# Anthropic Builds Momentum, xAI Speeds Up, and OpenAI Balances Churn With Buildout

*By AI News Digest • April 18, 2026*

Anthropic pushed Claude Opus 4.7 from model upgrade into a workflow product, xAI signaled faster training and product iteration, and OpenAI's day split between leadership departures and giant infrastructure progress. New research also raised harder questions about AI dependence and enterprise ROI.

## Frontier product race

### Anthropic turns Claude Opus 4.7 into a workflow story

Anthropic introduced Claude Opus 4.7 as its most capable Opus model yet, saying it handles long-running tasks with more rigor, follows instructions more precisely, and verifies its own outputs before reporting back [^1]. Code Arena, which evaluates agentic coding on real workflows like building live websites and apps, ranked it #1 overall, +37 points over Opus 4.6 and +46 over the next non-Anthropic model, with #1 spots on the React and HTML leaderboards [^2]. Anthropic also launched Claude Design, a research-preview tool for making prototypes, slides, and one-pagers by talking to Claude, powered by Opus 4.7 [^3]. *Why it matters:* Anthropic is pairing model gains with a task-specific interface, and early expert feedback suggests the jump is noticeable in practice; Jeremy Howard said 4.7 is the first model that feels aligned with his work intent after five hours of use [^4].

### xAI emphasizes speed in both training and product shipping

Elon Musk said xAI has shipped a more recent 0.5T checkpoint, has a 1T model roughly five days from finishing initial training, and now has a model factory working well enough to produce improved base models about every two weeks [^5]. On the product side, Grok 4.3 beta is being presented as natively multimodal, able to turn a website screenshot into code, use an Ubuntu shell and persistent file layer to generate artifacts, and create PDFs, slides, and other file formats, while Musk describes it as an early beta that should improve almost daily [^6][^7][^8][^9][^10]. *Why it matters:* xAI is signaling that it wants faster model-training cadence and faster product iteration at the same time [^5][^9].

## OpenAI's split-screen moment

### Leadership exits arrive as Stargate keeps moving

Three OpenAI leaders left on April 17: Kevin Weil, VP for Science; Srinivas Narayanan, CTO of B2B Applications; and Bill Peebles, Head of Sora [^11]. In his departure note, Peebles said he was leaving after helping build Sora zero-to-one, citing early evidence of object permanence and a move to high-fidelity 1080p multi-shot generation seven months earlier than skeptics expected [^12]. At the same time, the $500 billion Stargate initiative shows visible development across all seven surveyed US sites and appears on track for 9+ GW by 2029 [^13]. *Why it matters:* The contrast is notable: leadership turnover is happening while OpenAI and its partners continue a multi-site compute buildout that Greg Brockman says is aimed at demand from the compute-powered economy [^14].

## Friction beneath adoption

### New study finds AI help can reduce persistence

A paper from MIT, Oxford, Carnegie Mellon, and other labs reports that AI assistance can improve performance at first but hurt independent problem-solving soon afterward [^15]. Across three math and reading experiments involving about 1.2K participants, people using a GPT-5-based assistant finished early questions faster but, after roughly 10 minutes without AI, solved less, stalled more, and quit sooner; the sharpest drop came when the model was used for direct answers rather than hints [^15][^16]. *Why it matters:* The finding suggests that how people use AI may matter as much as whether they use it at all [^15][^16][^15].

### AI budgets keep rising even as usage lags

Goldman Sachs says companies are blowing past AI inference budgets by orders of magnitude, with engineering inference costs nearing 10% of total headcount costs and potentially moving toward salary parity within several quarters [^17]. The same discussion cited KPMG data showing average planned 12-month AI spending of $178M in the US, $245M in Asia-Pacific, and $157M in EMEA, even as compute shortages and a 48% rise in GPU spot prices over two months push costs higher and enterprise data suggests eight in ten workers are still avoiding AI tools or not using them at all [^17]. *Why it matters:* The spending boom is becoming harder to separate from ROI pressure, and Gary Marcus argued that competitive fear is now part of the story [^17][^18].

## Research to watch

### MIT's wristband points to finer robot control

MIT engineers built a wristband that uses ultrasound to image wrist muscles and tendons, then uses AI to infer the position of all five fingers across 22 hand poses from a smartwatch-sized device [^19]. In demos, wearers used it to direct a robotic hand to play piano, and the team says it has tested the system across eight volunteers and is now collecting data from hundreds more users for surgical-robot training [^19]. *Why it matters:* This points to a more precise human-to-robot control interface, with the team describing it as a possible universal remote for robots and virtual worlds [^19].

---

### Sources

[^1]: [𝕏 post by @claudeai](https://x.com/claudeai/status/2044785261393977612)
[^2]: [𝕏 post by @arena](https://x.com/arena/status/2045177492936532029)
[^3]: [𝕏 post by @claudeai](https://x.com/claudeai/status/2045156267690213649)
[^4]: [𝕏 post by @jeremyphoward](https://x.com/jeremyphoward/status/2044942799511191559)
[^5]: [𝕏 post by @elonmusk](https://x.com/elonmusk/status/2045293644693770630)
[^6]: [𝕏 post by @XFreeze](https://x.com/XFreeze/status/2045265210429108578)
[^7]: [𝕏 post by @tetsuoai](https://x.com/tetsuoai/status/2045180268605260148)
[^8]: [𝕏 post by @techdevnotes](https://x.com/techdevnotes/status/2045092507881451802)
[^9]: [𝕏 post by @elonmusk](https://x.com/elonmusk/status/2045285099680252417)
[^10]: [𝕏 post by @techdevnotes](https://x.com/techdevnotes/status/2045076941519081921)
[^11]: [𝕏 post by @ZeffMax](https://x.com/ZeffMax/status/2045248266384838800)
[^12]: [𝕏 post by @billpeeb](https://x.com/billpeeb/status/2045225014807670949)
[^13]: [𝕏 post by @EpochAIResearch](https://x.com/EpochAIResearch/status/2045258390147088764)
[^14]: [𝕏 post by @gdb](https://x.com/gdb/status/2045279841482928271)
[^15]: [𝕏 post by @rohanpaul_ai](https://x.com/rohanpaul_ai/status/2044860162016760104)
[^16]: [𝕏 post by @GaryMarcus](https://x.com/GaryMarcus/status/2045156208424968371)
[^17]: [𝕏 post by @HedgieMarkets](https://x.com/HedgieMarkets/status/2045220809669546306)
[^18]: [𝕏 post by @GaryMarcus](https://x.com/GaryMarcus/status/2045361707326046377)
[^19]: [𝕏 post by @rowancheung](https://x.com/rowancheung/status/2045158931367072104)