Why even Claude Opus 4.7 and GPT-5.5 fail at quantitative finance.

Claude Opus 4.7 delivers a step-change improvement in agentic coding over its predecessor. GPT-5.5 Pro achieves 95.2% on AIME 2025 olympiad math. Gemini 3.1 Pro processes 1M-token context windows for whole-codebase analysis. Yet all three struggle to break 68% accuracy on institutional quantitative finance workflows requiring temporal consistency, numerical precision, and low-SNR signal extraction from market microstructure.

That 22–27 point gap is not a rounding error. It is the reason building production financial AI requires fundamentally different infrastructure than what works for coding or academic reasoning. This article is a structured audit of why that gap has not closed despite unprecedented scaling — and how Nyquist addresses it via a specialised multi-agent architecture on NVIDIA NIM microservices.

Frontier leaderboards look spectacular from the outside. GPT-5.5 hit 94.6% on AIME 2025 and 74.9% on SWE-Bench Verified. Claude Opus 4.1 pushed SWE-Bench to 74.5% before Opus 4.7 continued that trajectory. On ARC-AGI-2, leading models have reached 52.9%. These are genuine capability advances — and they are almost entirely irrelevant to production quantitative finance.

According to QuantBench 2026, Scale SEAL's finance subset, LiveBench, and institutional testing by major hedge funds, the picture on real quant tasks looks materially different:

The breakdown by institutional task category is more telling — the aggregate score hides the structural failure modes:

The critical observation: Opus 4.7's "step-change in agentic coding" is measured on software-engineering tasks — GitHub issue resolution, refactoring, debugging. None of those benchmarks evaluate quantitative alpha generation or statistical inference on noisy financial time-series. Researchers specifically testing LLMs for look-ahead bias found Llama 3.1 and DeepSeek showed spectacular apparent trading returns on pre-cutoff data — but on post-cutoff data, performance collapsed by 22 percentage points or more. That is not model degradation. That is the difference between memorisation and inference.

Before examining each failure mode empirically, it is worth establishing why these problems are structural rather than addressable through additional compute or data.

The transformer architecture has a mathematically provable limitation on numerical composition. Research published at NeurIPS shows that transformer layers with log-precision arithmetic — which describes all production LLMs — can be simulated by constant-depth threshold circuits, placing them in the computational complexity class TC⁰. This means transformers cannot even accurately solve linear equalities for sufficiently large instances under standard complexity assumptions. A separate result using communication complexity proves transformer layers are fundamentally incapable of composing functions when the domains are large enough — with empirical evidence the failure appears at surprisingly small scales. These are not benchmark observations. They are formal proofs about the architecture.

The implication is direct: next-token prediction over a distribution of text does not converge toward correct arithmetic over floating-point price series. No amount of RLHF, extended-thinking tokens, or chain-of-thought prompting changes this — because the constraint is architectural, not training-data-based. As Two Sigma's Chief AI Innovation Officer put it in their Q1 2026 outlook: "AI is not a thing you deploy. It's becoming the operating system for how quantitative research and investing work" — meaning it serves as an augmentation layer, not a replacement for deterministic computation.

The broader industry data confirms the posture. NVIDIA's 2026 State of AI in Financial Services survey of 800+ professionals found 65% of firms actively use AI in production, 84% consider open-source models important to their strategy, and 89% report AI is increasing revenue or reducing costs — but the applications driving those numbers are document processing, fraud detection, AML, and customer service. Not alpha generation. Not portfolio construction. Not backtesting.

The aggregate gap decomposes into five distinct failure modes, each with a different root cause and a different fix. None is closed by another generation of scaling.

Mapped across the leading models, the failure profile is consistent — and so are the required fixes:

It is instructive to look at how the most sophisticated quantitative firms are responding to these limitations in 2026, because their behaviour reveals where the real value of LLMs lies — and where it does not. Two Sigma's 2026 AI outlook describes the shift precisely.

Their CTO explicitly warns against "buying the hype" and emphasises "critical review" of AI capabilities. The firm sees agentic AI as the "leading edge of transformation" but treats governance and safety controls as the primary unsolved problem.

The key insight from Two Sigma, D.E. Shaw, AQR and their peers is that LLMs are being deployed as research acceleration and interface layers, not direct alpha generators. That maps precisely to the task table above: document understanding (88–94%), natural-language-to-code translation (good), hypothesis generation (useful) — versus direct signal generation (18–26%) and backtesting (28–39%), which remain in deterministic quantitative infrastructure.

Instead of building custom GPU infrastructure requiring $3M–$6M annual operating costs plus $2M–$4M capital expenditure, Nyquist deploys production AI via NVIDIA NIM (NVIDIA Inference Microservices) — a containerised, enterprise-grade stack providing institutional-quality performance without the infrastructure burden. NIM is not managed inference; it is a complete optimised stack.

