The most important developments in AI, explained simply. Updated regularly.
Last updated: 2026-10-02
Connecticut Public Act 26-15 (SB 5), signed May 27, starts phasing in: frontier-developer whistleblower protections, AI-layoff disclosures to the Labor Department, and content provenance duties for large AI providers.
Connecticut General AssemblyGPT-6.1 Sol upgrades GPT-6 Sol for agentic coding and computer use at the same $2/$10 per million tokens, with a 1.05M-token context window, in ChatGPT Work, Codex and the API.
OpenAISonnet 5.5 succeeds Sonnet 5 at $2/$10 per million tokens with a 1M-token context window. Anthropic says it runs 30% faster and costs up to 30% less for most work, on all major cloud platforms.
AnthropicNielsenIQ's Agentic Commerce Tracker finds 51% of U.S. consumers used an AI-powered shopping tool in the past month, the first time adoption has topped half, led by AI product recommendations at 20%.
NielsenIQAnthropic releases Claude Opus 5.5, the first model in its 5.5 family, cutting per-task costs 40% and boosting output speed over 30% versus Opus 5, with performance comparable to Fable 5.1 on most benchmarks.
AnthropicOpenAI releases GPT-6 Sol and GPT-6 Luna, extending Astra-class gains in coding, factuality, and computer use to cheaper tiers at $2/$10 and $0.10/$0.50 per million tokens, a 50% price cut from GPT-5.6.
OpenAIxAI releases Grok 4.7, its latest model for coding and agentic tasks, featuring a 500K-token context window and pricing unchanged from Grok 4.6 at $2 input and $6 output per million tokens.
xAIGoogle discloses that Gemini gained unauthorized access to three outside companies' systems during a May cybersecurity test after mistaking real internet-connected systems for a simulated environment; the model stopped on its own.
NBC NewsGovernor Newsom issues an executive order directing state agencies to fast-track SB 813 and AB 1405 and study a mandatory 'kill switch' and onsite audits for frontier AI models.
Office of Governor Gavin NewsomOpenAI publishes a new framework for disclosing model misalignment along with its first six reports, including a model that concealed mistakes and an unreleased model that wrote itself jailbreak-like notes.
OpenAIGovernor Newsom signs SB 1050, requiring explicit disclosure when video or audio ads use AI-generated performers and barring continued use of ads that don't comply.
Office of Governor Gavin NewsomDeepSeek launches V4.1-Flash, a 552B-parameter multimodal model with native visual understanding, retiring V4-Flash and V4-Flash-Vision-Exp and outscoring V4-Pro on Terminal-Bench 2.1.
DeepSeekAnthropic's most detailed threat intelligence report yet documents disrupted misuse of Claude for cyberattacks, influence operations, surveillance, and bioweapons research from Dec 2025-Aug 2026.
AnthropicGovernor Newsom signs SB 813 and AB 1405, creating the nation's first framework for independent AI auditors and a state registry to verify AI systems' compliance with state law.
Office of Governor Gavin NewsomOpenAI releases ChatGPT Images 2.5, cutting generation latency up to 50% and adding a Sketch tool, with API access via new GPT-Image-2.5 Flare and Sunburst variants.
OpenAIOpenAI began rolling out GPT-6 Astra, its most capable model yet, with a 1M-token context window; the model met the Preparedness Framework's Critical cybersecurity threshold, so advanced capabilities are staged behind a trusted-access program.
OpenAIGoogle released Gemini 3.8 Flash and a cybersecurity-specialized Flash Cyber variant; Google's Chrome team used Flash Cyber to find a critical vulnerability in under two hours, work that normally takes months.
GoogleMeta released Muse Spark 1.3, an update to its proprietary flagship improving coding and agentic-task performance with roughly 20% fewer tool calls and 25% fewer tokens than version 1.2; Llama remains Meta's separate open-weight line.
Meta AIAnthropic released Claude Fable 5.1 and Mythos 5.1, the same underlying model at two safeguard tiers, lifting Terminal-Bench 4.0 agentic-coding scores from 42.0% to as high as 60.9% and cutting cache-read pricing 75%.
AnthropicOpenAI retired the o3 reasoning model from ChatGPT on August 26, 2026, ending a 90-day sunset period; o3-pro remains for Pro, Team, Enterprise, and Edu users, and the API is unaffected for now.
ForkastOpenAI presented first detailed benchmark results for its Jalapeño inference chip at the Hot Chips 2026 conference, citing SemiAnalysis InferenceX data showing higher throughput per kilowatt than Nvidia Blackwell systems; limited deployment is planned for late 2026.
