Latest model releases, tools and events in artificial intelligence
Cursor shipped three changes in a week. Google Workspace plugins let agents read, write and act across Gmail, Drive and Calendar. Cursor for iPad arrived on all paid plans with pinned sidebar chats for watching several agents at once. Cursor Start launched at Rs 649 per month for Indian developers, including Grok 4.5 and Composer.
DeepSeek released V4-Flash version 0731 at $0.14 per million input tokens and $0.28 output, with a 1M context window and both thinking and non-thinking modes. Cache hits cost $0.0028 per million. The pricing documentation carries an explicit warning: DeepSeek plans to raise API prices, with a significant increase expected.
OpenAI dropped the price of GPT-5.6 Luna to $0.20 per million input tokens and $1.20 per million output tokens. Luna launched at $1 per million input tokens, so this is an 80% cut. The rest of the family holds at Terra $2/$12 and Sol $5/$30 per million tokens.
Anthropic disclosed what it found while examining three real-world incidents surfaced through its cybersecurity evaluations. The publication is part of the company's ongoing safety reporting rather than a product announcement.
Google opened public preview for Gemini Robotics ER 2 with two endpoints: gemini-robotics-er-2-preview for advanced spatial reasoning and multi-robot coordination, and a streaming variant optimised for real-time use through the Live API. The previous generation, gemini-robotics-er-1.6-preview, shuts down on August 31, 2026.
Anthropic published a statement outlining where it stands on open-weights models. The same day it extended its collaboration with Cognizant to deploy Claude across enterprise clients.
Anthropic released Claude Opus 5 on July 24. It brings frontier-class agentic coding and computer use while keeping Opus pricing unchanged at $5 per million input and $25 per million output tokens — half the price of Claude Fable 5. Opus 5 has a 1M-token context window, up to 128K output tokens, thinking on by default and a May 2026 knowledge cutoff. It is now the default on Claude Max and the strongest model on Claude Pro.
On July 22, AMD and Anthropic announced a strategic partnership to deploy up to two gigawatts of AMD Instinct MI450-series GPUs, with the first gigawatt online in the first half of 2027. AMD intends to invest up to $5 billion in Anthropic, while Anthropic commits to buying tens of billions of dollars in AI server chips over the deal. The two also launched an engineering collaboration using Claude to optimize Instinct workloads and ROCm software — a direct challenge to Nvidia's grip on AI infrastructure.
Google moved gemini-3.6-flash and gemini-3.5-flash-lite to general availability. 3.6 Flash brings better token efficiency and stronger code and agentic planning at a lower price point; Flash-Lite targets low-latency subagent work. A breaking change ships with them: temperature, top_p and top_k are deprecated for both models.
Google's Gemini 3.5 Pro — announced at I/O in May with a 2M-token context window and a Deep Think reasoning layer — slipped again in mid-July. After targeting a July 17 launch, Google reportedly delayed it when the model missed internal quality goals on hallucination rates and real-world reliability, with DeepMind said to have scrapped and rebuilt the base model. The faster Gemini 3.5 Flash remains Google's current stable model.
On July 16, the European Commission issued two binding decisions under the Digital Markets Act requiring Google to give competing AI assistants equal access to Android. Google must open 11 system features to third-party AI — including on-device model access, background task execution and screen context on par with Gemini — with most changes due by August 2027. A second decision obliges Google to share anonymized search data with rival search engines and AI chatbots from January 2027. Google has strongly opposed the move, warning of security risks.
On July 16, China's Moonshot AI released Kimi K3, a 2.8-trillion-parameter open-weight Mixture-of-Experts model with a 1M-token context window — the largest open-weight model ever shipped. Within hours it debuted at #1 on LMArena's Frontend Code Arena with a score of 1679, overtaking Claude Fable 5 on the benchmark that most directly measures production coding value. It signals that open models from China are now competing at the frontier despite US chip-export limits.
On July 9 OpenAI publicly released its GPT-5.6 family in three tiers: Sol (the top model), Terra (mid-range) and Luna (fast and cheap). Sam Altman said Sol is about 54% more token-efficient on coding than the previous generation. Alongside the models OpenAI shipped ChatGPT Work — an agent built to carry out whole jobs rather than just answer questions — and GPT-Live, a full-duplex voice mode that listens and speaks at the same time.
