Happening Now

AI summary · Generated from this article

Moonshot AI has released Kimi WebBridge, a Chrome extension and local background service that lets AI agents control your browser the way a person would — clicking, typing, scrolling, and extracting data — without routing your session through the cloud. It works inside Chrome and Edge using Chrome DevTools Protocol, the same low-level interface developers use for browser debugging. Most competing browser automation tools send your authenticated sessions and form entries through cloud servers. Kimi WebBridge keeps everything on your device. The extension supports multiple AI ecosystems including Claude Code, Cursor, Codex, and Hermes, making it agent-agnostic rather than a closed platform. It is powered by Kimi K2.6, which currently leads the SWE-Bench Pro benchmark at 58.6%, ahead of GPT-5.4 and Claude Opus 4.6. For anyone already running Claude Code or Cursor who wants browser automation without surrendering session privacy, WebBridge is a concrete alternative worth testing against real workflows.

AI summary · Generated from this article

Google's Threat Intelligence Group has confirmed the first real-world case of a criminal threat actor using artificial intelligence to discover and weaponize a zero-day exploit. The unknown group developed a working bypass for two-factor authentication targeting a popular open-source web administration tool. The attack was stopped before causing damage, but the forensic evidence left behind marks a significant turning point in cybersecurity. The exploit was embedded in a Python script containing hallucinated CVSS scores and heavily annotated docstrings, artifacts consistent with LLM-generated code rather than human authorship. The vulnerability itself was a semantic logic flaw rooted in a hard-coded trust assumption, exactly the type of high-level bug that traditional static analysis tools miss but that large language models are well-suited to identify. The flaw has since been patched following responsible disclosure. Security researchers warn this is likely the first confirmed case among many. Nation-state actors from North Korea, China, and Russia are already integrating AI into offensive operations at scale, compressing the timelines between vulnerability discovery and active…

AI summary · Generated from this article

Anthropic has temporarily increased Claude Code weekly usage limits by 50 percent for all paid subscribers. The boost is already live with no setup required and applies to Pro, Max, Team, and seat-based Enterprise accounts through July 13 at 6PM PDT. Free plan users are not included. The increase stacks on top of a permanent doubling of per-session 5-hour limits that Anthropic announced on May 6, meaning developers now have more headroom both within individual sessions and across the full week. The changes are powered by a new compute deal with SpaceX, giving Anthropic access to over 220,000 NVIDIA GPUs at the Colossus 1 data center in Memphis. Peak-hour throttling for Claude Code has also been permanently removed for Pro and Max users. Developers running long agentic tasks stand to benefit most from the combined changes. The key question after July 13 is whether Anthropic makes any portion of the weekly increase permanent, as it did with the 5-hour session limit expansion.

AI summary · Generated from this article

OpenAI is offering enterprises two months of free Codex usage during a 30-day promotional window, targeting engineering teams currently using GitHub Copilot and Cursor. CEO Sam Altman announced the offer on X, describing Codex as the best AI coding product available. Eligible ChatGPT Business workspaces can also earn up to $500 in promotional credits by activating new Codex seats. Codex runs on GPT-5.3-Codex, which sets new benchmarks on SWE-Bench Pro across four programming languages. Unlike autocomplete tools, Codex functions as an agentic software engineer capable of handling full features, refactors, migrations, and CI/CD tasks. OpenAI is also shifting to token-based pricing rather than per-seat licensing, so teams pay only for actual usage. With Codex crossing three million weekly users and enterprise adoption growing sixfold since January, competitive pressure on Anthropic's Claude Code and GitHub Copilot is significant. The two-month free window gives teams a real opportunity to test Codex against production codebases before committing to billing.

AI summary · Generated from this article

Google’s unannounced Gemini Omni video-generation system appears to have leaked through interface screenshots and alleged sample videos ahead of Google I/O 2026. TestingCatalog spotted “Powered by Omni” in Gemini’s video interface, while demos included a professor writing trigonometric identities and two men eating spaghetti. Reports suggest Omni may offer templates, remixing and chat-based editing, potentially alongside Toucan and Veo 3.1. Google has not confirmed whether Omni is a new multimodal model or an interface for existing technology; I/O runs May 19–20.