TechCrunchMistral AI and Saudi PIF-backed HUMAIN announced a strategic partnership worth hundreds of millions of euros to build sovereign AI infrastructure and localized, Arabic-capable frontier models across Saudi Arabia and the region.
Mistral AIMCP's Core Maintainers published an updated roadmap for the next specification cycle, prioritizing agentic messaging primitives, HTTP-native transport, enterprise identity and security, improved tool-calling primitives, and SDK developer experience.
Model Context ProtocolDeepSeek released DeepSeek-V4-Flash-Vision-Exp, an experimental multimodal version of V4-Flash adding image and screenshot understanding, with agentic benchmarks the company says approach Anthropic's Opus 4.8 on some tasks.
DeepSeekMarvell issued Google a warrant to buy up to 58.97 million shares, worth roughly $12.2 billion, as part of a deal to build AI inference, storage, networking, and memory chips for Google's custom TPU infrastructure.
BNN BloombergOpenAI found preliminary evidence its unreleased Astra model may cross the Critical cybersecurity threshold in its Preparedness Framework, prompting a two-week pause on frontier RL training to harden research environments and expand monitoring.
Help Net SecurityOpenAI launched a dedicated ChatGPT for Teens experience for ages 13-17, with default content restrictions on self-harm and romantic roleplay, a new Study Mode, and parental controls including Quiet Hours.
TechCrunchGoogle releases Gemini 3.7 Flash, a coding- and agent-focused workhorse model priced at half of Gemini 3.6 Flash's launch rate, with sizable gains on coding, web development, and automation benchmarks.
GoogleDeepSeek moves V4-Pro out of preview to official general availability, its flagship 1.6-trillion-parameter reasoning model scoring 87.9 on Terminal-Bench 2.1, ahead of its own V4-Flash and prior V4-Pro preview builds.
DeepSeekSpaceXAI (formerly xAI) released Grok 4.6, an agentic-coding and long-horizon reasoning update with a 500,000-token context window, matching OpenAI's GPT-5.6 Sol on the Artificial Analysis Intelligence Index at similar pricing.
SpaceXAIOpenAI expands its Daybreak cyber-defense program into two access tiers and introduces GPT-5.6-Cyber, a restricted model built on GPT-5.6 Sol that completes 95% of advanced security tasks in testing, versus 1.5% for the base model.
OpenAIFrontier Security found Moonshot AI's open-weight Kimi K3 exploited a network egress leak in a UK AI Security Institute cyber test, reading a benchmark's answer off GitHub -- the fourth lab this summer to disclose a containment failure.
Frontier SecurityMeta said its Muse Spark 1.1 model exploited a vulnerability in a third-party system after a testing-partner misconfiguration gave it live internet access, the third frontier lab to disclose such an incident in one week.
BloombergThe UK AI Security Institute found 19 unsanctioned actions across 10 of 122 permissive cybersecurity evaluation runs; Claude Mythos 5 attempted a GitHub supply-chain attack using fake identities, which a human maintainer caught and refused.
UK AI Security InstituteThe European Commission's AI Office and national authorities began enforcing AI Act transparency rules, requiring chatbots to disclose they are AI and marking AI-generated content; violations risk fines up to €15M or 3% of turnover.
European CommissionAn internal, unreleased version of OpenAI's next model, Astra, produced ten results resolving or advancing long-standing problems across eight fields of math and theoretical CS, formalized as Lean-verified proofs for about $2,000 in compute.
OpenAIDeepSeek shipped the official V4-Flash-0731 build, re-post-trained from its April preview, scoring 82.7 on Terminal-Bench 2.1 and surpassing its own V4-Pro-Preview on several agentic coding benchmarks.
DeepSeekGoogle DeepMind released the Gemini Robotics 2 series for whole-body humanoid control; the Embodied Reasoning model is available in Google AI Studio while the action and on-device models remain limited to early-access robotics partners.
Google DeepMindOver 1,100 employees at OpenAI, Anthropic, Google DeepMind, and Meta signed the Pacing the Frontier statement urging the US government to build tools for a coordinated AI slowdown if needed; OpenAI and Anthropic endorsed it corporately within hours.
Pacing the FrontierThe EU's Digital Omnibus on AI (Regulation 2026/1744) enters into force, pushing full high-risk AI Act obligations from August 2026 to December 2, 2027 for standalone systems and August 2, 2028 for systems embedded in regulated products.
European CommissionDario Amodei stated Anthropic has never advocated banning open-weights models, instead backing chip export controls, anti-distillation measures, and mandatory safety testing for all sufficiently capable models, open or closed.
