This is a cost-and-architecture guide for running earnings-call summarisation across the eight production LLMs available in Q2 2026, framed around five representative recent earnings calls (NVDA, MSFT, GOOGL, META, AAPL) from the most recent full reporting season. The verifiable axis here is unit cost, computed from published vendor list pricing; accuracy is discussed qualitatively as a class of failure modes, not as a scored leaderboard. We do not publish per-model accuracy numbers, because we have not run a reproducible, blind-graded benchmark on these calls, and inventing accuracy figures would be worse than omitting them. For the methodology of building a defensible finance LLM eval, see the finance LLM eval-harness guide. The choice of model depends on whether the downstream consumer is a portfolio manager (accuracy-bound) or a daily news desk (cost-bound).
Setup
Five representative calls, all held in July or August 2025 for quarters that closed in June or July 2025:
- NVIDIA Q2 FY2026 (August 27, 2025), ~17,840 input tokens.
- Microsoft Q4 FY2025 / Q2 calendar 2026 (July 30, 2025), ~16,210 input tokens.
- Alphabet Q2 2025 (July 23, 2025), ~19,540 input tokens.
- Meta Q2 2025 (July 30, 2025), ~18,920 input tokens.
- Apple Q3 FY2025 (July 31, 2025), ~14,120 input tokens.
The eight models under discussion:
- Claude Opus 4.8 (Anthropic, claude-opus-4-8)
- Claude Sonnet 4.6 (Anthropic, claude-sonnet-4-6-20250514)
- Claude Haiku 4.5 (Anthropic, claude-haiku-4-5-20250514)
- GPT-5.5 (OpenAI, gpt-5-5-2025-05)
- GPT-5.4 mini (OpenAI, gpt-5-mini-2025-05)
- Gemini 2.5 Pro (Google, gemini-2.5-pro)
- Gemini 2.5 Flash (Google, gemini-2.5-flash)
- DeepSeek V4-Flash (DeepSeek, deepseek-chat-v4-flash)
A defensible prompt for the task: "Summarise the following earnings call into eight sections, guidance, segment performance, capex, capital return, AI commentary, risk factors, analyst Q&A highlights, forward catalysts. Cite each fact with a verbatim quote in single quotes."
Cost: the one axis we can publish
Cost is computable directly from published vendor list pricing for input tokens (the dominant cost; output is 1–2k tokens, negligible) as of June 18, 2026. The $/call figures below assume an ~18,000-token call:
| Model | $/M input | $/call (~18k tokens) |
|---|---|---|
| Claude Opus 4.8 | 5.00 | $0.090 |
| GPT-5.5 | 5.00 | $0.090 |
| Claude Sonnet 4.6 | 3.00 | $0.054 |
| Gemini 2.5 Pro | 1.25 | $0.0225 |
| Claude Haiku 4.5 | 1.00 | $0.018 |
| GPT-5.4 mini | 0.75 | $0.014 |
| Gemini 2.5 Flash | 0.30 | $0.005 |
| DeepSeek V4-Flash | 0.14 | $0.003 |
The cost spread is roughly 36× from DeepSeek V4-Flash to Claude Opus 4.8. For per-ticker volume economics on this workload, see the companion cost-per-ticker breakdown, which derives the figures from the Earnings Call Summarization Cost engine.
Accuracy is a benchmark, not a guess
We are not publishing per-model accuracy scores for these calls. A credible accuracy comparison requires a fixed task set, a blind grading rubric applied by graders who do not know which model produced which output, multiple runs to bound variance, and a public artefact others can re-run. We have not done that for these five calls, so any accuracy number we printed would be invented — exactly the failure this hub refuses to commit. For the disciplined version, the eval-harness guide documents the methodology without printing a single unverified accuracy figure.
What we can describe is the class of failures earnings-call summarisation produces, which is model-agnostic and worth designing around regardless of which model you pick.
Failure modes to design around
Fabricated numbers: a model emits a revenue or guidance figure that does not appear in the transcript. This is the highest-stakes failure for a finance pipeline; route every numeric claim through a verifier such as the Hallucination Detector that checks each number against the source.
Misattributed quotes: a CFO statement attributed to the CEO, or vice versa. Matters for downstream legal review and citation chains; smaller models tend to do this more often, but treat it as a universal risk and verify attributions.
Inverted directional claims: "growth accelerated" reported as "growth decelerated", or "raised guidance" reported as "lowered guidance". A catastrophic failure mode for trading desks; flag every directional claim for human review regardless of model.
