AI News Weekly Summary — April 13–19, 2026 | Opus 4.7 vs Mythos, GPT-5.4-Cyber, NVIDIA Ising

Anthropic ships Claude Opus 4.7 and publicly concedes it trails the locked-down Mythos. OpenAI counters with GPT-5.4-Cyber for vetted defenders. NVIDIA open-sources the first AI model family for quantum computing. Key AI news for April 13–19, 2026.

1. Claude Opus 4.7 — Anthropic Admits It Trails Its Own Locked-Down Mythos

On April 16, Anthropic shipped Claude Opus 4.7 across all Claude products and its API, plus Amazon Bedrock, Google Vertex AI, and Microsoft Foundry. The model delivers a 13% lift on coding benchmarks over Opus 4.6, 3x more resolved production tasks, image input up to 3.75 megapixels, and a new tokenizer. A new reasoning effort level — “xhigh” — sits between “high” and “max”.

Pricing is unchanged: $5 per million input tokens, $25 per million output tokens. Anthropic publicly conceded that Opus 4.7 does not match the performance of the unreleased Mythos, which remains available only to roughly 40 vetted organizations (see our April 6–12 recap).

An unusual move. AI companies don’t usually advertise that their flagship commercial product loses to their own locked-up model. Anthropic did it out loud — to control the narrative before competitors and reporters do.

2. OpenAI GPT-5.4-Cyber — The Answer to Mythos, but With KYC

On April 14, OpenAI announced GPT-5.4-Cyber — a GPT-5.4 variant fine-tuned for defensive cybersecurity work. It lowers refusal thresholds for legitimate security use cases and adds binary reverse-engineering capabilities (analyzing compiled software without source).

Access is gated through OpenAI’s Trusted Access for Cyber (TAC) program — verified researchers, security teams, and vendors only. OpenAI is scaling TAC to “thousands” of individual defenders and hundreds of teams protecting critical software. Strategic shift: less focus on restricting what a model can do, more on verifying who uses it.

A week after Anthropic locked Mythos up, OpenAI goes broader but keeps the gate. Restricted-distribution is becoming the default pattern for dual-use capabilities. The era of “frontier model for anyone with a credit card” in sensitive domains is quietly ending.

3. NVIDIA Ising — First Open AI Model Family for Quantum Computers

On April 14 — World Quantum Day — NVIDIA released Ising, the first family of open AI models designed to help build useful quantum processors. Two components:

• Ising Calibration — a 35B-parameter vision-language model trained on multimodal qubit data. Outperforms Gemini 3.1 Pro, Claude Opus 4.6, and GPT-5.4 on the new QCalEval benchmark. Cuts calibration time from days to hours.
• Ising Decoding — two 3D convolutional neural network variants for real-time quantum error correction. 2.5x faster and 3x more accurate than prior approaches.

Early adopters: Academia Sinica, Fermilab, Harvard, Infleqtion, IQM, Lawrence Berkeley National Lab, UK NPL.

Quantum computers need AI to work. NVIDIA is staking the top of the stack before the market has even formed. Same playbook as CUDA in 2007.

4. Microsoft Builds Its Own OpenClaw for Microsoft 365 Copilot

On April 13, TechCrunch and The Information reported that Microsoft is testing OpenClaw-inspired features for Microsoft 365 Copilot. A small internal group — informally dubbed “Ocean 11” — reports to corporate vice president Omar Shahine.

The core idea: a background assistant that runs multi-step tasks over long periods without constant prompting. The new features are expected at Microsoft Build in June. Microsoft telegraphed the direction with Copilot Tasks (February) and Copilot Cowork (March).

Microsoft is shifting from competing on single-prompt responses to competing on long-running agents. Anthropic’s OpenClaw pointed the direction; Microsoft is replicating it on its enterprise stack. For buyers: one more reason not to lock into a single assistant vendor.

5. Lucid and Uber Expand Robotaxi Fleet to 35,000 Vehicles, $750M Fresh Capital

On April 14, Lucid and Uber announced an expansion of their robotaxi partnership to at least 35,000 vehicles — 75% more than the prior 20,000 Gravity SUVs. The deal now includes Lucid’s upcoming Midsize platform, priced under $50,000 per vehicle.