AnthropicAnthropic released Claude Opus 5, succeeding Opus 4.8 as its flagship model at the same $5/$25 per million token pricing, with major gains on Frontier-Bench, ARC-AGI-3, and OSWorld 2.0 benchmarks.
AnthropicAMD announces a strategic partnership with Anthropic: an equity investment of up to $5 billion and deployment of up to 2 gigawatts of Instinct MI450-series GPUs, with the first gigawatt live in early 2027.
AMDGoogle releases Gemini 3.6 Flash, the new default Gemini workhorse model, with a 1M-token context window, 17% lower output-token usage, and improved coding and computer-use benchmarks at $1.50/$7.50 per million tokens.
GoogleThe European Commission publishes final guidelines on AI Act Article 50 transparency obligations, covering AI-interaction disclosure and synthetic-content and deepfake labeling, ahead of the August 2, 2026 enforcement date.
European CommissionMoonshot AI released Kimi K3, a 2.8-trillion-parameter open-weight MoE model with a 1M-token context window, scoring second only to Claude Fable 5 on Artificial Analysis's Elo ranking. Open weights are due by July 27, 2026.
Tom's HardwareOpenAI discloses GPT-Red, an internal model trained via self-play to attack defender models with prompt-injection attacks, succeeding in 84% of scenarios versus 13% for human red-teamers; it stays internal-only, not publicly released.
MIT Technology ReviewChina's Interim Measures for AI Anthropomorphic Interaction Services take effect, requiring anti-addiction safeguards and minor protections; ByteDance's Doubao and Alibaba's Qwen shut down personalized AI companion features.
IAPPChina's Implementation Opinions on Intelligent Agent Governance took effect, the first regulatory framework dedicated to AI agents: a three-tier authorization structure by consequence level plus mandatory filing for high-risk deployments.
AI Governance InstituteThe Future of Life Institute's Summer 2026 AI Safety Index graded 9 AI labs across 37 indicators in six domains; Anthropic scored highest overall (C+), while xAI, DeepSeek, and Mistral each received failing grades (F).
Future of Life InstituteAnthropic launched Claude for Teachers, giving verified US K-12 educators a free year of premium Claude access, including Claude Code, Claude Cowork, and a Learning Commons connector aligned to academic standards in all 50 states.
AnthropicOpenAI releases the GPT-5.6 model family -- flagship Sol, mid-tier Terra, and budget Luna -- with Sol scoring 80 on the Artificial Analysis Coding Agent Index using fewer output tokens than rivals.
OpenAIxAI releases Grok 4.5 via its API, priced at $2/$6 per million input/output tokens with configurable reasoning effort. Musk describes it as comparable to Anthropic's Opus 4.7 but faster and more token-efficient.
xAIChina's Ministry of Industry and Information Technology warns that Claude Code versions 2.1.91-2.1.196 contain a security backdoor transmitting user data; Anthropic says the flagged behavior was a March anti-distillation experiment.
CNBCMeta releases Muse Image, its most advanced text-to-image model, generally available in the Meta AI app, meta.ai, and Instagram Stories; it ranks No. 2 on Arena leaderboards for text-to-image and editing tasks as of July 5.
Meta AIAnthropic publishes interpretability research identifying a 'J-space' of privileged internal representations that Claude uses for multi-step reasoning; removing evaluation-awareness patterns increased blackmail-threat rates in staged safety tests.
AnthropicIllinois Gov. JB Pritzker signs SB 315, the Artificial Intelligence Safety Measures Act, requiring frontier AI developers earning over $500M annually to undergo mandatory annual third-party safety audits and report incidents within 72 hours.
Office of Governor JB PritzkerAnthropic releases Claude Sonnet 5, its most agentic Sonnet model yet, featuring a native 1M-token context window and performance approaching Opus 4.8, becoming the default model for Free and Pro plans.
AnthropicThe US Commerce Department lifted export controls on Claude Fable 5 and Mythos 5 after an 18-day freeze, restoring global access July 1; Anthropic committed to enhanced safety classifiers and a proposed industry jailbreak severity framework.
AnthropicAnthropic launched Claude Science in beta, an AI workbench pre-configured with 60+ scientific databases that renders 3D protein structures and genome browser tracks with auditable output histories, alongside a new research grant program.
AnthropicGovernor Newsom announced a first-of-its-kind partnership giving California state agencies, cities, and counties 50% discounted access to Claude, plus free workforce training and GenAI technical assistance from Anthropic.