Latency
Latency is the other axis that matters in production but is environment-dependent (origin region, time of day, rate-limit queueing) and not something we measured here under a controlled harness, so we do not print per-model second counts. As a planning rule: smaller models stream faster, frontier models are slower, and P95 during US business hours runs well above the median. For overnight batch summarisation (the typical earnings-call pipeline), latency is rarely the binding constraint; for live coverage during the call window, budget for P95 and add a circuit breaker.
Cost at scale
A daily news-desk pipeline processing 60 earnings calls per quarter at peak season runs:
- 60 calls × 18,000 tokens × $5/M = $5.40/quarter on Claude Opus 4.8.
- 60 calls × 18,000 tokens × $0.30/M = $0.32/quarter on Gemini 2.5 Flash.
The $5.08/quarter delta is irrelevant at desk scale. The same delta scales to $508/quarter at a portfolio of 6,000 calls (full Russell 3000 coverage, four quarters), where it matters.
Recommendations by use case
Portfolio-manager memo (accuracy-bound): a frontier model (Claude Opus 4.8 or Sonnet 4.6, or GPT-5.5). Validate the model choice on your own task set using a blind-graded harness rather than trusting a vendor leaderboard.
News-desk daily summary (cost-bound): Gemini 2.5 Flash or Claude Haiku 4.5. At under $0.02/call the API cost is negligible; the binding constraint is review time.
Hybrid (most production setups): route via a confidence-weighted ensemble. Run a cheap model first; if its self-reported confidence is below threshold or the call exceeds a complexity score, escalate to a frontier model. This buys most of the quality lift at a fraction of the all-frontier cost.
Audit / compliance pipeline: mandatory dual-read with at least one frontier model. The cost is required when the downstream is regulated communications.
Per-call difficulty
Dense, segment-rich, AI-themed calls (NVIDIA-class) are harder for every model than calls with a uniform geographic-segment structure (Apple-class). A research desk that summarises 50 calls per quarter will encounter four or five NVIDIA-class calls where any single-model summary should be flagged for review. The cost of a second-pass review on those calls is negligible against the risk of misreporting capex guidance.
What did not work
A single-prompt, single-model summary is no substitute for either source-verification or explicit fact-extraction. The most defensible pipelines combine: (a) a small model for first-pass extraction, (b) a frontier model for synthesis, (c) a deterministic verifier that cross-references each numerical claim against an SEC filing or press release. A single-model summary measures only step (b) in isolation.
Caveats
Vendor pricing changes frequently. The cost numbers above are June 18, 2026 list prices; verify current pricing at the source URLs cited. The accuracy discussion is qualitative by design — for scored model comparisons, run a reproducible, blind-graded harness rather than trusting any unverified figure.
Connects to
- Eval Harness for Finance LLMs — how to score models without printing unverified accuracy figures.
- Earnings Call Summarization Cost (per ticker) — the volume cost economics for the same workload.
- Model Selector for Finance: interactive routing across the eight models above.
- Hallucination Detector — runs a verifier pass on any model's earnings-call summary.
- Token Cost Optimizer: minimises per-call cost without dropping quality.
- Prompt Patterns for Earnings Calls — companion piece on prompt engineering.
References
- Anthropic. Pricing. https://www.anthropic.com/pricing, accessed June 18, 2026.
- OpenAI. Pricing. https://openai.com/api/pricing — accessed June 18, 2026.
- Google Cloud. Vertex AI Generative AI Pricing. https://cloud.google.com/vertex-ai/generative-ai/pricing, accessed June 18, 2026.
- DeepSeek. API Pricing. https://api-docs.deepseek.com/quick_start/pricing — accessed June 18, 2026.
- Lin, S., Hilton, J., & Evans, O. (2022). "TruthfulQA: Measuring How Models Mimic Human Falsehoods." ACL 2022. DOI: 10.18653/v1/2022.acl-long.229.
- Manakul, P., Liusie, A., & Gales, M. J. F. (2023). "SelfCheckGPT: Zero-Resource Black-Box Hallucination Detection for Generative Large Language Models." EMNLP 2023. DOI: 10.18653/v1/2023.emnlp-main.557.
- NVIDIA Corporation. (2025). Q2 Fiscal Year 2026 Earnings Call Transcript. August 27, 2025.
- Microsoft Corporation. (2025). Q4 Fiscal Year 2025 Earnings Call Transcript. July 30, 2025.
- Alphabet Inc. (2025). Q2 2025 Earnings Call Transcript. July 23, 2025.
- Meta Platforms Inc. (2025). Q2 2025 Earnings Call Transcript. July 30, 2025.
- Apple Inc. (2025). Q3 Fiscal Year 2025 Earnings Call Transcript. July 31, 2025.
- Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). "On the Dangers of Stochastic Parrots." FAccT 2021. DOI: 10.1145/3442188.3445922.