AI Daily 17 Apr: Mythos Shakes DC, Opus 4.7 Goes Live

Anthropic’s most powerful AI model is too dangerous to release publicly but not too dangerous for JPMorgan, Apple and the Pentagon. Claude Opus 4.7 is available to everyone today, Google upgrades Deep Think for scientists, and Meta launches its first closed-source model.

Anthropic’s Mythos model moved into the political spotlight today as CEO Dario Amodei prepared to meet White House chief of staff Susie Wiles in a bid to resolve the company’s dispute with the Pentagon. The model at the centre of the standoff, Claude Mythos Preview, scored 93.9% on SWE-bench Verified and 94.6% on GPQA Diamond during testing, and was capable of finding thousands of zero-day vulnerabilities across every major operating system and browser. Anthropic concluded the model was too capable to release publicly and instead launched Project Glasswing, a controlled programme under which a select group of firms can use Mythos for defensive cybersecurity purposes. Partners include JPMorgan Chase, Amazon, Apple, Cisco, Google and Palo Alto Networks, which called the project a “game changer” for finding hidden vulnerabilities.

The White House is preparing to grant a modified version of Mythos to major federal agencies, with the Office of Management and Budget establishing protective protocols before deployment. Anthropic co-founder Jack Clark confirmed discussions with the Trump administration are ongoing, even after the Pentagon severed business ties with the company following a contract dispute earlier this year. The meeting scheduled for today signals a potential breakthrough in what Bloomberg has described as a high-stakes standoff between the most capable publicly known AI system and the US government’s appetite for it.

Anthropic also shipped Claude Opus 4.7 as a generally available model, positioning it as the most capable commercial Claude model for anyone who wants access today. Opus 4.7 brings notable improvements to advanced software engineering, with users reporting it handles the hardest coding tasks with greater reliability than its predecessor. Vision is also upgraded: the model processes images at substantially higher resolution. A new “xhigh” effort level sits between high and max, giving developers finer control over the trade-off between reasoning depth and response latency. Opus 4.7 is available across all Claude products, the API, Amazon Bedrock, Google Cloud Vertex AI and Microsoft Foundry, priced at $5 per million input tokens and $25 per million output tokens, unchanged from Opus 4.6.

Google DeepMind released a significant upgrade to Gemini 3 Deep Think, its specialised reasoning mode, broadening it from mathematics and competitive coding into wide scientific domains. The updated model achieves gold medal-level results on the written sections of the 2025 International Physics Olympiad and Chemistry Olympiad, and scores 50.5% on CMT-Benchmark, a measure of advanced theoretical physics reasoning. Google developed the upgrade in close partnership with working researchers to handle problems where data is messy, guardrails are absent and there is no single correct answer. A practical demonstration shows a user sketching an object by hand, uploading the image, and having Deep Think model the geometry and generate a file ready for 3D printing.

Access to the updated Deep Think is now available to Google AI Ultra subscribers through the Gemini app, with researchers and enterprises able to apply for early API access. The update follows the broader launch of Gemini 3.1 Pro, which posted 77.1% on ARC-AGI-2 and led on 13 of 16 standard benchmarks.

Meta launched Muse Spark, its first closed-source AI model, developed under the direction of chief AI officer Alexandr Wang, who joined the company nine months ago to lead Meta Superintelligence Labs. Formerly known by the internal codename “Avocado,” Muse Spark sits near the top of agentic benchmark rankings, trailing Gemini 3.1 Pro and GPT-5.4 but ahead of most other models in the field. Meta says improved training techniques and rebuilt infrastructure allowed it to produce a model as capable as a mid-size Llama 4 variant using an order of magnitude less compute. The launch places Meta firmly back in the frontier model race, and arrives alongside an upward revision of the company’s 2026 AI capital expenditure guidance to between $115 billion and $135 billion, roughly double its 2025 spending.

Cursor released version 3 of its AI coding editor alongside Composer 2, a proprietary frontier coding model built entirely in-house by the Anysphere team. Composer 2 scores 61.3 on CursorBench, a 39% improvement over the previous version, and runs at over 200 tokens per second using custom GPU kernels, becoming the default model in Cursor’s Auto mode. The release comes as Cursor reached $2 billion in annual recurring revenue in March, making it one of the fastest-scaling developer tools on record. Existing Business and Pro subscribers receive the upgrade at no extra cost, and the new model is immediately available in the editor.

Snap announced it is cutting 1,000 employees, representing 16% of its global workforce, attributing the reduction directly to advances in artificial intelligence. CEO Evan Spiegel confirmed in a letter to staff that AI now generates more than 65% of Snap’s new code, and that the company is reallocating resources to its highest-priority initiatives. Snap expects annual savings of more than $500 million as a result. The company’s shares rose roughly 7% on the announcement. Disney also disclosed layoffs in the same week, contributing to what analysts are describing as an accelerating AI-driven restructuring wave across the technology sector.

Here is everything else worth knowing from today’s AI news.

• Microsoft Copilot Studio multi-agent coordination is now generally available, bringing agent-to-agent orchestration across Microsoft 365, Fabric and Teams. Microsoft
• Meta has extended its Broadcom chip deal through 2029, covering future generations of its in-house MTIA accelerator as it scales infrastructure to support Muse Spark and beyond. TechWire Asia
• Windsurf Wave 13 adds Arena Mode, letting developers compare AI models side by side in a blind test directly inside the IDE. Codeium Blog
• Novo Nordisk has agreed a partnership with OpenAI to integrate AI across its business from drug discovery to supply chain, with full deployment planned by end of 2026. VentureBeat
• The Stanford AI Index 2026 report finds generative AI reached 53% population adoption within three years, faster than the personal computer or the internet. Stanford HAI
• The EU AI Act’s transparency rules, including mandatory labelling of AI-generated content, take effect on 2 August 2026. Businesses operating in the EU should review compliance ahead of the summer deadline. AI Act Timeline
• The AI-crypto sector now spans 919 projects with a combined market cap of $22.6 billion, according to data presented at RenderCon 2026 in Hollywood this week. The Block
• SMB employees save an average of 5.6 hours per week using AI tools, though managers save more than double what individual contributors save, according to a new Business.com survey. Business.com

This is a daily news update for informational purposes only. AI products and policies change rapidly. Verify details directly with providers before making decisions. Nothing here is financial or legal advice.

AI Daily is Cristoniq’s daily guide to developments in artificial intelligence — published every evening.