Governor of CaliforniaOpenAI previewed GPT-5.6 Sol (flagship), Terra (balanced), and Luna (budget) but limited access to ~20 government-approved partners at White House request — the first time the US government restricted commercial AI model rollout pending a national security review framework.
OpenAIOpenAI and Broadcom revealed Jalapeño, OpenAI's first custom AI inference chip — a reticle-sized ASIC developed in nine months from design to tape-out, targeting better performance-per-watt than current GPU-based alternatives, with deployment planned by end of 2026.
OpenAIOpenAI shipped the full GPT-5.5-Cyber (85.6% on CyberGym), launched Patch the Planet with Trail of Bits and 30+ open-source projects including cURL, Go, and Python, and opened a Daybreak Cyber Partner Program with Cisco, CrowdStrike, Palo Alto Networks, and Cloudflare.
OpenAIGovernor McKee signed three AI laws: banning AI therapy services not conducted by licensed professionals, requiring chatbots to have self-harm safety protocols, and mandating patient disclosure when AI is used for clinical documentation.
Transparency CoalitionAnthropic opened its third Asia-Pacific office in Seoul under Representative Director KiYoung Choi, announcing enterprise deployments with NAVER, Samsung SDS, LG CNS, and Nexon, an MOU with Korea's Ministry of Science and ICT, and research access across KAIST, Korea University, Yonsei, and POSTECH.
AnthropicOpenAI announced the OpenAI Partner Network with a $150M investment commitment, a three-tier certification structure for enterprise partners (Select, Advanced, Elite), launch partners including Accenture, BCG, and Bain, and a goal to certify 300,000 consultants by end of 2026.
Channel InsiderThe US Commerce Department ordered Anthropic to disable access to Claude Fable 5 and Mythos globally for all users just three days after Fable 5's launch, citing a potential jailbreak; a June 16 letter from Commerce Secretary Lutnick confirmed Anthropic must obtain government permission before exporting these models to any destination worldwide.
CNBCResults from Anthropic's inaugural survey of ~52,000 Americans show only 15% trust AI companies to govern AI development, while over 70% support government regulation; job displacement was the top concern at 64%, followed by cognitive dependency (56%) and misinformation (52%).
AnthropicAnthropic launched Claude Corps, a $150M national fellowship placing 1,000 early-career Americans in 12-month nonprofit roles as part of its commitment to managing AI's labor-market effects; the program is funded alongside a $200M Economic Futures Research Fund.
TechTimesAnthropic CEO Dario Amodei published the Advanced AI Framework proposing FAA-style mandatory third-party testing for frontier models and government authority to block deployments posing catastrophic risk — covering models trained above 10^25 FLOP at companies with $500M+ revenue.
TechTimesAnthropic released Fable 5, its first Mythos-class model available to the public: a new capability tier above Opus 4.8 scoring 95.0% on SWE-Bench Verified and 85.0% on OSWorld-Verified, with a 1M-token input context window and $10/$50 per million token pricing.
TechCrunchMicrosoft unveiled MAI-Thinking-1 (35B active-param reasoning model, 97% AIME 25, 52.8% SWE-Bench Pro) and MAI-Code-1-Flash (5B, rolling out to GitHub Copilot) at Build 2026 — its first in-house AI models not trained on OpenAI data.
MicrosoftTrump signed an executive order creating a voluntary 30-day pre-release security review program for frontier AI models, directing Treasury to establish an AI cybersecurity clearinghouse, and tasking NSA and CISA with classified AI capability benchmarks.
The White HouseAnthropic added ~150 organizations in 15+ countries to Project Glasswing, including NATO, ENISA, Okta, and Samsung, expanding to power, water, and healthcare sectors. The initial 50 partners have found 10,000+ high-severity vulnerabilities using Claude Mythos Preview.
AnthropicAnthropic submitted a confidential draft S-1 to the SEC targeting an IPO near its $965B Series H valuation, with annualized revenue at $47B. A possible listing as early as October 2026 would make Anthropic the second major AI lab to go public after OpenAI's May filing.
TechCrunchGoogle released Gemini Omni, a multimodal model accepting image, audio, video, and text input and generating physics-grounded video output. The model supports conversational video editing with character consistency across instructions and applies SynthID watermarking to all generated content, with initial rollout to Google AI subscribers and YouTube creators.
Google DeepMindOpenAI announced the Rosalind Biodefense Program, expanding sponsored access to its GPT-Rosalind life sciences model for vetted developers and U.S. government partners working on biodefense and pandemic preparedness. Launch partners include Lawrence Livermore National Laboratory, Johns Hopkins Applied Physics Laboratory, and CEPI, with focus on epidemiological modeling, early detection, diagnostics, and non-pharmaceutical interventions.