• Pre-optimised inference containers — each model packaged with TensorRT-LLM acceleration, Triton orchestration, and quantisation tuned per GPU architecture; production benchmarks show NIM leads vLLM, SGLang, and HuggingFace TGI on throughput.
• Multi-LoRA deployment support — multiple LoRA adapters per base model with automatic routing, critical for regime-specific inference.
• Standardised API layer — OpenAI-compatible REST plus gRPC for latency-critical applications.
• Enterprise SLAs — NVIDIA AI Enterprise at $4,500/GPU/year includes 24/7 support, guaranteed uptime, security patches, and SOC 2 Type II compliance.
• Model catalog — 90+ pre-optimised models (Llama, Mistral, DeepSeek, Qwen) plus domain-specific financial models.
• Deployment flexibility — on-prem (bare metal or VMware), public cloud (AWS SageMaker, Azure ML, GCP Vertex AI), or NVIDIA DGX Cloud.
• Automatic optimisation — FP8 quantisation, continuous batching, PagedAttention, speculative decoding — configured without manual tuning.

NVIDIA's 2026 financial-services survey found 89% of institutions report AI increasing revenue or reducing costs, with 84% saying open-source models are central to strategy. NIM is the production-grade bridge between open-source flexibility and enterprise deployment — the reason 65% of firms have moved from experimentation to active production. NVIDIA has published financial-services reference architectures Nyquist directly extends:

1. GPU-accelerated risk analytics — RAPIDS cuDF and cuML containerised in NIM, delivering 10–100× speedup on VaR and CVaR vs. CPU — from 15–30 minutes to sub-second.
2. AML and fraud detection — RAPIDS for Apache Spark achieves up to 20× improvement on entity resolution and transaction monitoring.
3. Intelligent document processing — RAG over 30 years of SEC EDGAR filings, earnings transcripts, and regulatory filings at <2-second response on 50M+ document corpora.
4. Agent-based trading orchestration — NeMo Agent Blueprint for multi-agent coordination with built-in circuit breakers, audit logging, and human-in-loop escalation.

Nyquist specifically extends architecture #4 as the foundation for the 36-agent fleet, adding the hybrid symbolic-neural compute layer that eliminates all LLM arithmetic.

Three structural reasons make general-purpose frontier models economically and operationally prohibitive at institutional quant scale: latency (1.5–8s with 1M-token context is unacceptable when downstream execution must finish in under 500ms), cost (Opus 4.7 at $15/M output tokens, GPT-5.5 Pro at $30/M; at 900K+ daily inferences, API costs alone exceed $5M/year), and capacity allocation (general models spend parameters on poetry and trivia instead of financial domain expertise — which is why a 13B domain-specific model can outperform a 405B general model on financial tasks, a pattern established by BloombergGPT).

FleetAgent classes and their deployment.

Each agent class targets a specific failure mode by shrinking the surface area: small, domain-trained models hit Sharpe-contribution targets that 405B general-purpose models structurally cannot.

ComputeThe hybrid symbolic-neural layer.

The most consequential architectural decision: all numerical operations are routed to deterministic symbolic compute engines. Zero arithmetic is performed by LLMs. The ESMA/Turing report on LLMs in finance (2025) notes that "hallucinations in an LLM that underpins [trading systems] might lead to reputational damage for the firm" — with financial calculation errors classified as materially more dangerous than NLP hallucinations. Tool-use approaches, where arithmetic is delegated to a Python interpreter rather than computed in natural language, are more robust regardless of how much financial text the model saw.

1. Agents generate computation plans in structured natural language ("calculate annualised Sharpe ratio for returns array X, 252 trading days").
2. Plans are parsed and executed by NumPy / pandas / SymPy with type-checked interfaces.
3. Results returned to agents for contextual interpretation and downstream decision.
4. All intermediate computation artefacts logged for regulatory audit.

This eliminates the 16–28% hallucination rate on financial calculations — not reduces it, eliminates it.

RoutingContinuous learning & regime-aware adapters.

Markets are non-stationary — models trained on 2023 data degrade measurably by 2026. Monthly LoRA retraining per agent class addresses distribution shift without full-retraining costs. FinLoRA (ICLR 2026) demonstrated LoRA methods achieve 40.1-point average accuracy gains over base models on financial tasks, with fine-tuning executable on 4× A5000 GPUs for under $100 per run. A Hidden Markov Model — not an LLM, running at sub-10ms latency — continuously monitors realised volatility, cross-asset correlation, and bid-ask dynamics to detect the current regime. Each agent class carries three regime-specific LoRA adapters:

• Low-volatility regime — slow mean-reversion signals, value-factor emphasis, low turnover.
• High-volatility regime — momentum and risk-off positioning, tighter VaR limits.
• Crisis regime — liquidity-constrained optimisation, correlation-breakdown handling, elevated human-in-loop escalation.

Keeping the regime detector out of the LLM ensures the regime signal itself is not subject to the temporal hallucination problems above. NIM's native multi-LoRA support routes each inference to the appropriate adapter without separate deployments or additional memory footprint.

