GPT-6 Inbound, Opus 4.7 Ships, OpenAI Targets Hackers

Wednesday, April 22, 2026 by stevekimoi
GPT-6 Inbound, Opus 4.7 Ships, OpenAI Targets Hackers

GPT-6 Inbound, Opus 4.7 Ships, OpenAI Targets Hackers

Last Week in AI, April 14–21, 2026

A dense week for model releases. Anthropic shipped Opus 4.7 on Thursday, then admitted the same day that their best model is still locked away. OpenAI released a security-tuned GPT variant and GPT-6 edged into launch range, while Anthropic went from performance complaints to a server outage to an $800B valuation offer in 72 hours. Under it all: a Stanford report confirming that most enterprise workers are still actively avoiding AI tools.

Key Takeaways

  • GPT-6 ("Spud") is imminent. Pre-training finished March 24. Polymarket gives it a 78% chance of launching before April 30. If the standard evaluation window holds, it drops within days, not weeks.
  • OpenAI released a cyber-permissive model. GPT-5.4-Cyber is fine-tuned for offensive/defensive security research, available to verified teams only. This is the template for how frontier capabilities get deployed to high-risk domains going forward.
  • Anthropic shipped Opus 4.7 and admitted it's not their best. Released April 16 with a 13% coding lift and 3x more production tasks resolved. Same price as Opus 4.6. But on the same day, Anthropic confirmed Claude Mythos still outperforms it and remains locked.
  • Anthropic is valued at $800B, on paper. Investor offers have arrived at more than double the February valuation of $350B, but Anthropic hasn't taken a new round. The gap between what investors will pay and what the company needs right now is itself a signal.
  • Worker resistance to AI is real and growing. 54% of enterprise workers bypassed their company's AI tools in the past 30 days and did the work manually. Another 33% haven't touched AI at all. That's ~87% of the workforce actively avoiding the technology their companies bought.
  • TSMC posted its fourth straight record quarter. Revenue beat forecasts with a 58% profit jump. The chip layer is the most reliable business in AI right now, regardless of which model wins.

GPT-6 and the Cybersecurity Bet

OpenAI Releases GPT-5.4-Cyber

On April 14, OpenAI released GPT-5.4-Cyber, a variant of GPT-5.4 specifically fine-tuned for cybersecurity tasks. Access is gated behind identity verification through its Trusted Access for Cyber programme, which is expanding to thousands of individual researchers and hundreds of security teams. The underlying implication: OpenAI is now willing to ship models trained on offensive techniques, as long as the recipient pool is verified. That's a significant policy shift. Builders working on security tooling should get on the waitlist now. This is the most capable openly-distributed security model from a frontier lab.

GPT-6 ("Spud") Window Opens

Pre-training on GPT-6, internally codenamed Spud, completed on March 24. Sam Altman described the launch as "a few weeks" away on that same date. Polymarket currently assigns a 78% probability of release by April 30, with the standard 3–6 week evaluation window putting the earliest possible date at April 14. This is not rumour. Altman confirmed training completion publicly. Watch for a surprise drop.

Anthropic Releases Opus 4.7, Then Admits Mythos Is Still Better

Claude Opus 4.7 Ships April 16

Anthropic released Claude Opus 4.7 on April 16, two months after Opus 4.6. The headline numbers: 13% improvement on coding benchmarks, 3x more production tasks resolved, high-resolution vision up to 3.75 megapixels, and a new tokenizer. Pricing stays flat at $5/$25 per million tokens, available across Claude API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. For developers doing heavy software engineering work, this is a meaningful step up. The reported improvement in autonomous task completion is the number that matters most for agentic use cases.

But Anthropic Concedes It's Not Their Ceiling

The same day Opus 4.7 launched, Axios reported that Anthropic explicitly acknowledged the new model trails Claude Mythos Preview in overall capability. Mythos, which generated 181 working browser exploits vs Opus 4.6's 2, remains under Project Glasswing, available only to vetted defensive cybersecurity partners. The public gets the second-best model while the most capable one stays locked. That's a deliberate policy choice, not a capacity constraint.

Anthropic: From Outage to $800 Billion

Performance Backlash, Then a Server Outage

April 14–15 was a rough stretch for Anthropic. A Fortune report surfaced a wave of user complaints about degraded Claude performance (slower responses, less thorough outputs) with accusations that Anthropic had quietly reduced compute allocation without transparency. The next day, Claude went down entirely for a period that Anthropic acknowledged publicly. Two bad days for a company that has staked its differentiation on reliability and trust.

Investors Offer $800 Billion

Despite the turbulence, investor appetite hasn't cooled. Anthropic has received offers valuing it at $800 billion, more than double the $350 billion it was valued at in February 2026. The company hasn't taken new funding, and MCP crossed 97 million installs in March with every major AI provider now shipping compatible tooling. The business case for Anthropic's infrastructure-layer bet is holding even when individual model performance draws scrutiny.

The Numbers That Frame Everything

Q1 2026 Was the Biggest VC Quarter Ever

Venture investors poured $300 billion into 6,000 startups in Q1, up 150%+ year over year and an all-time record. Four of the five largest venture rounds in history closed in Q1: OpenAI ($122B), Anthropic ($30B), xAI ($20B), and Waymo ($16B). Foundational AI startup funding in Q1 alone doubled all of 2025. The capital is flowing, but it's concentrating at the top.

TSMC Posts Fourth Record Quarter

TSMC's Q1 profit surged 58%, beating estimates and extending its streak of record quarters. AI chip demand is the floor holding up the entire ecosystem. Model companies come and go, but the compute layer keeps printing.

Quick Hits

  • Novo Nordisk + OpenAI: The pharma giant announced a strategic partnership with OpenAI to accelerate drug discovery and integrate AI across global operations by end of 2026. Another signal that frontier AI is entering regulated science pipelines, not just enterprise software.
  • Stanford AI Index: Only 10% of Americans say they're more excited than concerned about AI in daily life; among AI experts, that figure is 56%. The perception gap between builders and the public is at its widest.
  • Worker AI resistance: 54% of enterprise workers bypassed company AI tools in the past 30 days and did the work manually; 33% haven't used them at all. Mandates without trust-building aren't working.
  • Claude Code updates: New /tui fullscreen rendering, mobile push notifications, smarter resume and Remote Control support, plus MCP stability fixes.

The throughline this week: frontier AI is bifurcating. One tier of models ships publicly at competitive prices (Opus 4.7, GPT-5.4-Cyber), while the most capable versions stay locked behind access programmes. Builders on the public tier are getting genuinely better tools, but the gap between what's deployed and what exists is widening. GPT-6 is the next inflection point. If Polymarket's 78% probability holds, that gap resets within days.

Last Week in AI is published every Monday by the Lablab team.

S
Steve Kimoi

Software Developer