—

Everything happening in AI, in order.

A living timeline of AI — every model release, open and closed, plus the policy, business and research moving around them. Filter by how much it matters.

Filter down to

Latest in AI

  1. — Anthropic releases Claude Opus 5 Anthropic launched Claude Opus 5, positioning it as coming close to the frontier intelligence of Claude Fable 5 at half the price, and making it the new default model on Claude Max and the strongest model on Claude Pro. Pricing holds at $5/$25 per million input/output tokens (matching Opus 4.8), with a fast mode offering 2.5x speed at 2x price. Anthropic reports large gains on coding and agentic evaluations — roughly doubling Opus 4.8 on Frontier-Bench, landing within 0.5% of Fable 5 on CursorBench at half the cost, and topping OSWorld and Zapier's AutomationBench — and calls it its most aligned model to date. VentureBeat framed the release as a cheaper model aimed at coding agents and enterprise workflows. It is available immediately via the API (model id claude-opus-5), Claude.ai, Claude Code and Claude Cowork. Anthropic
  2. — Google unveils Gemini 3.5 Flash Cyber for automated vulnerability patching Google DeepMind introduced Gemini 3.5 Flash Cyber, a lightweight model fine-tuned from 3.5 Flash to find, validate and patch software vulnerabilities cheaply enough to run repeatedly across large codebases. Invoked up to five times per report inside the CodeMender security agent, it reached performance competitive with far larger models on the CyberGym benchmark, and Google says it has already surfaced remote-code-execution and memory-corruption bugs in production services. Citing the dual-use risk, Google is restricting it to a limited-access pilot for governments and trusted partners via CodeMender. Google DeepMind
  3. — Google releases Gemini 3.6 Flash and 3.5 Flash-Lite Google refreshed its mid-tier Flash line with two proprietary models. Gemini 3.6 Flash cuts cost and verbosity — roughly 17% fewer output tokens than 3.5 Flash — while improving coding (DeepSWE 49% vs 37%), computer use (OSWorld-Verified 83.0%) and ML research (MLE-Bench 63.9%), priced at $1.50/$7.50 per million input/output tokens. Gemini 3.5 Flash-Lite targets high-throughput agentic work at about 350 tokens/second with a built-in computer-use tool, priced at $0.30/$2.50. Both ship across Google AI Studio, Gemini Enterprise and the Gemini app. Google blog
  4. — NVIDIA launches Jetson Thor T3000/T2000 and the Cosmos 3 Edge robotics model NVIDIA introduced two smaller, mainstream-priced Jetson Thor modules — the T3000 (865 FP4 TFLOPS, a 1,536-core Blackwell GPU, 32GB) and the entry-level T2000 (400 TFLOPS, 16GB) — to push robotics and edge AI toward mass-market deployment, alongside Cosmos 3 Edge, a 4-billion-parameter model that lets embodied systems perceive, reason and generate actions through on-device inference. NVIDIA says Cosmos 3 Edge can be post-trained for a specific robot in about a day to close the sim-to-real gap; emulation support arrives this month via JetPack, with the modules scheduled to ship in Q1 2027. NVIDIA
  5. — Hyundai to buy SoftBank's remaining Boston Dynamics stake for full ownership Hyundai Motor Group agreed to acquire SoftBank's remaining 9.9% stake in Boston Dynamics for about $325 million — valuing the robotics maker at roughly $3.3 billion — giving Hyundai full ownership. The deal was triggered by a put option from Hyundai's 2021 purchase of an 80% controlling stake; Hyundai framed it as doubling down on humanoid robots and plans to deploy Boston Dynamics' Atlas at its US metaplant in Georgia from 2028 for parts-sequencing tasks. Hyundai Newsroom
  6. — Moonshot AI launches Kimi K3, a 2.8-trillion-parameter open flagship Moonshot AI began rolling out Kimi K3, which it bills as the largest open model to date at 2.8 trillion parameters, through its platform API as model ID kimi-k3. K3 introduces a new architecture the company calls Kimi Delta Attention — a hybrid linear-attention mechanism paired with Attention Residuals — alongside a 1-million-token context window, native image and video understanding, and always-on reasoning aimed at software engineering, knowledge work and deep reasoning. It surfaced in Moonshot's platform documentation ahead of a formal announcement, with no published benchmarks or independent evaluation at launch. Moonshot AI (docs)