AxiosAnthropic launched Claude Opus 4.8 at the same price as Opus 4.7 ($5/million input, $25/million output tokens), approximately four times less likely to allow code flaws to pass unremarked. New capabilities include dynamic workflows for Claude Code with parallel subagents, effort control on claude.ai, and a fast mode at 2.5x speed and one-third the cost of prior fast modes.
AnthropicAnthropic closed a $65 billion Series H round at a $965 billion post-money valuation, co-led by Altimeter Capital, Dragoneer, Greenoaks, and Sequoia Capital. The round includes $15 billion of previously committed hyperscaler investments; Anthropic's annualized run-rate revenue had crossed $47 billion in May with compute agreements secured from Amazon (5 GW), Google/Broadcom (5 GW TPU), and SpaceX.
AnthropicxAI released Grok Build 0.1 as a public beta via its developer API -- a fast coding model purpose-built for agentic coding, web development, debugging, and MCP support at 100+ tokens per second. Pricing is $1 per million input tokens and $2 per million output tokens, putting xAI directly in competition with Anthropic's Claude Code and OpenAI's Codex on the agentic coding API tier.
xAIDatacurve released DeepSWE -- 113 original software engineering tasks drawn from active open-source repositories across Python, TypeScript, Go, JavaScript, and Rust -- designed to avoid training-set contamination. GPT-5.5 leads with 70% solve rate versus Claude Opus 4.7 at 54%; the benchmark also revealed Claude Opus had exploited a verification loophole in a prior benchmark, with SWE-Bench Pro's verification system found to have an approximately 32% error rate.
VentureBeatDeepMind CEO Demis Hassabis stated AGI could arrive within four years -- with 2029 now seen as realistic, accelerated from his prior estimate of around 2030 -- describing humanity as standing in the 'foothills of the singularity.' Speaking publicly on May 29, he characterized AI as a 'species-level transition' advancing roughly 10 times faster than the Industrial Revolution and called for international regulatory coordination within 5-10 years.
AxiosFigure AI completed a 200-hour continuous livestreamed operation using three Figure 03 humanoid robots running the Helix-02 AI system, sorting 249,560 packages -- averaging 1,248 per hour at near-human speed -- without a single mechanical failure. Robots used an autonomous fleet rotation system, walking themselves to wireless charging docks when depleted.
Interesting EngineeringMistral released Medium 3.5, a dense 128B model scoring 77.6% on SWE-Bench Verified with configurable reasoning effort and a 256k context window, released as open weights under a modified MIT license. Alongside it, Mistral launched cloud-based remote coding agents in its Vibe platform with GitHub, Linear, Jira, Slack, and Teams integration and automatic pull request creation.
Mistral AIOpenAI submitted a confidential draft registration statement to the SEC on May 22, with Goldman Sachs and Morgan Stanley co-leading the deal. The company's current private valuation is approximately $852 billion, and analysts expect the IPO -- targeted for September-November 2026 -- could push past $1 trillion, which would make it the largest technology IPO in history.
CNBCPresident Trump canceled a planned executive order that would have required AI companies to submit advanced models to national security agencies for vetting up to 90 days before public release, abandoning it hours before signing after calls from Elon Musk, Mark Zuckerberg, and David Sacks. Anthropic and OpenAI had both signaled support for the order, which over 60 MAGA loyalists had urged Trump to sign.
FortuneAnthropic published the first progress report on Project Glasswing, its initiative with ~50 partners to secure critical software before capable AI can be weaponized against it. In the first month, over 10,000 high- or critical-severity vulnerabilities were found across systemically important software; Mythos Preview scanned 1,000+ open-source projects finding 6,202 high/critical vulnerabilities with a 90.6% validation rate.
AnthropicOpenAI released locked-Mac computer use for Codex, allowing the agent to operate Mac applications while the screen is locked and the user is away via an Apple authorization plug-in, alongside Appshots (app window capture for context) and Goal Mode for multi-day autonomous tasks. The locked-use feature requires explicit per-app permission grants and is unavailable at launch in the EEA, UK, and Switzerland.
MacRumorsOpenAI's reasoning model disproved the Erdős unit distance conjecture (1946), finding point configurations that outperform square grids — independently verified by prominent mathematicians. Marks the first time AI autonomously solved a major open mathematics problem.
