The GPT-5.5 Pro ECI score isn’t new. This hub confirmed it at 159 on April 29. What the May 2026 Epoch index update adds isn’t a new top score. It’s a picture of the frontier that looks different from the one we had a month ago, not because the leader changed, but because the gap between the leader and the pack narrowed to the point where it no longer functions as a competitive moat.
This piece works through what that compression means, and what the efficiency trend behind it implies about how long any lead can hold.
The Index State, as of May 2026
Per Epoch AI’s independent evaluation, GPT-5.5 Pro scores 159 on the ECI, with a 90% confidence interval of approximately 155 to 161. GPT-5.5 is reported at 158. Gemini 3.1 Pro is reported at 157. Claude Opus 4.7 is reportedly ranked at approximately 156, according to Epoch AI’s May 2026 index update, placing it in rough parity with GPT-5.4, the prior generation leader.
A note on certainty: the GPT-5.5 Pro score (159) and FrontierMath results are confirmed via Epoch AI’s own publications. The Gemini 3.1 Pro and Claude Opus 4.The range is the picture; the precise rankings within it carry more uncertainty than the headline numbers suggest.
The ECI is a composite metric drawing on 37 benchmarks. GPT-5.5 Pro’s FrontierMath performance is the most granular confirmed data point available: 52% on Tiers 1 through 3, and 40% on Tier 4, both new records per Epoch AI’s reporting.
What a Three-Point Gap Actually Means on a 37-Benchmark Composite
Three points on a composite index sounds small because it is. The 90% confidence interval for GPT-5.5 Pro alone spans six points (155 to 161). That means the entire reported range from GPT-5.5 Pro at 159 down to Claude Opus 4.7 at roughly 156 falls within the measurement uncertainty band of the top model.
In practical terms: you cannot reliably say which of these four models is “better” from the index alone. The benchmarks don’t discriminate at this resolution. What you can say is that all four are operating at a capability level that, until recently, only the single top-ranked model reached.
That’s the clustering story. It isn’t a story about parity in the sense of identical capability, the models perform differently across specific task types, and practitioners building on them will notice those differences. It’s a story about composite benchmark rankings ceasing to be a useful selection signal for most enterprise use cases.
The Claude Opus 4.7 Position and What It Signals
The parity finding is notable for what it says about Anthropic’s competitive trajectory. Claude Opus 4.7 was confirmed as generally available on Amazon Bedrock as of late April; prior coverage on this hub noted its ranking as second on the ECI with “reportedly” framing. The May index update, if the ~156 score holds under further confirmation, would place Opus 4.7 in parity with GPT-5.4, not GPT-5.5 Pro, but the model that led the index one generation ago.
Reaching the prior generation’s top score with a current release is meaningful. It means Anthropic has closed the gap that existed when GPT-5.5 Pro first debuted. Whether Opus 4.7 closes that gap further with subsequent releases depends partly on the efficiency trend below.
The Efficiency Trend as the Strategic Variable
According to Epoch AI’s analysis, pre-training compute efficiency is improving at approximately 3.0x per year.It carries an “approximately, per Epoch AI’s analysis” qualifier accordingly.
If the figure is directionally accurate, the implication is significant. A 3.0x per year efficiency gain means a model trained at a given compute budget in 2026 achieves roughly three times the effective training value of the same budget in 2025. That rate of improvement is faster than the rate at which any single lab can maintain a capability advantage through raw compute alone.
Frontier models have historically been differentiated partly by training scale. If efficiency is improving rapidly enough that smaller compute budgets can achieve results comparable to larger ones from the prior year, the barrier to closing benchmark gaps drops. Labs that were behind on raw compute may catch up faster than prior trends suggested. Labs currently leading may find their leads compress faster than their planning cycles assumed.
This is the mechanism behind the clustering phenomenon. It’s worth watching whether the May 2026 index state is a snapshot of a temporary convergence or the beginning of a persistent pattern.
What Actually Differentiates Models in a Clustered Frontier
When the benchmark can’t tell you which model is better, something else has to. For enterprise buyers, that list is long and underappreciated.
Pricing. GPT-5.5 Pro, Claude Opus 4.7, and Gemini 3.1 Pro are priced differently across their API tiers and enterprise contracts. At production volume, small per-token differences become large line items.
Availability and SLA. Where a model runs, on which cloud, under what latency guarantees, and with what uptime commitment, these differ meaningfully across vendors and are not captured in ECI scores.
Safety tier and guardrail architecture. Anthropic’s Constitutional AI approach and OpenAI’s safety tier system produce models with different behavioral profiles in edge cases. Enterprises in regulated industries care about this. The benchmark doesn’t measure it.
Tooling and ecosystem. API maturity, SDK support, fine-tuning options, and integration with existing enterprise tooling vary significantly. A model that scores 157 on the ECI with excellent LangChain integration and production-grade batch processing may be more deployable than a 159-score model with thinner tooling support.
One practical consideration the index doesn’t address: latency at production scale. A high ECI score reflects accuracy on benchmark tasks, not throughput under load. Enterprise teams operating at volume should validate inference performance in their own environments, benchmark results don’t transfer directly to production latency characteristics.
The Forward Picture
The May 2026 index update confirms what the April briefs were beginning to show: the frontier has compressed. That compression has a mechanism (efficiency gains), a current state (four models within three points), and implications for how the competitive dynamic plays out going forward (differentiation on factors other than raw capability).
The question the next update will answer is whether this is a stable plateau or whether one model breaks away again. Given the efficiency trend, a breakaway would require a qualitative architectural advance, not just more compute. That’s a different kind of race than the one the industry ran in 2024 and 2025.