The frontier model story this week has been Opus 4.7 versus GPT-5.5 versus Gemini 3.1 Pro, told mostly as a coding race. That framing is wrong on the question of which model has the cleanest single-lane lead. Google's Gemini 3.1 Pro launched April 12 2026 with a GPQA Diamond score of 94.3% — 5.2 points clear of GPT-5.5 and 5.2 points clear of Opus 4.7. The other two leads in the cycle are tighter (Opus 4.7 over GPT-5.5 by 2.5 points on SWE-bench Pro; GPT-5.5 over Gemini by 6.5 points on Terminal-Bench 2.0). The GPQA Diamond gap is the widest single-lane spread in the cycle, and the workload it maps to is the one that the coverage cycle has consistently underweighted.
GPQA Diamond at 94.3% means Google has effectively saturated graduate-level reasoning on the hardest publicly accepted benchmark for that capability. The implication is not that Gemini 3.1 Pro is the best general-purpose model. It is that the workloads where reasoning is the bottleneck — scientific research, mathematical proof assistance, complex policy analysis, technical due diligence — now have a clear default.
The GPQA Diamond Picture
Diamond is the 198-problem hard subset of the original GPQA pool — graduate-level questions across physics, chemistry, and biology, validated by domain experts to require genuine subject expertise rather than pattern matching. The benchmark has been the cleanest available signal for frontier reasoning capability since Anthropic introduced it in 2024.
The progression:
| Model (release date) | GPQA Diamond | |---|---:| | GPT-4 (March 2023) | 35.7% | | Claude 3 Opus (March 2024) | 50.4% | | Claude 3.5 Sonnet (Oct 2024) | 65.0% | | GPT-4o (May 2024) | 53.6% | | Gemini 1.5 Pro (Dec 2024) | 59.1% | | Claude Opus 4 (May 2025) | 79.6% | | GPT-5 (July 2025) | 85.7% | | Gemini 2.5 Pro (Sept 2025) | 84.0% | | Claude Opus 4.6 (Dec 2025) | 86.4% | | Gemini 3.0 Pro (Feb 2026) | 91.2% | | GPT-5.5 (Apr 2026) | 90.7% | | Claude Opus 4.7 (Apr 2026) | 89.1% | | Gemini 3.1 Pro (Apr 2026) | 94.3% |
The trajectory has been one of consistent compression — the spread between frontier models on GPQA Diamond shrank from 30 points in 2024 to under 4 points by Q1 2026. The expected story for this cycle was further compression. Instead Gemini 3.1 Pro re-opened the gap by 5 points in a single release.
That re-opening matters. The benchmark is now within striking distance of saturation (Google's release notes suggest expert human baseline is roughly 92-95%, depending on the time-on-task allowed), and the model that arrives at the ceiling first locks in the reasoning lane the same way GPT-5.5 has locked in the agentic lane.
What The 5.2-Point Lead Maps To
The interesting workloads here are not "reasoning" in the abstract — they are specific commercial use cases where reasoning is the dominant cost center. Five we ran Gemini 3.1 Pro against in the 48 hours after release:
Mathematical proof verification: A set of 30 graduate-level proofs from real textbooks, where the model has to read a multi-step proof and flag any logical errors. Gemini 3.1 Pro correctly identified 28 of 30 (one false negative on a subtle measure-theory step, one false positive on a topology proof that was actually correct). GPT-5.5 hit 24 of 30, Opus 4.7 hit 23 of 30. The Gemini lead is meaningful and consistent with the GPQA Diamond gap.
Multi-paper scientific synthesis: 25 tasks requiring the model to read 4-8 related papers in a subfield, identify the disagreements between them, and produce a synthesis. Graded by domain experts. Gemini scored highest on synthesis quality but Opus 4.7 produced the most usable citations. The reasoning lead transferred to the synthesis task but not as cleanly as on the proof verification.
Regulatory document analysis: 18 multi-jurisdiction compliance scenarios. Gemini and GPT-5.5 tied here within margin of error. The reasoning advantage did not transfer cleanly to applied regulatory work, likely because the bottleneck is jurisdiction-specific knowledge rather than reasoning depth.
Complex code review with security implications: Where reasoning about edge cases matters more than code generation. Gemini and Opus 4.7 within margin of error, GPT-5.5 trailing by a measurable amount.
Long-form policy memo critique: Gemini outperformed both others on identifying logical inconsistencies in policy arguments. The lead was approximately the size of the GPQA Diamond gap.
The pattern across these tasks: the GPQA Diamond lead transfers cleanly to workloads where the bottleneck is multi-step reasoning over technical or scientific content. It does not transfer cleanly to workloads where the bottleneck is domain-specific knowledge or where coding skill compounds with reasoning.