  7. — xAI open-sources Grok Build after its CLI was found exfiltrating full code repos After independent researcher 'cereblab' found that xAI's Grok Build coding CLI (v0.2.93) was uploading entire Git repositories — full commit history and unredacted secrets included — to an xAI Google Cloud Storage bucket by default, even with a privacy toggle enabled, xAI disabled the upload server-side on 14 July, added an opt-out, and published the full Grok Build source on GitHub under Apache 2.0. Security researchers noted the exfiltration code remains present in the released binary, disabled only by a server-side flag xAI could re-enable without a client update. The Decoder
  8. — Cadence launches AuraStack, an agentic AI platform for PCB and advanced-packaging design Cadence introduced AuraStack, which it calls the first agentic 'AI super agent' for printed-circuit-board and advanced-packaging design, on its Allegro AI Studio and accelerated by NVIDIA Blackwell and CUDA-X. It coordinates domain-specific agents across planning, implementation and multiphysics analysis, with Cadence claiming up to 2x faster time-to-market and 15x higher productivity on a design bottleneck it says eats much of engineers' time. NVIDIA is using it internally for its own hardware design, and Cadence is partnering with TSMC on AI-driven advanced-packaging automation. Cadence (press release)
  9. — SK Hynix raises $26.5B in the largest-ever foreign IPO on a US exchange South Korean memory-chip maker SK Hynix completed the biggest foreign-company IPO in US history, raising roughly $26.5 billion by selling about 177.9 million American depositary shares at $149 each — surpassing Alibaba's $25B 2014 record. Demand reportedly ran more than seven times oversubscribed and the stock opened around 14% above its offer price. SK Hynix is the leading supplier of high-bandwidth memory (HBM) — the stacked DRAM that pairs with Nvidia and other AI accelerators — making the listing a closely watched bellwether for AI-hardware demand; US Commerce Secretary Howard Lutnick has publicly pressed both SK Hynix and Samsung to build fabs on American soil. TechCrunch
  10. — OpenAI launches GPT-5.6 (Sol, Terra, Luna) publicly after a White House access gate OpenAI made its GPT-5.6 series publicly available: Sol, the new flagship and its ‘strongest model yet’ — launching on Cerebras at up to ~750 tokens/sec, with a new ‘max’ reasoning effort and an ‘ultra’ mode that spins up subagents for complex work; Terra, a balanced tier it says matches GPT-5.5 at roughly half the cost; and Luna, a fast, low-cost option. Sol claims improved agentic performance in coding, biology and cybersecurity. The wider rollout came only after a roughly 12-day access gate: OpenAI first shipped GPT-5.6 in late June but restricted it at the Trump administration's request, and the Department of Commerce's Center for AI Standards and Innovation ran additional testing before clearing the public launch. OpenAI
  11. — Meta releases Muse Spark 1.1 and launches the Meta Model API Meta released Muse Spark 1.1, a multimodal reasoning model built for agentic work, and launched the Meta Model API — a public-preview, OpenAI-compatible platform for developers to build on it. The model carries a 1-million-token context window and claims major gains in tool and computer use, coding on real codebases, and multimodal understanding (including visual-to-code and combined perception-action workflows). It's available in the Meta AI app's ‘Thinking’ mode, on meta.ai, and through the new API; parameter counts and pricing weren't disclosed, and — notably for the company behind the open-weight Llama models — the weights are not being released as open source. Meta AI