AI summary · Generated from this article

Anthropic is consulting religious leaders from multiple faith traditions as part of a structured effort to encode moral values into its Claude AI model. After hosting approximately 15 Christian leaders at a private two-day summit at its headquarters, the company expanded its outreach to include representatives from Hindu, Sikh, Jewish, Greek Orthodox, Baha'i, and Latter-day Saint communities at a New York roundtable called the Faith-AI Covenant. This effort is grounded in Claude's 29,000-word foundational document, which states that Anthropic wants Claude to behave as a deeply and skillfully ethical person. The company favors cultivating genuine values over rigid rules, and its own constitution acknowledges a core fear that its efforts to instill good enough values in Claude could fail. Religious insights are expected to filter into system prompts, reward modeling, and human feedback pipelines rather than directly into training data. Critics warn that no universal moral framework exists across traditions, making this approach inherently incomplete. Whether these consultations produce measurable behavioral changes in Claude or serve primarily as legitimacy-building remains an open…

AI summary · Generated from this article

Lovable has shipped a major aesthetics update that brings real design control to its AI-driven app builder. The release lets users define typography, color, and layout preferences before the AI Agent writes any code, directly addressing the visual sameness that has defined vibe-coded apps since the category emerged. A pre-build design preview lets users approve visual direction before the Agent commits to building, shifting the workflow from reactive iteration to deliberate creative agreement. The update includes a Themes panel for setting reusable brand tokens across projects, expanded Visual Edits for adjusting margins, padding, shadows, and icons without coding, and AI image generation that does not consume platform credits. The release also integrates Claude Opus 4.5, delivering 20 percent fewer errors and stronger design planning at no additional cost. For non-technical founders and marketing teams, this closes a meaningful gap. Functional apps are now table stakes in the vibe coding space, and visual differentiation is where competition is heading next.

AI summary · Generated from this article

OpenAI launched Daybreak on May 11, 2026, a cybersecurity initiative designed to shift software defense from reactive patching toward resilience built directly into development. The program combines frontier AI models, the Codex agentic harness, and a partner network including Cloudflare, CrowdStrike, Palo Alto Networks, and others covering vulnerability discovery through supply chain defense. Daybreak operates across three model tiers: standard GPT-5.5 for general use, a Trusted Access version for verified defensive workflows like malware analysis and vulnerability triage, and GPT-5.5-Cyber for authorized red teaming and penetration testing. Codex Security builds editable threat models from live codebases and has already helped fix over 3,000 critical vulnerabilities. OpenAI is also establishing a Frontier Risk Council to govern dual-use concerns as model cyber capabilities advance. Daybreak competes directly with Anthropic's Project Glasswing, which has already signed Apple, Microsoft, Google, and Amazon. That rivalry is likely to accelerate useful tooling for defenders faster than either company would deliver independently.

AI summary · Generated from this article

Anthropic has launched agent view in Claude Code, a unified dashboard that consolidates all parallel Claude Code sessions into a single command-line interface. Before this update, developers managing multiple AI agents simultaneously had to juggle terminal tabs, tmux grids, and constant context-switching to track what each session was doing. Agent view displays every active session in a roster showing its current state, last activity, and whether it is working, waiting for input, or completed. Developers can open it by pressing the left arrow from any session or running claude agents in the terminal. Sessions continue running in the background when detached, and attaching to one delivers a recap of activity missed while away. The feature requires Claude Code v2.1.139 or later and is available as a research preview on Pro, Max, Team, Enterprise, and API plans. This release reflects Anthropic's broader push to position Claude Code as an agent operations layer for software teams rather than a simple coding assistant, making large-scale parallel delegation more practical.

AI summary · Generated from this article

A new AI model called Gemini Omni surfaced inside Google's live Gemini app before any official announcement, discovered by users who spotted UI strings reading "Create with Gemini Omni" and "Powered by Omni" in the video generation tab. The leak emerged just days before Google I/O 2026, scheduled for May 19–20, where major AI announcements are expected. The model appears to be a unified system capable of handling text, images, and video in one architecture, potentially replacing Google's current split approach using Veo 3.1 for video and separate models for other modalities. Early outputs show strong prompt adherence, improved scene coherence, and notably capable in-chat editing features like object swapping and watermark removal. However, compute costs are steep, with two short video clips consuming 86 percent of a user's daily AI Pro quota. The editing capabilities and unified architecture are what make this leak significant beyond typical pre-announcement hype. If Gemini Omni delivers on what early demos suggest, it could meaningfully reshape how creators work across text, image, and video in a single workflow.