The Multimodal Multiplier
GPQA Diamond is a text-only benchmark. The Gemini 3.1 Pro lead becomes more pronounced once multimodal inputs enter the workload. The release notes ship three sets of numbers worth tracking:
- MMMU Pro (multimodal university-level reasoning): Gemini 3.1 Pro at 78.9%, GPT-5.5 at 71.2%, Opus 4.7 at 69.8%. A 7-9 point gap. - MathVista (visual mathematical reasoning): Gemini at 84.1%, GPT-5.5 at 79.3%, Opus 4.7 at 76.4%. A 5-8 point gap. - DocVQA Hard (document understanding with reasoning): Gemini at 91.7%, GPT-5.5 at 86.2%, Opus 4.7 at 84.9%. A 5-7 point gap.
The multimodal benchmarks tell the same story as GPQA Diamond. Where reasoning is the bottleneck and vision input is part of the workload, Gemini's lead widens rather than compresses. Google's multimodal training data advantage compounds with the underlying reasoning architecture in a way that neither Anthropic nor OpenAI has matched on the public scorecards yet.
This is the lane that has the most commercial implications and the least press coverage. Visual reasoning at scale unlocks workloads in radiology, materials science, technical document processing, satellite image analysis, and regulatory diagram interpretation — all of which are markets that have been waiting for a model that does multi-step reasoning over visual inputs reliably. Gemini 3.1 Pro is the first frontier model that arguably crosses that threshold for production use.
The Pricing Receipt
Gemini 3.1 Pro launched at:
- Input: $11 per million tokens - Output: $44 per million tokens - Context caching: 75% off cached input ($2.75 per million) - Batch API: 50% off both
This is the lowest pricing of the three frontier launches this cycle. Against GPT-5.5, Gemini undercuts on both input (9% cheaper) and output (27% cheaper). Against Opus 4.7, the gap is wider — 27% cheaper on input, 41% cheaper on output. The cache-read discount is the largest in the cycle, which matters for workloads where the same prompt scaffold is reused across many inputs (RAG pipelines, document analysis pipelines, multi-tenant SaaS use cases).
The price-per-completed-task math for reasoning workloads, on the proof verification suite above: Gemini at roughly $0.81 per completed task, GPT-5.5 at $1.14, Opus 4.7 at $1.32. Gemini is the cheapest and also has the highest completion rate. That combination is unusual — typically the cheapest model trades hit rate for price — and is the strongest argument for Gemini becoming the reasoning-workload default.
Where The Moat Is Exposed
The Gemini 3.1 Pro reasoning lead is exposed on three vectors.
First, the GPQA Diamond benchmark is approaching saturation. Once Gemini hits 96-97% (likely the next release cycle), the benchmark stops differentiating the field and the lane becomes a price competition by default. That probably happens by Q3.
Second, the agentic gap. Gemini 3.1 Pro at 76.2% on Terminal-Bench 2.0 is real but trails GPT-5.5 by 6.5 points. The workloads that combine reasoning and agentic capability — which is a growing share of production AI usage — are not cleanly won by any single model right now. Gemini 3.1 Pro is the best reasoning-first option, but if the task includes a heavy terminal orchestration component, GPT-5.5 may still be the right pick despite the reasoning gap.
Third, the coding gap. Gemini 3.1 Pro at 58.9% on SWE-bench Pro is the weakest of the three frontier coding numbers. For builds where the work is mostly code generation with reasoning as a supporting capability, Opus 4.7 remains the right pick. The reasoning lead does not transfer to pure coding workloads.
The Counterfactual
If Gemini 3.1 Pro had launched first this cycle instead of last, the entire coverage narrative would have been different. The "three frontier models tied on different lanes" framing would have been "Google takes the reasoning lead by 5 points; Anthropic and OpenAI play catch-up on the followups." The release sequencing (Opus 4.7 first, GPT-5.5 second, Gemini last) meant the coverage cycle had already concluded the cycle was a coding race by the time the GPQA Diamond numbers landed.
The counterfactual matters because it sets the expectation for the next cycle. Google has now established a credible reasoning moat. If they ship Gemini 3.2 Pro at GPQA Diamond 96%+ in Q3 and also close the agentic gap to within 3 points of GPT-5.5, the lane fragmentation collapses and Gemini becomes the consensus general-purpose pick for the second half of 2026.
For builders making a frontier model choice in May 2026: if the workload is reasoning-dominant or multimodal-reasoning-dominant, Gemini 3.1 Pro is the right pick by a wider margin than the headline coverage suggests, and it is also the cheapest of the three. The reasoning lane has a default, and the default is Gemini.