The numbers are close enough to matter.
According to Epoch AI’s Capabilities Index documentation, the ECI combines multiple benchmark evaluations onto a unified scale to enable direct model comparisons. The April 24, 2026 update places GPT-5.4 Pro at 158, Gemini 3.1 Pro at 157, Claude Opus 4.7 at 156, and Qwen3.5-Omni, developed by Alibaba, at 155. Claude Opus 4.7’s score is confirmed through an Epoch AI official source. Scores for GPT-5.4 Pro and Gemini 3.1 Pro carry qualified status while the primary index URL remains inaccessible, but no contradicting data exists and the general ranking pattern is corroborated through multiple channels. These are not preliminary numbers. This is the published state of the frontier.
Stanford’s 2026 AI Index frames it without ambiguity: the US-China AI performance gap has “effectively closed,” with models from both countries trading the top position on performance rankings multiple times over the past year. Stanford measures this as a 2.7 percentage point difference on benchmark aggregates, a margin that, in practical terms, is within normal evaluation variance.
That’s the data. Here’s the paradox.
The capital efficiency problem
US AI development has attracted investment at a scale that is genuinely without precedent in the technology industry. The specific funding figures circulating in this week’s trade coverage lack primary source confirmation and don’t appear here as facts. The directional reality, that US AI investment dwarfs Chinese AI investment by a substantial multiple, is not seriously disputed. The paradox is that despite this imbalance, the capability gap has closed.
This matters because the investment thesis for US AI dominance has always rested on a simple assumption: more compute, more data, more capital equals a more capable model. The ECI data, and Stanford’s independent framing of the same trend, suggests that assumption is under stress. Qwen3.5-Omni sits one point below Claude Opus 4.7 on a benchmark scale that aggregates multiple evaluations. One point. Whether that reflects genuine architectural efficiency gains from Chinese labs, different research priorities, or the inherent limits of benchmark discrimination at the frontier is an open question. What’s not open is the result.
GPT-5.4 Pro uses what OpenAI describes as “more compute to think harder”, a framing that underscores the continued US bet on scaling. The benchmark data shows that bet is producing the world’s highest-rated model. It also shows the margin is narrow.
What enterprise buyers actually need to decide
For the CIO or enterprise AI strategist evaluating model selection in April 2026, benchmark parity is both clarifying and complicating.
It’s clarifying because it removes a shortcut. You can no longer point to a substantial US capability lead and use that as the primary selection criterion. Four models within three ECI points are, for most enterprise use cases, functionally comparable on the tasks those benchmarks measure.
It’s complicating because the real differentiators are now the harder-to-evaluate factors:
Pricing and API economics. Model pricing has been volatile. OpenAI’s pricing moves this year, covered in earlier TJS reporting on the agentic pricing shift, illustrate how quickly the economics of a model choice can change. Enterprise teams signing multi-year commitments need to model pricing scenarios, not just current rates.
Regulatory and compliance posture. Qwen3.5-Omni scoring within three points of GPT-5.4 Pro does not mean Qwen3.5-Omni is an equivalent choice for a financial institution operating under EU AI Act requirements or a US defense contractor. The governance stack, data residency, audit trails, conformity assessment, vendor contractual commitments, remains materially different across these models. Benchmark parity is not regulatory equivalence.
Enterprise reliability and support. Benchmarks measure model performance on curated test sets. They do not measure uptime history, API stability, support response time, or how the vendor handles a failure in production. For enterprise teams, the support infrastructure around a model is often more consequential than the model’s score on SWE-Bench. No arXiv technical papers have been published for GPT-5.4 Pro or Gemini 3.1 Pro yet, an absence that matters for technical evaluation teams.
Supply chain and geopolitical risk. A Chinese-developed model at benchmark parity raises questions that have nothing to do with capability. For regulated industries, for US federal contractors, and for enterprises with sensitive IP, the vendor’s jurisdiction of incorporation is a material procurement factor. Benchmark scores are jurisdiction-neutral. Procurement decisions are not.
What the convergence means for investors
The capital efficiency question is not academic. If benchmark parity has arrived at a fraction of the US investment level, the question for investors is what the marginal return on additional US AI capital actually is.
There are two reasonable interpretations. The first is that benchmarks have hit a ceiling and don’t discriminate the real capability gap, that the ECI’s four-point spread understates meaningful differences in complex reasoning, context reliability, and production-scale performance. This is a defensible position, and it’s consistent with the observation that technical papers for these frontier models often reveal architectural differences that aggregate benchmarks obscure.
The second interpretation is that the frontier is genuinely converging, that the diminishing returns on compute scaling are real, and that the next phase of AI competition will be won on application-layer differentiation, enterprise integration depth, and regulatory trust, not on raw benchmark scores. This interpretation suggests the investment thesis needs updating.
Neither interpretation is settled by the ECI data alone. Both are now live questions that weren’t seriously entertained eighteen months ago.
The open question that matters most
Stanford’s framing is the right one to hold: the gap has effectively closed on benchmark measures, and models have traded places at the top. That’s meaningfully different from saying one country’s models are definitively better.
The next eighteen months will test which interpretation is right. Watch for arXiv papers from OpenAI and Google that provide technical transparency on GPT-5.4 Pro and Gemini 3.1 Pro architecture, their absence today is notable. Watch for the Epoch AI primary index URL to resolve and confirm or revise the specific scores. Watch for enterprise deployment data that shows whether benchmark-close models produce equally-close outcomes in production environments.
Enterprise buyers don’t need to bet on an interpretation today. They do need to stop making selection decisions based on a capability gap that the data no longer supports.