AI summary · Generated from this article

OpenAI has opened its ChatGPT advertising platform to any verified U.S. business, moving from a closed pilot program launched in February 2026 to a fully self-serve model in roughly three months. The new Ads Manager at ads.openai.com supports both CPC and CPM bidding, a Conversions API, and pixel-based conversion tracking. Agency partners including Dentsu, Omnicom, and Publicis are integrated alongside ad tech platforms like Adobe, Criteo, and StackAdapt. Ads appear only to Free and Go tier users, are clearly labeled as sponsored, and are matched to conversation topics without sharing user data with advertisers. The platform is promising but early. Independent click-through estimates sit around 1.3 percent compared to Google Search's 29.2 percent, third-party measurement remains unavailable, and advertisers currently rely on OpenAI's own aggregated reporting. The intent signals ChatGPT captures mid-conversation are genuinely valuable, but the ad format remains minimal. Modest testing budgets make sense for now, though securing independent verification will determine how seriously performance budgets shift from established platforms over time.

AI summary · Generated from this article

OpenAI replaced GPT-5.3 Instant with GPT-5.5 Instant as the default ChatGPT model on May 5, 2026. Unlike previous updates, this is a fully retrained base model rather than a fine-tune, featuring a unified architecture that handles text, images, audio, and video in a single system. It is the first rebuild of this scale since GPT-4.5. The headline improvement is a 52.5% reduction in hallucinated claims on high-stakes prompts, alongside a 37.3% drop in inaccuracies on user-flagged conversations. Responses run roughly 30% shorter, with less overformatting and fewer unnecessary emojis. A new memory sources feature lets users see and manage which past chats, files, or Gmail data shaped a given reply, with enhanced personalization launching first for Plus and Pro subscribers. Developers using the chat-latest API alias should test their prompts carefully, as default model changes affect behavior beyond accuracy alone. OpenAI's self-reported benchmark gains are notable, but independent verification will determine whether these improvements hold up in real-world use.

AI summary · Generated from this article

Google has expanded its Gemini API File Search tool with three updates: native multimodal retrieval, custom metadata filtering, and page-level citations. Previously limited to text, the tool now indexes images and text together within a single store using Gemini Embedding 2, which maps both modalities into a shared vector space. Unlike OCR-based approaches that discard visual signal, this model embeds images natively, eliminating the need for separate retrieval pipelines when working with charts, diagrams, or product photos. Metadata filters let developers scope queries to specific data slices without duplicating documents across stores, reducing latency and noise at scale. Page-level citations tie every response to a specific source page and include downloadable image references for multimodal stores, making outputs auditable for compliance-sensitive industries like legal and healthcare. The open question is retrieval precision on genuinely mixed-modality queries at scale. Early results from Harvey and Supermemory are promising, but both operate in text-heavy domains. Developers building with visually rich documents should test this update now.

AI summary · Generated from this article

Google has expanded Pomelli, its experimental AI marketing platform built in collaboration with Google DeepMind, with a new feature called Catalog. The update allows small and medium-sized businesses to point the tool at a product list URL and have it automatically ingest an entire inventory in one pass, rather than uploading items individually. The Catalog feature generates brand-consistent product shots, ad creatives, and social posts across the full inventory using Pomelli's existing Business DNA system, which extracts brand identity directly from a merchant's website. Output quality relies on the platform's Photoshoot engine, which produces studio-quality imagery for free, compared to professional photography costs that typically run between $500 and $5,000 per product. Pomelli remains free during its public beta with no usage limits. A Websites feature is also in early development, suggesting Google may be building toward a full end-to-end marketing and web presence platform for SMBs, putting it on a direct collision course with Wix, Squarespace, and similar services.

AI summary · Generated from this article

Google has released Multi-Token Prediction drafters for the entire Gemma 4 model family, delivering up to 3x faster inference without any degradation in output quality. The update covers all four model sizes, from the lightweight E2B and E4B edge models to the larger 26B and 31B variants, and is available under the Apache 2.0 license on Hugging Face and Kaggle. The speed gains come from speculative decoding, where a lightweight four-layer drafter proposes multiple candidate tokens simultaneously while the larger target model verifies them in a single forward pass. The drafter shares the target model's KV cache and uses its activations directly, keeping overhead minimal while still achieving meaningful throughput improvements. On Apple Silicon with the 26B model, batch sizes of four to eight unlock around 2.2x speedups locally. The open question worth watching is how acceptance rates vary across task types. Structured outputs and code will see the largest gains, while creative or unpredictable generation will benefit less. Google has not yet published per-task acceptance rate data, which will matter for developers deciding whether to enable drafters by default in production.

