack integration allowing teams to @mention Claude directly in channels. Available on Pro and Max plans.

Google AI for Developers now lists Gemini 3.5 Flash as generally available, stable, and ready for scaled production use. The model targets agentic execution, coding, long-horizon workflows, sub-agent deployment, and rapid iteration, with current docs listing 1M context and 65K max output.

Runway moved from minutes-based to credits-based pricing. Standard plan ($15/mo) now includes 625 credits per month.