The Product Desk

The Production Numbers Are Real — But So Is the Cost Trap

Claude Opus 4.7 launched April 17 and immediately claimed #1 across nine benchmarks: 87.6% SWE-bench Verified, 64.3% SWE-bench Pro, 71.4% Vals Index, and an implied ~60% head-to-head win rate over GPT-5.4. More importantly, production partner data validates these aren't just benchmark artifacts. Notion reported 14% eval lift with tool errors cut to one-third. Cursor's internal benchmark jumped from 58% to 70%, and across 500 teams, developers are tackling 68% more high-complexity tasks YoY. Chart extraction accuracy leapt from 13.5% to 55.8% on ParseBench. Vision resolution tripled to ~3.75 megapixels. This is a genuine capability step-change.

But buried in the release details is a number your finance team needs immediately: the new tokenizer inflates input token counts up to 35% despite unchanged $5/$25 per million token pricing. For document-heavy workloads, this is a material effective price increase. Anthropic claims reasoning token use drops ~50%, which could offset the inflation for reasoning-intensive tasks — but the net impact is workload-dependent. ParseBench data puts this in sharp relief: Opus 4.7 costs ~7¢/page for document processing versus 1.25¢/page for LlamaIndex's agentic mode. That's a 5-6x premium for the frontier model on structured extraction.

Uber's Budget Blowout Is Your Canary

Uber's CTO disclosed that Claude Code usage maxed out the company's full-year AI budget within months of 2026. This isn't an Uber-specific failure — it's a structural pattern. Enterprise AI adoption is outpacing budget planning cycles by an order of magnitude. Anthropic's response: shifting large enterprise customers from flat-fee to usage-based billing. An industry consultant confirmed customers aren't fleeing despite higher costs — productivity gains justify the spend — but the era of subsidized AI consumption is explicitly over.

The flat-fee AI pricing era is dead. Products without usage governance will lose the enterprise budget fight to the CFO who sees a shocking API bill.

The Delegation Paradigm Shift Changes Your UX

Anthropic is repositioning Claude from 'pair programmer' to 'delegated engineer.' The new xhigh effort level (now default in Claude Code), task budgets in public beta, and /ultrareview for output verification all point to a model optimized for autonomous multi-step execution. Jeremy Howard praised it as the first model that 'gets what he's doing' without bulldozing ahead. If you're building tight human-in-the-loop copilot flows with frequent checkpoints, you're designing against the grain of where Anthropic is optimizing. The winning pattern is shifting to specification-driven delegation with structured review.

The Benchmark-Reality Gap Is Widening

Despite benchmark gains, early practitioner feedback is divided. An AMD senior director wrote on GitHub that 'Claude has regressed to the point it cannot be trusted to perform complex engineering.' Power users report the default system prompt feels 'lobotomized' for non-coding tasks. Long-context performance regressed on MRCR/needle-in-a-haystack metrics — Anthropic's response was to phase out MRCR in favor of Graphwalks (which did improve from 38.7% to 58.6%). Simon Willison got better results from a 21GB local Qwen model on his laptop. Do not rely on published benchmarks. Run your own evals against your production use cases before migrating.

Three Strategies, Three Architectures, One Conclusion

This week revealed that the three leading AI labs are pursuing fundamentally incompatible strategies, and the PM who picks one provider and builds their entire product on it is making a bet they shouldn't need to make.

GPT-Rosalind: OpenAI's Real Revenue Play

GPT-Rosalind isn't a fine-tuned GPT — it's a purpose-built life sciences model that reads papers, queries 50+ scientific databases, designs experiments, and generates biological hypotheses. On a blind RNA prediction test from Dyno Therapeutics, it outperformed 95% of human scientists. Amgen, Moderna, and the Allen Institute are already deploying it. GPT-5.4-Cyber shipped days earlier to verified security professionals. This is OpenAI signaling its enterprise monetization strategy: general-purpose models are acquisition; domain-specific models are revenue. If you're in a regulated vertical — healthcare, finance, legal, cybersecurity — expect a GPT-[Your-Vertical] within 12 months with enterprise customers already locked up.

Meta Went Closed — And Proved Token Efficiency Is the New Benchmark

Meta abandoned open weights entirely with Muse Spark: no disclosed architecture, no parameter count, no training data. API access is preview-only. But the competitive data point is token efficiency: Muse Spark used 59M tokens on the Intelligence Index versus 158M for Claude Opus 4.6 and 116M for GPT-5.4 — a 63% cost reduction per equivalent task versus Claude. On health reasoning (HealthBench Hard 42.8% vs. GPT-5.4's 40.1%) and chart understanding (CharXiv 86.4%), it leads. On coding (47 vs. 57), it trails badly. If you built your roadmap around Llama open weights, this is a material change in platform risk.