AI summary · Generated from this article

Ollama 0.23 adds native support for Claude Desktop to its ollama launch ecosystem, letting developers replace Anthropic's default inference with Ollama's open cloud models inside both Claude Cowork and Claude Code. Previously, this required manual environment variable configuration or proxy setups. Now a single command handles authentication and model discovery automatically. The release works by wiring into Claude Desktop's existing third-party inference feature. Running ollama launch claude-desktop prompts for an Ollama API key, after which cloud-hosted models appear directly in the model selector. Recommended options include kimi-k2.6 for long-horizon coding tasks and glm-5.1 for agentic engineering. Models require at least a 64k token context window to work reliably with Claude Code, and proper tool calling support is essential for real file read and write operations. For developers watching Anthropic API costs accumulate, this integration offers a practical alternative. Third-party options can cost significantly less, and the setup can be fully reversed with a single restore command.

AI summary · Generated from this article

OpenClaw, the free and open-source autonomous AI agent built by Austrian developer Peter Steinberger, now lets users sign in with an existing ChatGPT Plus or Pro account to access GPT-5.4 directly inside the agent with no separate API billing. The March 2026 update adds an OpenAI Codex OAuth integration, meaning a standard ChatGPT subscription does double duty across both the ChatGPT website and a self-hosted, persistent AI agent. The integration routes requests through OpenAI's sanctioned Codex OAuth path rather than pay-per-token API keys, unlocking a 1,050,000-token context window and 128,000 max output tokens. ChatGPT Plus subscribers receive a 5-hour weekly Codex usage quota, while Pro subscribers get higher or unlimited access. A known bug in some versions incorrectly routes GPT-5.4 requests and falls back to an older model, and the authentication flow requires an interactive terminal session, complicating headless server setups. For anyone already paying for ChatGPT, this makes running a capable, long-context autonomous agent significantly more affordable and predictable than metered API access.

AI summary · Generated from this article

OpenAI has added an unexpected feature to its Codex developer platform: animated desktop companions called Codex Pets. These small pixel creatures float on your screen and display the current status of your active coding agent, letting developers monitor background tasks without switching back to the app. The feature is live now and available in the Codex app settings. Eight predefined pets are available out of the box, and users can generate custom ones using the hatch-pet skill. Pets signal whether Codex is running, waiting for input, or ready for review, and they can be toggled via a slash command or keyboard shortcut. The feature is not available in the EEA, the United Kingdom, or Switzerland at launch. The practical case for Pets grows as Codex sessions get longer and more complex. If OpenAI expands the state-signaling vocabulary beyond basic status indicators to surface approvals, errors, or specific agent decisions, the feature could become a genuinely useful tool for anyone managing parallel agent workflows.

AI summary · Generated from this article

CVE-2026–3854 let authenticated users with repository push access potentially achieve remote code execution on vulnerable GitHub Enterprise Server instances through a crafted git push. Wiz Research traced the flaw to unsanitized semicolons in push options embedded in GitHub’s X-Stat header, enabling field injection and an unsandboxed pre-receive-hook path. GitHub patched GitHub.com on March 4, 2026, said it found no exploitation, and fixed GHES in versions 3.14.25, 3.15.20, 3.16.16, 3.17.13, 3.18.7, and 3.19.4.

AI summary · Generated from this article

Mistral AI has launched Workflows, an orchestration layer now in public preview that helps enterprise teams run AI-powered processes reliably in production. The release marks a strategic shift for Mistral, moving beyond model development toward full production infrastructure. It targets a persistent failure point in enterprise AI: systems that work in development but break down when deployed at scale. Workflows is built on Temporal's durable execution engine, the same infrastructure used by Netflix, Stripe, and Salesforce. It natively handles retries, state persistence, timeouts, and human-in-the-loop approvals. Workers run inside the customer's own environment while Mistral hosts the control plane, a split architecture designed to address data sovereignty requirements in regulated industries like banking and healthcare. Early adopters include ASML, La Banque Postale, and CMA-CGM. Mistral plans to release a more managed version for developers who need less infrastructure control, with further roadmap details expected at its AI Now Summit in Paris on May 27 and 28.