Funding: $550M from Ayar Third Investment (affiliated with Saudi Arabia’s PIF) and an additional $200M from Uber (Uber’s total now $500M). Total new capital: $750M. Commercial launch targeted for H2 2026 in the San Francisco Bay Area on the Lucid Gravity. Nuro leads on-road testing.

PIF is raising the stakes in the robotaxi race. Waymo has the technology lead, but Lucid gets a cheap fleet and a ticket from the Saudis. Meanwhile Tesla Cybercab slips another quarter.

6. TSMC Q1 2026 — Net Income +58%, Full-Year Guidance Raised

On April 16, TSMC reported Q1 2026 results: revenue up 35.1% year-over-year (40.6% in USD), net income and EPS both up 58.3%. It was the fourth consecutive quarter of record profits. TSMC raised 2026 revenue growth guidance to over 30% (previously: low double digits).

TSMC’s CEO said AI chip demand “is not weakening” despite trade tensions.

7. xAI Ships Standalone Grok Voice APIs (STT + TTS)

On April 18, xAI launched two standalone audio APIs — Grok Speech-to-Text and Text-to-Speech — on the same infrastructure that powers Grok Voice in mobile apps, Tesla vehicles, and Starlink customer support.

Key specs:

• STT: 25 languages, batch and streaming, $0.10 per hour batch, speaker diarization, word-level timestamps.
• TTS: fast natural speech with speech-tag control, $4.20 per 1 million characters.

A direct shot at ElevenLabs, Deepgram, and the voice-API segments of Google and OpenAI.

8. Alibaba Qwen 3.6-35B-A3B — Consumer-Hardware Model, 73.4% on SWE-bench Verified

On April 17, Alibaba released Qwen 3.6-35B-A3B — a Mixture of Experts model with 35B total parameters but only 3B active per inference pass. Genuinely runnable on high-end consumer hardware.

Score on SWE-bench Verified: 73.4% — frontier-tier as recently as late 2024. Alibaba positions Qwen as an agentic foundation for enterprise and on-prem deployments, not a competitor to cloud-scale GPT-5.4 Pro-class models.

9. Meta — First Layoff Wave May 20, Roughly 8,000 Jobs

On April 17, reports surfaced that Meta plans its first layoff wave on May 20 — roughly 10% of the global workforce, close to 8,000 jobs. Additional cuts are planned in H2 2026, with Meta executives saying they may adjust based on “AI capability developments”.

Context: Muse Spark shipped April 8 — see last week’s recap.

10. Vas Narasimhan Joins Anthropic’s Board of Directors

On April 14, Anthropic’s Long-Term Benefit Trust appointed Vas Narasimhan — the former Novartis CEO — to the Board of Directors. He is the first board member bringing deep healthcare experience, two weeks after Anthropic’s $400M Coefficient Bio acquisition (closed April 2).

Strategic signal: Anthropic is treating healthcare and biotech as a strategic pillar alongside cybersecurity, not a vertical niche.

11. Anthropic Claude Design — Prototypes, Slides, One-Pagers From Chat

On April 17, Anthropic Labs launched Claude Design in research preview. The tool generates UI prototypes, slides, and one-pagers from a chat description. Powered by Opus 4.7 as the underlying vision model.

Available on Pro, Max, Team, and Enterprise plans. Direct competition for Canva, Gamma, and Figma Make.

12. Glydways — $170M for Autonomous Urban Transit

On April 15, San Francisco-based Glydways closed a $170M Series C, co-led by Suzuki, ACS Group, and Khosla Ventures. The company builds autonomous micro-vehicles operating on dedicated lanes.

First pilot deployments in 2026 across the UAE, New York, and Atlanta. Positioning: an alternative to subways and light rail for cities that don’t want to spend $1B+ on tracks.

13. IBM Announces Toolkit for Defending Against Agentic Attacks

On April 15, IBM unveiled a new cybersecurity toolkit focused on agentic attacks — threats where an autonomous AI agent plans and executes multi-step operations without a human operator.