On June 24, Google added Computer Use to Gemini 3.5 Flash, letting the model operate a computer's interface — clicking, typing and navigating apps — to complete tasks as an autonomous agent. It brings agentic UI automation to Google's fastest stable model. The larger Gemini 3.5 Pro, with Deep Think reasoning and a 2M-token context window, is still expected in July.
Anthropic released Claude Fable 5 on June 9 — the first publicly available Mythos-class model. State-of-the-art across software engineering, vision, scientific research, and autonomous task execution. Simultaneously launched Claude Mythos 5, restricted to approved organizations for government cybersecurity use. Free on Pro, Max, Team, and Enterprise plans through June 22; from June 23 priced at $10/$50 per 1M tokens.
All GitHub Copilot plans switched to usage-based billing on June 1. Every interaction now costs based on the model used and tokens consumed. New currency: GitHub AI Credits (1 Credit = $0.01). Pro plan includes 1,500 credits/month, Pro+ includes 7,000 ($39/mo), Max includes 20,000. Business and Enterprise customers receive extra credits through September 1, 2026.
Anthropic raised $65 billion in a Series H round at a $965 billion post-money valuation, led by Altimeter, Dragoneer, Greenoaks, and Sequoia. On June 1 the company confidentially filed an S-1 for an IPO. Revenue run-rate hit ~$47B as of May 2026, surpassing OpenAI's ~$852B valuation and making Anthropic the most valuable AI startup in history.
Anthropic launches Opus 4.8 just 41 days after 4.7. USAMO 2026 math score jumps from 69.3% to 96.7%. Dynamic Workflows in Claude Code fan out up to 1,000 parallel subagents for repo-scale tasks. Fast mode drops from $30/$150 to $10/$50 per 1M tokens — a 3x cost reduction. Same base price: $5/$25 per 1M tokens.
ByteDance disclosed plans to spend up to $70 billion on AI data centers and infrastructure in 2026 — nearly triple its 2025 capex of $25 billion. The company plans to self-fund most of the spending from approximately $50 billion in 2025 profits. The figure puts ByteDance in the same capex tier as Microsoft and Google.
Google's most product-dense I/O in years. Gemini 3.5 Flash replaces the previous flagship at Flash speed and $1.50/1M tokens. Gemini Spark launches as a proactive 24/7 personal agent running on cloud VMs. Android XR smart glasses with a real display go hands-on. Google AI Ultra replaces the $250/month plan at $100/month.
OpenAI co-founder and former Tesla AI lead Andrej Karpathy has joined Anthropic pretraining team. He is launching a new group focused on using Claude itself to accelerate pretraining research. Karpathy works under pretraining lead Nick Joseph. The hire is seen as a major signal of Anthropic momentum in attracting elite technical talent.
A California federal jury unanimously dismissed all claims in Elon Musk's $150 billion lawsuit against OpenAI and Sam Altman. The verdict came in under two hours: statute of limitations had expired on all counts. Musk announced plans to appeal.
OpenAI and Dell Technologies partnered to deploy Codex in hybrid and on-premises enterprise environments via the Dell AI Data Platform — OpenAI's first explicit on-prem distribution deal. Codex now has 4M+ weekly developer users.
Cursor shipped Composer 2.5, powered by Kimi K2.5, achieving ~63% on SWE-Bench — matching Claude Opus 4.7 — at $0.50/task vs. $7/task for Opus. Also ships multi-repo cloud agent environments, Dockerfile-based environment setup, and enterprise audit logs.
Google I/O 2026 opens this week with a new Gemini model announcement expected on May 19. Google has signaled major updates across the Gemini family and the broader Google AI product portfolio. The event follows the Google Android Show 2026 where Gemini Intelligence for multi-step task automation on Android was announced.
AWS blocked new Amazon Q Developer IDE plugin signups from May 15, 2026, announcing its replacement: Kiro, a spec-driven agentic IDE. Kiro uses "Specs" (natural-language requirements that drive full implementation) and "Hooks" (auto-triggers on file save or commit). IDE plugin support ends April 30, 2027.
HeyGen discontinued legacy unlimited avatar video plans on May 15, 2026. All new HeyGen subscriptions are now credit-based. Avatar V, capable of studio-quality output from 15 seconds of footage, and the Seedance 2.0 video integration were launched alongside the pricing restructure.
Notion released Developer Platform 3.5 on May 13, 2026. Workers runtime lets developers deploy serverless JavaScript logic inside Notion. Database Sync enables two-way connections to external databases. Bidirectional Webhooks allow Notion to receive and send events. Plan Mode enables autonomous multi-step AI agents within Notion workflows.