AI summary · Generated from this article

NVIDIA has released Nemotron 3 Nano Omni, an open multimodal model designed to replace fragmented enterprise AI stacks that rely on separate vision, audio, and language models. The 30-billion-parameter model unifies all three modalities into a single system, targeting agent workflows like document intelligence, computer use, and audio-video reasoning. The model uses a hybrid Mamba2 Transformer Mixture-of-Experts architecture that activates only 3 billion of its 30 billion parameters per forward pass, delivering inference economics closer to a small dense model. It supports a 131K token context window and achieves 9.2x higher throughput than comparable open omni models on video tasks. It leads six multimodal leaderboards and is available under NVIDIA's Open Model Agreement with full commercial rights. For teams running separate perception stacks today, this release is worth a serious evaluation. Open weights, published training recipes, and broad availability across AWS, Oracle Cloud, Hugging Face, and local runtimes like Ollama lower the barrier to adoption considerably.

AI summary · Generated from this article

Wiz researchers disclosed a critical command injection flaw in GitHub.com and GitHub Enterprise Server, tracked as CVE-2026-3854, that allowed any authenticated user to achieve remote code execution with a single git push command. The bug lives inside GitHub's internal X-Stat header protocol, where user-supplied push option values were embedded without sanitizing semicolons. Attackers could chain three injections to bypass sandboxing, redirect hook directories, and execute arbitrary commands as the git user on shared backend storage nodes. The cross-tenant exposure amplifies the severity significantly. Because GitHub.com runs on shared infrastructure, a compromised git user had filesystem access to every repository on the affected node, spanning millions of repos across unrelated organizations and users. GitHub patched GitHub.com within two hours of the report, but 88 percent of Enterprise Server instances remain unpatched and should upgrade to version 3.19.3 immediately. Wiz discovered the flaw using AI-augmented reverse engineering on closed-source binaries, surfacing a vulnerability GitHub's own engineers missed. This signals a meaningful shift in offensive security research…

AI summary · Generated from this article

Warp, the modern terminal used by nearly one million developers, has open sourced its client under the AGPLv3 license after five years as a closed-source product. The move is deliberate, timed to coincide with the rise of AI coding agents, and positions Warp as an agentic development environment rather than just a terminal. OpenAI is the founding sponsor of the open source repository. The contribution model is what makes this release technically distinct. Rather than simply accepting community pull requests, Warp has built a structured pipeline where its cloud agent platform, Oz, handles issue triage, spec writing, and initial implementation. Human reviewers at Warp make final decisions on what ships. The server infrastructure and Oz itself remain closed source, meaning the client is open but the orchestration engine powering it is not. Whether sustained community contribution follows the launch excitement is the real test. Teams building internal developer tooling should study the agent-first contribution architecture regardless, as it represents a concrete and documented approach to scaling open source development with AI agents.

AI summary · Generated from this article

Cognition has introduced Devin for Terminal, a local coding agent that runs directly inside your terminal and works with your local files. Unlike the full cloud version of Devin, this one is built for faster, more interactive coding tasks like bug fixes, feature work, code review, and quick automation. The big idea is that Devin now has two workflows. Devin Cloud is still meant for longer async tasks that run in a virtual machine, while Devin for Terminal is for hands-on coding sessions where you want the agent close to your actual project environment. It also includes Skills, which let users create custom slash commands using Markdown files. The strongest angle is the local-to-cloud handoff. You can start locally, let Devin understand the codebase and shape the plan, then escalate bigger tasks to Devin Cloud when needed. That could make Devin for Terminal more than just another Claude Code or Codex CLI competitor, especially for teams already using Devin.

AI summary · Generated from this article

Lovable has launched a native mobile app for iOS and Android, bringing its AI-powered full-stack app builder beyond the desktop. The platform, known for tools like Agent Mode and multiplayer workspaces, now lets founders, designers, and non-technical builders prompt, iterate, and ship web applications from anywhere. The app is not a lightweight companion. It supports voice and text prompting, autonomous background agent execution, push notifications, and full cross-device project continuity. Lovable's agent independently reads codebases, writes code, runs tests, and resolves errors without supervision, claiming a 91 percent reduction in errors. The underlying stack remains React, TypeScript, Tailwind CSS, and Supabase, with GitHub export intact. For the broader AI development space, this signals a shift in how agentic coding is practiced. As agents handle more execution, the human role becomes direction and review, tasks that translate naturally to mobile. Competitors like Bolt.new and Replit now face a builder that competes on availability, not just capability.