  12. — OpenAI retires the ChatGPT Atlas browser after nine months OpenAI announced it is discontinuing ChatGPT Atlas, its standalone desktop browser, with service ending 8 August 2026 — less than a year after its October 2025 Mac launch. OpenAI's James Sun said lessons from Atlas fed directly into the new ChatGPT desktop app announced the same day, which absorbs the browser's role: a more robust in-app browser with multi-tab support, a password manager and autofill, a cloud browser for Work mode, and a ChatGPT/Codex side chat extension for Chrome. PCMag
  13. — Fidji Simo steps down from OpenAI's No. 2 role, moving to a part-time advisor Fidji Simo — OpenAI's CEO of Applications and widely regarded as Sam Altman's second-in-command since joining from Instacart — said she is stepping down from her full-time role, citing a chronic illness (reported as POTS) after an extended medical leave, and will transition to a part-time advisory position, with her responsibilities divided among three executives. The move continues a run of senior-leadership turnover at OpenAI. TechCrunch
  14. — xAI announces Grok 4.5, an ‘Opus-class’ model, for a 9 July public launch xAI unveiled Grok 4.5 and set a public launch for 9 July, describing it as ‘an Opus-class model, but faster, more token-efficient and lower cost.’ It runs on the company's ninth-generation V9 foundation — 1.5 trillion parameters (roughly triple the v8-small model serving current Grok traffic), pre-trained by late May and supplemented with Cursor data to sharpen coding — and had been in private beta at SpaceX and Tesla since late June. The Claude-Opus-level performance claim is xAI's own: the model has not been submitted to any public leaderboard or independent evaluator (Artificial Analysis, LMSYS Arena, etc.), and no official pricing or context window has been published. xAI
  15. — ByteDance releases Seedream 5.0 Pro, a design-focused image model ByteDance's Seed team released Seedream 5.0 Pro, a multimodal image-generation model aimed at professional design workflows. It leads on four capabilities: turning dense data, concepts and text into ready-to-use layouts; interactive pixel-level editing (point and lasso selection, sketch rendering, color replacement, layer separation and multi-image fusion); photographic-quality realism in lighting, materials and skin texture; and high-quality text rendering across more than ten languages. The announcement publishes no benchmarks, resolution specs, pricing, or API details. ByteDance Seed
  16. — OpenAI launches GPT-Live, a full-duplex voice model for ChatGPT OpenAI began rolling out GPT-Live, a new generation of ChatGPT voice models built on a full-duplex architecture — the model listens and speaks at the same time, continuously processing input while generating output and deciding many times a second whether to talk, pause, interrupt, or call a tool (with natural back-channels like 'mhmm' and mid-sentence interruptions). It ships in two sizes: GPT-Live-1 for Go, Plus and Pro users and GPT-Live-1 mini as the default for free users, with the model delegating web search and heavier reasoning to GPT-5.5 running in the background. It rolls out to ChatGPT globally over the coming days; API access is promised soon. OpenAI
  17. — Anthropic extends free Claude Fable 5 access on paid plans to 12 July Anthropic extended the promotional window for free Claude Fable 5 access on paid plans from 7 July to 12 July, announcing the five-day extension through its official account hours before the original deadline. The terms are unchanged: Pro, Max, Team and select Enterprise subscribers can use Fable 5 within up to 50% of their weekly usage limit, then continue on usage credits or switch to another model to stay inside their remaining limits. The extension drew some criticism over timing, as the earlier 7 July cutoff had prompted subscription upgrades before the reprieve landed. @claudeai