At the Google Android Show 2026, Google announced Gemini Intelligence — a new layer that automates multi-step tasks like shopping, bookings, and calendar management directly on Android devices. Coming to Samsung and Google phones in summer 2026. Google also announced Googlebooks — new Android-powered laptops from Acer, ASUS, and Lenovo with Gemini built in.
Cursor 3.4 introduces agentic development environments: agents now provision Dockerfile-defined environments, work across multiple repositories simultaneously, and run in parallel. Cursor 3.3 had added PR review inside the editor. BugBot transitions to usage-based billing from June 2026.
Krea AI released its Krea 2 foundation model, built from the ground up with a focus on aesthetics and style transfer. The launch accompanies Krea Nodes (build image workflows from text descriptions) and Krea Edit (region editing, relighting, image expansion). Seedance 2.0 is now available on all paid plans.
OpenAI officially retired DALL-E 2 and DALL-E 3 on May 12, 2026. Both are replaced by the gpt-image-2 model, which powers ChatGPT Images 2.0 with native 2K resolution, multilingual text rendering, and up to 8 images per prompt in thinking mode.
Pika 2.5 eliminates flickering artifacts and delivers stable character appearance across scenes. PikaStream 1.0 launches real-time AI video chat: users have live conversations with AI agents rendered as generated video characters. Pika is expanding into an AI social video app platform combining creation, sharing, and real-time AI interaction.
Luma AI released Ray3.14 with native 1080p video output (no upscaling), generation speed 4x faster than Ray2, and 3x lower cost per generated second. Character Seeds maintain consistent character appearance across multiple clips. Modify with Instructions enables text-guided post-generation edits without full regeneration.
Figma launched an MCP (Model Context Protocol) server enabling AI agents such as Cursor and Claude to create and edit designs directly on the Figma canvas. Built-in AI Image Tools now include Vectorize, Remove Background, Erase, and Expand — eliminating the need to export to Photoshop. Voice input and text on path are also now available.
Midjourney V8.1 delivers native 2K resolution output, generation speed 3 to 4 times faster than V7, and Raw mode for more literal prompt interpretation. Public profiles allow creators to share their work socially. Midjourney TV launches as a real-time stream of community generations. V8.2 has been announced as the next release.
Kling AI released v3.5 with native 1080p/60fps output, extended Pro-tier clip length up to 2 minutes (the longest single generation available), and improved motion physics for fluid dynamics and cloth simulation. Kuaishou is evaluating spinning off Kling AI as an independent company at a valuation of approximately $20 billion.
OpenAI replaced GPT-5.3 Instant with GPT-5.5 Instant as the default model in ChatGPT. The new model delivers 52.5% fewer hallucinations on high-stakes prompts and integrates memory from past conversations, Gmail, and saved files for Plus and Pro users.
NIST's Center for AI Standards and Innovation announced agreements with Microsoft, Google, and xAI to evaluate unreleased AI models for national security risks before launch. The move was prompted by Anthropic's restricted Mythos model. Anthropic is notably absent from the agreement.
AI music platform Suno is raising a Series D round at a $5 billion valuation, more than doubling its November 2025 valuation of $2.45 billion. The company now has 2 million paid subscribers and generates 7 million tracks per day.
Novo Nordisk announced a strategic partnership with OpenAI to integrate AI across drug discovery, clinical trials, manufacturing, supply chains, and commercial operations. The deal targets acceleration of new obesity and diabetes treatments.
Anthropic's Model Context Protocol (MCP) crossed 97 million installs in March 2026, transitioning from an experimental standard to foundational infrastructure for building AI agents. Every major AI provider now ships MCP-compatible tooling.
Anthropic has developed a new frontier model called Mythos which the company says is "far ahead" of all other models in cybersecurity capability. Due to the risks, Anthropic is restricting access to a select group of vetted organizations. The model prompted the White House to begin weighing a formal AI review process.
Google's Gemini 3.1 Pro has emerged as the overall benchmark leader across independent evaluations, topping 13 of 16 major benchmarks. Key scores: 80.6% on SWE-bench, 94.3% on GPQA Diamond (highest of any model), 2M token context window across text, image, audio and video.
Mistral released Small 4, a 119-billion-parameter open-source model with a 256k context window, configurable reasoning, and multimodal (vision) capabilities. It combines fast instruction-following, deep reasoning, and image understanding in a single unified model under an open license.
