AI Daily — May 02, 2026

The tools are working too well, and nobody priced for it. Uber blew its annual AI budget in four months amid heavy use of Claude Code by engineers, Sam Altman publicly walked back the cash-transfer experiment he funded with $14 million of his own money, and Spotify decided that proving you're a human is now a professional credential. The frontier isn't about which model is smartest anymore — it's about who can afford to use it, who gets credit for what it makes, and who pays when it goes wrong.

• Verified by Spotify (Spotify): Authenticity badge distinguishing human artists from AI-generated profiles, covering 99%+ of artists listeners actively search for (as of April 30, 2026).
• Codex for Work (OpenAI): Codex update extending beyond coding into Microsoft/Google/Salesforce suites, with 42% faster computer-use in initial benchmarks reported in April 2026, in-app Office editing, and a Cowork-style planning UI.
• Claude Security (Anthropic): Repo vulnerability scanner powered by Opus 4.7, paired with new integrations into Blender, Adobe Creative Cloud, Ableton, Splice, and Canva.
• ARC-AGI-3 Failure Analysis Package (ARC Prize): Open-sourced replays and reasoning traces from 160 GPT-5.5 and Opus 4.7 attempts, mapping three distinct world-model failure modes.
• Cursor Agent Harness Notes (Cursor): Public methodology for tuning agent runtime — CursorBench, latency, token efficiency, and real-world code "keep rate" — signaling that the harness is now the product.

The most honest sentence any tech executive has uttered this year came from Uber CTO Praveen Neppalli Naga, who told The Information: "I'm back to the drawing board because the budget I thought I would need is blown away already."

The Information cites Anthropic's Claude Code as a major factor. according to The Information, Uber rolled out access in December 2025; by February 2026, usage had nearly doubled, climbing from 32% of developers in December 2025 to 63% as of February 2026. By April, the annual AI line item was gone. As of May 2, 2026, 95% of Uber engineers used AI tools monthly, and 70% of committed code came from AI. Individual engineers running Claude Code as an agent reported $500–$2,000 in monthly API bills in early 2026.

What changes if this pattern holds: every CFO at every Fortune 500 company is now updating their AI tooling forecast, and consumption-based pricing — which looked elegant in pilots — starts looking like a pre-cloud nightmare at scale. Uber is already evaluating OpenAI's Codex as an alternative, per Yahoo Finance reporting. The signal to watch is simple: if Codex wins Uber's bake-off, it's the first major enterprise defection from Anthropic's coding-agent dominance, and every other enterprise gets new leverage in pricing negotiations.

What failure looks like: enterprises slap on rate limits, departmental budgets, and procurement gates — and the productivity gains evaporate alongside the bills.

For two years, the story was simple: closed frontier models lead, open-source trails. That story is getting complicated.

According to AI benchmarking firm Artificial Analysis, three open-weight models released this week — Moonshot's Kimi K2.6, Xiaomi's MiMo V2.5 Pro, and DeepSeek V4 Pro — now score 52–54 on its Intelligence Index, against 57 for Gemini 3.1 Pro Preview and Claude Opus 4.7, and 60 for GPT-5.5. All three are trillion-parameter mixture-of-experts systems with permissive licenses. The AINews digest called DeepSeek V4 Pro the first open-weight model that genuinely feels comparable to Codex or Claude Code for multi-turn agentic work, noting a 1M-token context window and inference compute reduced to roughly a quarter of comparable models at long context.

What changes if this trajectory continues: enterprises spooked by Uber-style budget shocks start seriously evaluating self-hosted deployments as a hedge. The South China Morning Post reports investor capital is already rotating into Chinese chipmakers on the thesis that efficient local models will drive demand for domestic inference stacks.

What failure looks like: the remaining gap on the hardest tasks — frontier science reasoning, hallucination resistance — proves stubborn, and open weights stay a backup plan rather than the default. Watch enterprise procurement RFPs over the next quarter; if they start specifying open-weight fallbacks, the moat is already cracking.

This sounds like a music story. It's an AI policy story.

Spotify introduced a "Verified by Spotify" badge on April 30, designed to confirm that an artist profile represents an actual human career — concert dates, merchandise, linked socials, sustained listener engagement. Profiles primarily representing AI-generated music or AI personas will not be eligible. At launch, Spotify said more than 99% of artists listeners actively search for would be verified (as of April 30, 2026).

The context: last summer, an indie band called The Velvet Sundown crossed 1 million plays before pressure forced its operators to admit every track was AI-generated, per NBC News. That's not an edge case anymore.

What changes if Spotify eventually ties verification to royalty weighting: human artists get a structural payment advantage, and every other streaming platform — Apple Music, Amazon Music, YouTube Music — has to either follow or look like an AI dumping ground. What failure looks like: the badge becomes a meaningless checkmark that everyone gets, and AI-generated content continues to siphon royalty pools through bot-amplified streams. The signal to watch is whether Spotify announces any payment-tier consequences — without those, this is decoration.

The timing is uncomfortable. The man building the tools reshaping the labor market just publicly abandoned the policy most often proposed to cushion the blow.

Speaking with The Atlantic's CEO Nicholas Thompson, Altman said: "I no longer believe in universal basic income as much as I once did." This is a notable reversal — Altman personally contributed $14 million to a $60 million study giving low-income participants $1,000 a month for three years (the study, published in 2025). Researchers found overall spending rose, but no "direct evidence of improved access to healthcare or improvements to physical and mental health," per coverage in Inc. and AOL.

His proposed replacement: "collective ownership that could be in compute or in equities or something else." The man whose company controls the compute is now proposing compute as the redistribution vehicle. House progressives aren't convinced — Rep. Alexandria Ocasio-Cortez told Semafor she's "skeptical about their willingness to pay or incur the taxes necessary to sustain such proposals."

What changes if "compute as a public good" gains traction: the redistribution conversation moves from cash to access, and OpenAI sits at the center of the policy architecture for the displacement its products cause. What failure looks like — and the more probable path — is that "collective ownership" stays a vibe at conferences, the experiments quietly end, and Uber's 8,000-engineer productivity story has no policy counterweight at all.

Bloomberg reported on May 1 that the Defense Department struck new agreements with Nvidia, Microsoft, Amazon Web Services, and Reflection AI for use of advanced AI tools on classified military networks. Reflection AI is the headline here — a non-traditional vendor joining the classified circle alongside the cloud incumbents.

Notably absent: Anthropic, which Bloomberg reports has been in dispute with Defense over military-use guardrails. That omission is the story. Policy posture, not benchmark scores, is now determining who gets access to the highest-margin AI contracts in the world.

What changes if Reflection's inclusion converts into operational deployments: classified AI moves from access agreements to program budgets, and the revenue mix at frontier labs starts visibly tilting toward defense. What failure looks like: the agreements stay paperwork, vendors burn cycles on compliance with no shipped capability, and Anthropic's policy stance turns out to have cost it nothing. Watch for named operational use cases in the next two weeks — that's the line between procurement theater and actual deployment.

A CTO staring at a blown budget while his engineers cheerfully ship 70% of the codebase, a billionaire announcing that cash for the displaced doesn't work and proposing tokens for his own product instead, and a streaming platform issuing humanity certificates to musicians. The future is here, it just sent you an invoice for $2,000 and a green checkmark proving you exist. Onward.

Forward this to the friend whose company just rolled out Claude Code — they're going to want to be ready for the budget meeting.