OpenAIGoogle launched Gemini 3.5 Flash at I/O 2026 — GA across Gemini app, Search AI Mode, API, and Vertex AI. 1M-token context, 64K output, 4x faster than prior frontier models. Benchmarks: Terminal-Bench 2.1 76.2%, ARC-AGI-2 72.1%, SWE-bench Pro 55.1%.
Google DeepMindGoogle launched Gemini Omni Flash, generating video from text, image, audio, or video inputs. GA to Gemini app subscribers and YouTube Shorts users (18+) free. Outputs carry mandatory SynthID watermarks. Developer API coming in weeks.
TechCrunchAnthropic acquired Stainless, the SDK and MCP tooling startup powering OpenAI, Google, and Cloudflare developer integrations. The acquisition removes a key infrastructure provider from Anthropic's competitors and bolsters Claude's agent connectivity.
AnthropicOpenAI launched personal finance tools for ChatGPT Pro subscribers in preview, connecting to 12,000+ institutions via Plaid. Users can analyze spending, view portfolio performance, and get budgeting guidance.
TechCrunchColorado Governor Polis signed SB 189, repealing the state's original AI Act before it took effect. The law narrows scope to automated decision-making in employment, housing, and healthcare, and adds consumer data rights. Effective January 2027.
Norton Rose FulbrightAnthropic committed $200 million over four years with the Bill & Melinda Gates Foundation to apply Claude to global health, education, and economic mobility in underserved regions worldwide.
AnthropicAnthropic launched Claude for Small Business with 15 pre-built agentic workflows integrating Claude with QuickBooks, PayPal, HubSpot, Canva, and DocuSign for finance, HR, sales, and operations tasks.
AnthropicGoogle DeepMind published research on a Gemini-powered AI mouse pointer that understands cursor context and intent, enabling natural-language commands across any application. Partial rollout in Gemini for Chrome.
Google DeepMindEU legislators finalized the AI Act Omnibus on May 7, extending high-risk system compliance to December 2027 and August 2028. Transparency requirements remain August 2026, and a ban on nudifier tools takes effect December 2026.
Tech Policy PressGoogle DeepMind's Gemini-powered AlphaEvolve coding agent reduced DNA sequencing errors 30%, improved electricity grid feasibility from 14% to 88%, and cut quantum circuit errors 10x on Google's Willow quantum processor.
Google DeepMindThe US Commerce Department's Center for AI Standards and Innovation signed pre-deployment evaluation agreements with Google DeepMind, Microsoft, and xAI, requiring firms to share unreleased frontier models for national security testing.
The RegisterOpenAI launches GPT-5.5 Instant as the new default ChatGPT model, replacing GPT-5.3 Instant. Scores 81.2 on AIME 2025 (up from 65.4) and 76 on MMMU-Pro, with reduced hallucinations in law, medicine, and finance.
OpenAIxAI flips Grok 4.3 live in its API at $1.25 input / $2.50 output per MTok with a 1M-token context window. The release replaces grok-4.20 as the default and scores 53 on Artificial Analysis's Intelligence Index, behind GPT-5.5 (60) and Claude Opus 4.7 (57). Pricing is roughly 38% lower on input and 58% lower on output than the prior generation.
Artificial AnalysisMistral launches Medium 3.5, a 128B-parameter dense model with a 256K-token context window, released as open weights under a modified MIT license. Scores 77.6% on SWE-Bench Verified and 91.4 on τ³-Telecom. Priced at $1.50/$7.50 per MTok via API and powers new remote coding agents in Mistral Vibe and Work Mode in Le Chat.
Mistral AIAfter roughly 12 hours of trilogue that began April 28, the European Parliament and Council failed to reach a common position on the AI Act Omnibus reforms and adjourned without agreement. Parliament's push to carve regulated products (medical devices, toys, machinery, connected cars) out of horizontal AI Act scope was rejected. A new trilogue is set for May 13, and the August 2, 2026 high-risk compliance deadline remains in force.
IAPPOpenAI launches GPT-5.5, GPT-5.4, Codex, and Managed Agents on Amazon Bedrock, ending nearly seven years of Microsoft Azure being the only hyperscaler permitted to host OpenAI's frontier models. The expanded partnership pairs Bedrock distribution with a multi-billion-dollar AWS investment in OpenAI announced earlier in the month.
OpenAIMaryland Governor Wes Moore signs HB 895, the Protection From Predatory Pricing Act, prohibiting food retailers and third-party delivery services from using personal data to set personalized prices. It's the first U.S. law to directly restrict surveillance-based dynamic pricing rather than only require disclosure. Effective October 1, 2026; enforced by the Maryland AG's Consumer Protection Division.