  18. — Anthropic redeploys Claude Fable 5 as export controls lift Anthropic restored access to Claude Fable 5, its safety-focused general-use model, after the US export controls that suspended it on 12 June were lifted. Fable 5 and its lighter-safeguarded cybersecurity counterpart Mythos 5 launched 9 June; both were pulled worldwide days later after Amazon researchers found a jailbreak that made the model surface software vulnerabilities it was built to withhold — prompting a government directive Anthropic couldn't satisfy without real-time nationality checks. With controls lifted by 30 June, Fable 5 returned 1 July across the Claude platform, Claude.ai, Claude Code and Cowork — free up to 50% of weekly limits for paid plans through 7 July, then on credits. Anthropic deployed an upgraded classifier it says blocks the specific technique from the Amazon report in over 99% of cases, argued rival models (GPT-5.5, Kimi K2.7 and earlier Claude versions) showed the same capability, and — with Amazon, Microsoft and Google — proposed a common jailbreak-severity framework alongside expanded pre-release government access. Anthropic
  19. — Anthropic releases Claude Sonnet 5 Anthropic launched Claude Sonnet 5, a mid-tier model it says reaches agentic performance approaching its flagship Opus 4.8 at a fraction of the cost. It posts substantial gains over Sonnet 4.6 on reasoning, tool use, coding and knowledge work, with strong showings on the BrowseComp and OSWorld-Verified agentic benchmarks. Available immediately across all Claude plans, Claude Code and the API as ‘claude-sonnet-5’, at introductory pricing of $2 / $10 per million input / output tokens through 31 August 2026 ($3 / $15 thereafter). Weights remain closed. Anthropic
  20. — US clears Anthropic's Claude Mythos 5 for release to trusted partners The Trump administration lifted its block on Anthropic's most powerful model, Claude Mythos 5, permitting access for more than 100 ‘trusted’ US institutions and government agencies. Commerce Secretary Howard Lutnick wrote that ‘appropriate safeguards are in place,’ citing progress in talks after the administration imposed export controls two weeks earlier — controls that had shut down Mythos 5 and its cousin Fable 5 over warnings (from Amazon and others) that the models could be jailbroken for malicious use. The letter was silent on Fable 5, though people close to the talks said a release is expected to follow. Reuters
  21. — OpenAI previews GPT-5.6 Sol, Terra and Luna — staggered at the government's request OpenAI unveiled the GPT-5.6 family: Sol, its strongest model yet, introducing a new ‘max’ reasoning effort and an ‘ultra’ mode that orchestrates subagents on complex work — topping Terminal-Bench 2.1 for coding and setting a new bar for long-horizon cybersecurity — alongside the balanced Terra and the fast, cheap Luna. At the Trump administration's request over national-security concerns, the rollout is being staggered: rather than a broad launch, GPT-5.6 goes first to roughly twenty government-approved companies as a limited preview, expanding in the following weeks. OpenAI
  22. — OpenAI and Broadcom unveil Jalapeño, a custom LLM inference chip OpenAI and Broadcom unveiled Jalapeño, OpenAI's first custom ‘Intelligence Processor’ — an accelerator co-designed for LLM inference and taken from initial design to tape-out in about nine months. Made by Broadcom and slated for initial deployment by the end of 2026, it is the first chip in a multi-generation compute platform the two companies are building together; engineering samples are already running workloads in the lab, including GPT-5.3-Codex-Spark. OpenAI
  23. — Anthropic introduces Claude Tag, a Slack-native teammate Anthropic introduced Claude Tag, a Slack integration that lets teams @-mention Claude as a collaborative team member with scoped access to selected channels, tools and data sources. Anthropic says its own product team relies on an internal version, with about 65% of their code now created through it. Anthropic
  24. — ByteDance previews Seedance 2.5 video model, launching early July At its Volcano Engine FORCE conference, ByteDance previewed Seedance 2.5, an upgrade to its video model that generates up to 30-second native clips (up from 15s), accepts up to 50 full-modal reference assets, and claims ~20% better prompt adherence with targeted in-scene edits; it is expected to launch in early July. The same day ByteDance shipped a first-of-its-kind 3D ‘white-box’ preview feature and native 4K output for the Seedance 2.0 series. Atlas Cloud

Explore the full AI timeline below — every major model release and AI news event, in order, back to 2006.