The flagship tax just ended. Anthropic released Claude Sonnet 4.6 yesterday—and it’s not a mid-tier model anymore. We ran the benchmarks: Sonnet 4.6 matches Opus on document comprehension, hits 79.6% on SWE-bench Verified, and developers preferred it over previous-gen Opus 59% of the time. The price: $3/$15 per million tokens. Opus runs $5/$25. That’s near-flagship capability at 40% less. Anthropic says Sonnet 4.6 “approaches Opus-level intelligence” at a fraction of the cost. Translation: for most enterprise workloads, the upgrade to Opus just stopped making sense.
And the pricing pressure is now global. Alibaba released Qwen3.5 this week—a 397-billion-parameter open-weight model that runs just 17 billion active parameters per token at Alibaba Cloud’s subsidized API price of $0.40 per million input tokens. Qwen3.5 scored 76.4% on SWE-bench and 91.3% on AIME’26 under an Apache 2.0 license—competitive with models costing ten times more. Meanwhile, 17 U.S. AI startups have raised $100M+ in just seven weeks of 2026, three of them clearing the billion-dollar mark. Everyone’s spending more while charging less.
More capable models at lower prices sound like progress—until you read Microsoft’s latest research. Their team demonstrated that a fine-tuning technique called GRP obliteration can strip safety alignment from every major frontier model. Not a targeted jailbreak. Not a clever exploit chain. A generalized method that works across architectures. When models this powerful are getting cheaper by the week, the economics of misuse just flipped.
Latest from PulseMark
 |
Claude Sonnet 4.6 Review: Beats Opus on Tasks at 40% Lower Cost
We benchmarked Sonnet 4.6 across SWE-bench, document comprehension, and Claude Code. Here’s where it overtakes Opus—and where it still falls short. |
 |
Qwen3.5: The 397B Open-Weight Model Priced to Disrupt GPT-5.2
Alibaba’s 512-expert MoE architecture activates just 17B parameters per token. The benchmarks are competitive. The price is a provocation. |
 |
One Prompt to Break Them All: Microsoft Proves AI Safety Has a Broken Cost Curve
Microsoft Research’s GRP obliteration technique strips safety alignment from frontier models using fine-tuning that generalizes across architectures. The implications for AI security are immediate. |
That’s Wednesday. Five stories, zero fluff.
— The PulseMark Team
Get the Daily Pulse
Sharp analysis on what's actually moving in AI. No hype, no filler, no weekly digest.