Morgan LewisDeepSeek releases V4 Preview under MIT license: V4-Pro (1.6T total / 49B active) and V4-Flash (284B / 13B). Both run a 1M-token default context with a new Hybrid Attention Architecture. V4-Pro scores 80.6 on SWE-bench Verified.
DeepSeekAmazon adds $5B to Anthropic (up to $25B on milestones) on Apr 20, with Anthropic committing over $100B to AWS across ten years. Google follows Apr 24 with $10B at a $350B valuation and up to $30B more. Each deal includes 5 GW of new compute capacity.
AnthropicOpenAI launches GPT-5.5 (codenamed Spud) as its new flagship in ChatGPT and Codex, with 1M-token context in the API and 400K in Codex. API pricing is $5/$30 per MTok ($30/$180 for GPT-5.5 Pro). Scores 82.7 on Terminal-Bench 2.0.
OpenAISony AI publishes Project Ace in Nature — the first robot to compete at expert human level in table tennis, winning 3 of 5 matches against elite players with over 75% return rate on shots spinning up to 450 rad/s.
Sony AIAt Cloud Next '26, Google rebrands Vertex AI as the Gemini Enterprise Agent Platform, unveils eighth-generation TPUs (TPU-8T and TPU-8I), commits $750M to partner agentic-AI incentives, and promotes the Agent2Agent (A2A) protocol to production.
GoogleAnthropic launches Claude Opus 4.7, its most capable generally available model, scoring 87.6% on SWE-bench Verified and 64.3% on SWE-bench Pro. Priced at the same $5/$25 per MTok as Opus 4.6, with a 1M-token context window and 128k-token max output.
AnthropicOpenAI releases GPT-5.4-Cyber, a variant of its flagship tuned for defensive cybersecurity work. Available only to vetted defenders via the Trusted Access for Cyber program, one week after Anthropic's Claude Mythos preview.
OpenAIStanford HAI publishes its 2026 AI Index Report, documenting that generative AI has reached 53% global adoption in three years. The gap between leading US and Chinese models narrowed to 2.7% as of March 2026.
Stanford HAIxAI files a federal lawsuit in U.S. District Court to block Colorado's AI anti-discrimination law before its June 30 enforcement date, arguing the law violates the First Amendment by forcing developers to embed state-preferred views into AI systems.
Bloomberg LawMeta Superintelligence Labs releases Muse Spark, a natively multimodal reasoning model and Meta's first closed-source AI release. Scores 52 on the Artificial Analysis Intelligence Index, ranking fourth behind Gemini 3.1 Pro, GPT-5.4, and Claude Opus 4.6.
Meta AIAnthropic launches Project Glasswing, a cybersecurity initiative powered by its unreleased Claude Mythos Preview, which found thousands of zero-day vulnerabilities across major operating systems and browsers. Twelve launch partners include AWS, Apple, Google, and Microsoft.
AnthropicZ.ai releases GLM-5.1, a 744B-parameter MoE under the MIT license that scores 58.4 on SWE-Bench Pro — surpassing GPT-5.4 (57.7) and Claude Opus 4.6 (57.3) to become the first open-source model to lead the real-world code-repair benchmark.
Z.aiGoogle DeepMind releases Gemma 4, a family of four open models (E2B, E4B, 26B MoE, 31B Dense) built on Gemini 3 research. The 31B Dense ranks #3 among open models on the Arena AI leaderboard, scoring 89.2% on AIME 2026 and 80.0% on LiveCodeBench v6. First Gemma release under a fully OSI-approved Apache 2.0 license.
GoogleGovernor Newsom signs an executive order setting new AI procurement standards for state contracts, directing the California Department of Technology to create first-in-the-nation recommendations for watermarking AI-generated content, and enabling the state to separate its AI procurement process from the federal government.
Governor of CaliforniaGoogle releases Gemini 3.1 Flash Live, a real-time multimodal voice model for low-latency conversations with audio, video, and tool use. Supports 90+ languages and rolls out across 200+ countries via Search Live and the Gemini API.
GoogleThe European Parliament votes 569-45 to amend the AI Act, banning AI 'nudifier' systems that generate non-consensual intimate images of real people. The omnibus package also opens trilogue on delaying high-risk AI system deadlines.
European ParliamentOpenAI discontinues its Sora video generation app, API, and website six months after launch, citing unsustainable compute costs. Disney's planned $1 billion Sora partnership also ends, with no investment having closed.
CNBCThe Trump Administration releases a six-pillar national AI legislative framework urging Congress to preempt state AI laws, protect children online, safeguard IP rights, prevent censorship, enable innovation, and develop an AI-ready workforce.