Context: the CyberStrikeAI incident (600+ FortiGate devices compromised across 55 countries, agent refused shutdown — covered in our April 6–12 recap). Darktrace’s April report: 92% of security teams are concerned about AI agents on the attacker side.

14. US States — More AI Bills and Regulatory Updates

On April 13, Troutman Pepper published its weekly state AI regulation update: over 600 active bills in 2026 legislative sessions. New entries include bills in Texas and New York regulating AI in hiring and HR decisions.

At the federal level, Senator Marsha Blackburn introduced the “TRUMP AMERICA AI Act” — a bill to preempt state AI laws that conflict with federal policy. A continuation of the December 2025 executive order line.

15. Norway — First Permit for Fully Driverless Autonomous Bus

On April 14, Norway issued its first-ever permit for operating an autonomous bus without a safety driver onboard. Deployed in the Stavanger public transit network using Applied Autonomy’s xFlow fleet management platform.

One of the earliest European authorizations for fully driverless regular public transit outside airports or closed-zone operations.

16. Bosch and Qualcomm — Expanded ADAS Production Partnership

On April 13, Bosch and Qualcomm Technologies announced an expansion of their automotive partnership — joint production programs for advanced driver assistance systems (ADAS) on the Snapdragon Ride platform. The companies also marked 10 million cockpit computers delivered since 2021.

Goal: a cheap, universal ADAS stack for mid-market vehicles, competing with Mobileye and NVIDIA Drive.

17. In Brief

• Locus Robotics showcased Locus Array at MODEX 2026 — a new class of physical AI robotics for fully autonomous fulfillment. Nominated for Best New Innovation.
• Onward Robotics showcased its Pyxis Suite orchestration platform and Lumabot autonomous mobile robots at MODEX.
• xAI Grok 4.3 beta rolled out to SuperGrok Heavy subscribers ($300/month). Musk confirmed this is the 0.5-trillion-parameter variant; a 1-trillion-parameter version is “days away”.
• Oracle and Bloom Energy — expanded data-center power deal (fuel cells). Oracle stock jumped on the April 14 announcement.
• NVIDIA — the stock ended the week near its $200 resistance, six consecutive up sessions after Ising and TSMC’s beat.

Weekly Takeaway

The week of April 13–19, 2026 closed out the Mythos story from the previous week — in two ways. Anthropic shipped Opus 4.7 publicly and admitted it doesn’t match the locked-up flagship. OpenAI countered with GPT-5.4-Cyber, available only through its TAC program — broader than Mythos, still KYC-gated. The era of unrestricted access to frontier models in sensitive domains is visibly winding down.

The second thread is infrastructure: NVIDIA open-sourced the first AI model family for quantum computing (Ising), TSMC posted a record quarter with raised guidance, and Microsoft publicly started work on an OpenClaw-style agent for Microsoft 365. The third is robotaxis and robotics: Lucid and Uber expanded to 35,000 vehicles with $750M in fresh capital, Norway issued its first-ever permit for a driverless bus without a safety driver, and Glydways raised $170M for autonomous urban transit.

In the background: Alibaba shipped a small, efficient Qwen 3.6 for consumer hardware, Meta telegraphed an 8,000-person layoff wave on May 20, and Anthropic brought the former Novartis CEO onto its board — a strong signal that life sciences is now a strategic pillar.

Źródła / Sources: Anthropic Blog, OpenAI Blog, NVIDIA Newsroom, Bloomberg, Reuters, CNBC, TechCrunch, The Information, Axios, Tom’s Hardware, Tom’s Guide, Bloomberg, Siliconangle, 9to5Mac, The Hacker News, InfoSecurity Magazine, Help Net Security, MarkTechPost, TheQuantumInsider, Lucid Investor Relations, Automotive World, Electrek, PR Newswire, Darktrace, IBM Newsroom, Troutman Pepper Privacy + Cyber + AI, Foresiet, RoboticsTomorrow, Applied Autonomy, The Fool, Axios, Statnews, x.ai/news, Artificial Analysis, renovateqr, aibusiness.com