AI summary · Generated from this article

Moonshot AI has released Kimi K2.6, a 1 trillion parameter open-weight model that directly challenges the dominance of closed-source AI coding tools like GPT-5.4 and Claude Opus. Built on a Mixture-of-Experts architecture with 32 billion active parameters per token, it supports a 262,144-token context window and is available on Hugging Face under a Modified MIT License for self-hosting and commercial use. On key benchmarks, K2.6 leads competitors in agentic coding tasks, scoring 58.6 on SWE-Bench Pro versus GPT-5.4's 57.7, and 54.0 on Humanity's Last Exam with tools. It sustains 12-hour continuous coding sessions across 4,000-plus tool calls and can coordinate swarms of up to 300 parallel agents. Pricing runs roughly 30 times cheaper than Claude Opus 4.6 at comparable workloads. For engineering teams running coding agents at scale, K2.6 represents a serious evaluation candidate. The gap between open-weight and closed-source frontier models is narrowing fast, and the cost difference alone makes ignoring it difficult to justify.

AI summary · Generated from this article

AWS has launched Claude Platform on AWS, bringing Anthropic's native Claude experience directly into existing AWS accounts without requiring Amazon Bedrock or separate vendor relationships. The move is part of a deepening partnership between the two companies, backed by Amazon's $25 billion investment in Anthropic and a $100 billion Anthropic commitment to AWS infrastructure over the next decade. The integration eliminates several long-standing enterprise friction points. Developers authenticate using existing IAM credentials, removing the need for separate Anthropic API keys. Billing consolidates onto the standard AWS invoice, and all Claude activity is captured in CloudTrail for compliance and audit purposes. Importantly, this is Anthropic's first-party platform accessed through AWS entry points, not a Bedrock-managed abstraction. The key trade-off to understand is data residency. Claude Platform on AWS processes data through Anthropic outside the AWS boundary, while Claude on Bedrock keeps data within AWS infrastructure. Organizations with strict regional data requirements should stick with Bedrock.

AI summary · Generated from this article

OpenAI released GPT-5.5, a new frontier model focused on stronger coding, computer use, agentic work, and research, alongside a higher-accuracy GPT-5.5 Pro tier. It scored 82.7% on Terminal-Bench 2.0 and supports 1 million API context tokens. OpenAI claims roughly 40% lower output-token usage on comparable Codex tasks. API access is “coming soon” at $5 per million input tokens and $30 per million output tokens—double GPT-5.4’s rates—while Pro costs $30 and $180, respectively.

AI summary · Generated from this article

Michael Truell, a 25-year-old MIT dropout and former Google intern, co-founded Cursor alongside three MIT classmates after identifying limitations in Microsoft's GitHub Copilot. The AI-powered coding assistant includes a built-in development environment that autocompletes, reviews, and edits code while learning individual developer styles. Cursor reached $100 million in annualized revenue faster than Slack or Dropbox, crossing $2 billion by February 2026, with 67 percent of Fortune 500 companies now using the platform. On April 22, 2026, SpaceX announced a deal giving it the right to acquire Cursor for $60 billion or pay $10 billion for a joint development partnership. SpaceX intends to combine Cursor's developer reach with its Colossus supercomputer to build competitive AI models, while also positioning itself as an AI company ahead of a summer IPO targeting a $1.8 trillion valuation. The deal raises unresolved questions around model neutrality, enterprise data security, and Cursor's ties to Chinese AI infrastructure through its Composer model.

AI summary · Generated from this article

OpenAI released GPT-5.5 on April 23, 2026, positioning it as their most autonomous and capable flagship model yet. Arriving just seven weeks after GPT-5.4, the release reflects the accelerating pace of AI development. OpenAI describes it as their smartest model to date, built to handle complex, multi-step tasks with minimal human guidance across coding, scientific research, and enterprise workflows. GPT-5.5 supports a 1 million token context window and was co-designed with NVIDIA GB200 and GB300 hardware. It leads across 14 benchmarks, including 82.7% on Terminal-Bench 2.0 and 98.0% on Tau2-bench Telecom. Priced at $5 per million input tokens, OpenAI argues token efficiency makes it cheaper per task despite the higher rate. The model is available via API and in ChatGPT for Plus, Pro, Business, and Enterprise users. OpenAI has flagged GPT-5.5 as High risk under its Preparedness Framework for cybersecurity and biological capabilities, making its broad release a notable industry moment worth watching closely.