Microsoft and OpenAI rewrote their foundational 2019 deal: Microsoft ends its Azure revenue share to OpenAI, OpenAI can now sell via Amazon Web Services and Google Cloud, and the famous "AGI clause" is removed. Anthropic's ARR has surpassed OpenAI's for the first time at $30B, which may have accelerated the renegotiation.
Google announced an investment of up to $40 billion in Anthropic, starting with $10 billion at a $350 billion valuation, with follow-on tranches tied to milestones. Google Cloud also committed 5 gigawatts of compute capacity over five years — the largest single AI investment ever recorded.
OpenAI launched GPT-5.5, its most capable model to date, with dramatic improvements in agentic coding via Codex, deep research, and the ability to operate software end-to-end. It rolled out to Plus, Pro, Business and Enterprise users in ChatGPT and the API on April 24.
Anthropic released Claude Design, its first dedicated design product, as part of Anthropic Labs. Built on Claude Opus 4.7, it generates slide decks, mockups and marketing materials from conversation — with automatic brand system generation from existing codebases or design files. Available at no extra cost for Claude Pro, Max, Team and Enterprise subscribers.
Adobe launched Creative Agent, an AI orchestration layer that accepts a single natural language prompt and drives coordinated workflows across Photoshop, Premiere, Lightroom, and Illustrator. Custom Models beta enables brand-specific fine-tuned generation. Sound generation was added to Adobe Firefly, extending the platform into audio creation.
xAI launched Grok 4.3 beta to SuperGrok Heavy subscribers — twice the size of Grok 4.20 at 1 trillion parameters. Early benchmarks show significant gains on reasoning and coding tasks. General availability is expected within weeks.
Anthropic opened a private beta for Agent Memory, a managed service that extracts key information from agent conversations and surfaces it when relevant — keeping context windows lean. Anthropic also announced a $100K developer hackathon on Claude Opus 4.7.
Anthropic's Claude Opus 4.7 has set a new record on SWE-bench Verified, scoring 87.6% — the highest ever for a publicly available model. Priced at $5/$25 per million tokens, it marks a step-change in agentic software engineering capabilities.
Apple announced a completely reimagined Siri powered by Google's Gemini model running on Apple's Private Cloud Compute. The new Siri handles multi-step tasks on-device for the first time and answers complex questions with far greater accuracy.
OpenAI released GPT-5.4 in Standard, Thinking and Pro variants. All three share a 1.05 million token context window — the largest OpenAI has ever released commercially. The model dynamically retrieves tool specs rather than loading every definition into the prompt.
Stanford's annual AI Index reveals that as of March 2026, Anthropic leads the overall model rankings, trailed closely by xAI, Google and OpenAI. The report finds people are adopting AI faster than they adopted the personal computer or the internet.
OpenAI announced it has acquired TBPN, the Technology Business Programming Network — a daily live tech and business show that has become a cult phenomenon in Silicon Valley. It is the company's first acquisition of a media company.
PwC's 2026 AI Performance study finds a sharp divide: a small group of companies focused on growth — not just productivity — is pulling far ahead. The report warns that most businesses are still in early experimentation, missing the compounding returns.
Anthropic confirmed the existence of Claude Mythos, described as the most capable model they have ever built. It will not be released to the public. Access is limited to ~50 partner organizations through Project Glasswing, focused on cybersecurity and advanced reasoning.
OpenAI has surpassed $25 billion in annualized revenue and is reportedly taking early steps toward a public listing, potentially as soon as late 2026. Anthropic is close behind, approaching $19 billion in annualized revenue.
Meta released Llama 4 Scout and Maverick — the first Llama models with Mixture-of-Experts architecture. Scout features 17B active parameters across 16 experts (109B total) and a 10-million-token context window. Available for commercial use.
Google released Gemma 4 in four variants from 2.3B to 31B parameters. The 31B Dense model ranks #3 globally on Arena AI among open models, making it the strongest open-weight model Google has shipped to date.
The EU AI Act entered full enforcement in March 2026, requiring all AI systems in the EU to meet transparency, safety and risk classification requirements. OpenAI, Anthropic and Google have now published their GPAI compliance documentation.
Chinese lab Zhipu AI released GLM-5.1 under the MIT license — a 744-billion-parameter mixture-of-experts model with 40B active parameters per forward pass and a 200K context window. The fully open release puts serious pressure on closed-source competitors.