NIM is not just a capability bridge — it is the difference between institutional-grade financial AI being accessible only to the top-10 hedge funds, or to the broader institutional market.

Break-even AUM using NIM is $80M–$120M; for custom infrastructure it is $250M–$400M. Operationally: 6–8 weeks to production vs. 6–9 months; 2–3 MLOps engineers vs. 7–12; monthly catalog updates allow adopting newly optimised models in hours. NIM also supports air-gapped on-prem deployment for regulated environments where data cannot leave the perimeter.

Q3 2026The 13B financial foundation model.

NIM provides deployment infrastructure; the fundamental pre-training gap requires a different solution. Nyquist is training a proprietary 13B-parameter financial foundation model for Q3 2026. The BloombergGPT experiment (2023) trained a 50B model on 363B tokens of proprietary financial data and achieved 25-point improvements over GPT-NeoX — BloombergGPT's 50B model outperformed GPT-3's 175B general model on financial tasks despite 3.5× fewer parameters. A 13B model trained on 15TB of financial-domain data is a more credible alpha-generation substrate than a 405B model trained primarily on Common Crawl.

• Training corpus (15TB) — 30 years of SEC EDGAR filings (50M+ docs, temporal versioning preserved); 25 years of earnings transcripts (2M+, speaker diarisation); 15 years of tick data across equities, futures, options, FX and crypto (10TB OHLCV + bid/ask, microsecond timestamps); licensed sell-side research; and the q-fin academic corpus.
• Base & size — Llama 4 Scout architecture (Apache 2.0); 13B fits a single H100 (80GB), enabling lower latency than larger models.
• Custom tokeniser — extended vocabulary for CUSIP/ISIN/Bloomberg tickers, corporate action types, SEC filing section headers; 32K context (1M is not needed for most quant tasks).
• Training — 512× H100 on DGX Cloud, 6–8 week window, Megatron-LM + NeMo 2.0; initial run $1.2M–$1.8M.

LoRA retraining scales linearly at 13B — monthly domain adaptation costs $30K–$70K per agent class; at 405B the same cycle would cost $750K–$1.5M, making continuous learning economically infeasible. Post-training, the model is packaged as a NIM container, deployed across all 36 agents as the base with agent-specific LoRA adapters, and A/B tested on live P&L over four weeks. Targets: 15–25% improvement in out-of-sample Sharpe attribution from signal agents; 30–40% reduction in false-positive rate from risk agents.

Academic benchmarks (SWE-Bench, AIME, MMLU-Pro, ARC-AGI) are structurally invalid for evaluating financial AI. They test static recall and symbolic reasoning on closed-form problems — neither captures the temporal, stochastic, regime-dependent nature of financial inference. The ESMA/Turing report notes standard AI metrics "do not capture the specific failure modes relevant to financial applications." The emerging institutional standard for 2026:

• Scale SEAL (finance subset) — document extraction, financial modelling, risk analysis; current leader Opus 4.7 at 71%. Useful for analyst-grade workflows, not quant trading AI.
• FinTSB (Tongji Finance Lab, 2025) — covers uptrends, downtrends, volatility, and black-swan regimes; preprocessing designed to prevent data leakage makes results meaningful for production.
• QuantBench 2026 — alpha generation by out-of-sample IC and rank-correlation stability across regime shifts; portfolio construction by Sharpe, Sortino, max drawdown, Calmar; risk by VaR breach rate and stress-test accuracy; zero-tolerance look-ahead detection.

Nyquist's per-agent production targets, measured weekly:

Crisis scenario regression testing. Every model update is tested on 2008 (Lehman, correlation breakdown), 2020 (COVID shock, liquidity crisis), and 2022 (rate shock, equity-bond correlation flip). A model must maintain positive risk-adjusted performance across all three before any production deployment. A human quant PM reviews 100 randomly sampled agent decisions weekly — any decision below 3/5 triggers model review and potential retraining.

The regulatory trajectory adds a dimension rarely discussed in AI capability debates. ESMA's February 2026 MiFID II briefing imposes Article 9 self-assessment (validate each strategy before deployment and after any "material change" — which every monthly LoRA cycle triggers), compliance staff comprehension (operationally incompatible with 60K–200K-token reasoning chains), and stress-testing at twice the firm's six-month peak volume (mandating on-demand stress-testable infrastructure, not general-purpose API endpoints). On-premises NIM under a defined governance framework provides a cleaner compliance path than opaque third-party API calls.

The honest constraints: NIM adds 15–30ms vs. bare-metal GB200, so ultra-low-latency HFT (<10ms) is out of scope; the 68% aggregate hides wide variance by task, regime, and data quality; the Q3 2026 timeline depends on DGX Cloud availability and training convergence; cost estimates track May-2026 AWS reserved and NVIDIA list pricing; LoRA gains diminish in stable regimes; QuantBench 2026 is an emerging, not universal, standard; ESMA's briefing is non-binding guidance; and GPT-4-class models later matched BloombergGPT on public benchmarks — so our 13B advantage will be validated against live P&L, not academic sets.