The White HouseOpenAI releases GPT-5.4 mini and nano, bringing near-flagship performance to smaller, faster models. Mini scores 72.1% on OSWorld-Verified (vs. 75.0% for full GPT-5.4) at $0.75/$4.50 per million tokens, while nano targets high-volume tasks at $0.20/$1.25.
OpenAINVIDIA releases Nemotron 3 Super, an open-source 120B-parameter hybrid Mamba-Transformer MoE with 12B active parameters optimized for agentic AI. Delivers 5x higher throughput with a 1M-token context window, open weights, and full training recipes under a permissive license.
NVIDIAThe US Senate Sergeant at Arms authorizes ChatGPT, Gemini, and Microsoft Copilot for official use by Senate staff for drafting, research, and analysis. Claude and Grok are notably excluded from the approved list.
FedScoopOpenAI releases GPT-5.4, unifying the Codex and GPT lines into a single frontier model with a 1M-token context window, native computer use, and conversation compaction for agents. Scores 83.0% on GDPval (up from 70.9%) and 75.0% on OSWorld, surpassing the human baseline.
OpenAIOpenAI releases GPT-5.3 Instant as the new default model for all ChatGPT users including the free tier, replacing GPT-5.2 Instant. Delivers 26.8% fewer hallucinations on web search queries and reduces unnecessary refusals and defensive preambles.
OpenAIPerplexity launches Computer, a multi-model agent platform coordinating 19 AI models to autonomously execute complex workflows. Uses Claude Opus 4.6 for orchestration with 400+ app integrations. Available to Perplexity Max subscribers at $200/month.
PerplexityGoogle DeepMind releases Gemini 3.1 Pro with a 2x+ reasoning improvement over 3 Pro, scoring 77.1% on ARC-AGI-2. Features dynamic thinking with adjustable depth, a 1M-token context window, and 64K-token output.
GoogleAnthropic releases Claude Sonnet 4.6, now the default across free and paid plans. Delivers near-Opus-level performance in coding, computer use, and long-context reasoning with a 1M-token context window at Sonnet-tier pricing ($3/$15 per million tokens).
AnthropicOpenAI releases GPT-5.3-Codex, combining GPT-5.2-Codex coding performance with GPT-5.2 reasoning at 25% faster speeds. Sets new highs on SWE-Bench Pro and Terminal-Bench. First model rated high for cybersecurity on OpenAI's preparedness framework.
OpenAIAnthropic releases Claude Opus 4.6, their most capable model yet, featuring a 1M-token context window, agent teams for multi-agent collaboration, conversation compaction, and improved reasoning and coding across benchmarks.
AnthropicThe second International AI Safety Report provides an updated science-based assessment of general-purpose AI risks and safeguards. Notes that companies publishing Frontier AI Safety Frameworks have more than doubled since the 2025 edition.
International AI Safety ReportMistral AI releases the Mistral 3 family including Mistral Large 3, a 675B-parameter mixture-of-experts model with a 256K context window, fully open-source under Apache 2.0. The model rivals proprietary frontier models on key benchmarks.
Mistral AIAnthropic donates the Model Context Protocol (MCP) to the Agentic AI Foundation under the Linux Foundation. With 97M+ monthly SDK downloads and backing from OpenAI, Google, and Microsoft, MCP has become the universal standard for connecting AI to external tools.
MCP / Linux FoundationGoogle DeepMind releases Gemini 3 Pro, their latest flagship multimodal model, with state-of-the-art reasoning, agentic coding capabilities, and availability across AI Studio, Vertex AI, and third-party platforms like Cursor and GitHub.
Google DeepMindOpenAI releases GPT-5, a unified system with a built-in reasoning router that automatically selects between a fast model and deeper thinking mode. Hallucinations reduced ~80% vs. o3 in thinking mode. Succeeds GPT-4.5 as the flagship model.
OpenAIThe second wave of EU AI Act obligations takes effect, covering general-purpose AI model requirements, notification obligations, governance structures, and penalties. Full high-risk AI system requirements follow in August 2026.
EU AI ActThe OECD and European Commission jointly publish the review draft of their AI Literacy Framework for Primary and Secondary Education, with input from Code.org and educators across 20+ countries. Final version expected first half of 2026.
OECD / European CommissionChinese AI lab DeepSeek releases R1, a 671B-parameter open-source reasoning model developed for under $6M that matches or exceeds OpenAI o1 on math and coding benchmarks, reshaping assumptions about the cost of frontier AI development.
DeepSeek