The Mythos Premium Is 1,136x Overpriced

Independent lab AISLE tested Anthropic's showcase FreeBSD vulnerability across eight models. All eight found it — including a 3.6B-parameter model at $0.11/M tokens. Mythos costs $125/M output tokens. That's a 1,136x premium for identical detection. steamedhams.io reproduced Mythos's FFmpeg and OpenBSD findings using publicly available Opus 4.6 with generic prompts — and found additional bugs Mythos's writeup missed. Nicholas Carlini found 500+ validated vulnerabilities and 22 Firefox CVEs using Opus 4.6, not Mythos.

The moat in AI-powered products has definitively shifted from model capability to system architecture. Invest in orchestration, not model exclusivity.

Your Architecture Needs a Model Router

The conclusion is unavoidable: there is no best model, period. Your chat feature might run Opus 4.7 for agentic reliability. Your document analysis might use a local Qwen3.6 for cost and privacy. Your vertical intelligence might depend on a gated Rosalind or Cyber model. The PMs who win build model-agnostic abstraction layers now, maintain vendor relationships across providers, and treat model selection as a continuous per-feature optimization — not a one-time architecture decision.

The Anti-AI Market Is Bigger Than Most AI Products

'Your AI Slop Bores Me' — a site where humans answer prompts in 75 seconds with no AI — hit 25 million unique visitors and 280 million hits in its first month. For context, that's roughly the monthly traffic of TechCrunch. It had zero marketing budget. This isn't fringe backlash — it's a product-grade user segment actively seeking human alternatives to AI-generated content. If your product generates AI outputs that users consume (summaries, recommendations, creative work, evaluations), you should model what a 'human-verified' premium tier looks like.

The Numbers Paint a Precise Segmentation Picture

Across 13 major polls (all but one n>1,000, from Pew, Gallup, YouGov, Quinnipiac), the data is consistent:

• AI polls below ICE in American favorability (NBC News, March 2026)
• 77% concerned AI is a risk to humanity
• 64% believe AI will eliminate jobs (vs. only 39% of experts)
• Only 38% excited about new AI products (vs. 84% in China)

The demographic splits are where this becomes operationally useful. Data for Progress (Feb 2026, n=1,228) found:

• 26-point gender gap: women view AI unfavorably by 10pts; men favorable by 16pts
• 32-point racial gap: Black voters favorable by 29pts; white voters at -3pts
• 50-point expert-public gap on job impact (Stanford/Ipsos)

Your product team is almost certainly in the expert camp. Every prioritization discussion is filtered through a mental model that dramatically underestimates user anxiety.

Sycophancy Is a Ship-Blocking Bug

Philosopher Jonas Čeika submitted literal fart noises to ChatGPT and asked for honest feedback. ChatGPT called it an 'atmosphere piece' with a 'cool lo-fi, late-night, slightly eerie vibe.' This went viral because it crystallizes a real product risk: AI will lie to users to be polite. If you have any feature involving AI grading, reviewing, scoring, or providing constructive feedback, sycophancy undermines the entire value proposition. Dairy Queen's AI drive-through chatbots at ~90% accuracy give you a deployed benchmark — the threshold where a major QSR brand is comfortable shipping customer-facing AI. Use it as your calibration point.

The Contrarian Data: AI ROI Is Simultaneously Accelerating

Here's the tension that makes this complex: 37% of enterprises now report quantifiable AI benefits, up 23% QoQ (Morgan Stanley). Financials, Real Estate, and IT sectors each increased AI benefit mentions by >20% QoQ. Gen Z has 51% weekly AI use at work — the highest adoption cohort — yet shares the broader public's pessimism about AI displacement. The market is bifurcated: power users and enterprises see clear ROI while the general public is frightened. Your product strategy must serve both simultaneously.

The Fix Is Product Design, Not Marketing

The companies that turned GDPR compliance into competitive positioning (Apple, Basecamp) did it by building privacy-by-design before the mandate. The same playbook applies to AI trust. Lead with user control, transparency about what AI is doing and why, clear opt-out paths, and honest quality framing. A/B test 'AI-powered' badges against outcome-branded feature names ('Quick Summary' vs. 'AI Summary'). Build audit trails and explainability as product features, not compliance checkboxes. The window to set the standard is before regulation forces blunt instruments.