Claude Opus 4.7 vs Gemini 2.5 Pro: 10-K Extraction Cost

On a 10-K extraction with 2-peer comparison, 2,000-token output, 50% cache hit rate, the Financial Document Token Estimator returns: Claude Opus 4.7 one-pass cost $0.1066, synthesis $0.2197; Gemini 2.5 Pro one-pass $0.0341, synthesis $0.0622, a 3.1× one-pass cost differential in Gemini's favour, widening to 3.5× on the 2-peer synthesis path. Cross-checked against the Model Selector for Finance on a high-quality / $200+ budget / 200k–1M context / sub-30s extraction profile, both Opus 4.7 ($180/mo selector estimate) and Gemini 2.5 Pro ($58/mo selector estimate) tie at score 98.0. The selector's monthly ratio is 3.1×, now in line with the per-call one-pass ratio (3.1×) and close to the synthesis ratio (3.5×) since Opus 4.7's $5/$25 rates narrowed the gap. The question is whether the cost gap is bought back by extraction-quality at the 96%+ tolerance threshold.

TL;DR

Same workload, two flagship-tier models, three reference points:

Metric	Claude Opus 4.7	Gemini 2.5 Pro	Ratio
One-pass extraction cost ($/filing)	$0.1066	$0.0341	3.1×
With 2-peer synthesis	$0.2197	$0.0622	3.5×
Selector monthly budget at reference workload	$180	$58	3.1×
Selector score (extract / high / 200k–1M / sub-30s)	98.0	98.0	tied

The two models tie on the selector's filter axes. Gemini Pro is ~3× cheaper. The pivot is extraction quality, which neither engine measures. The selector explicitly hands off to a 50-task eval harness on the buyer's own documents.

The 3× cost gap, decomposed

Both models price input and output tokens. The list rates at the engine's reference snapshot:

Component	Claude Opus 4.7	Gemini 2.5 Pro
Input per Mtok	$5.00	$1.25
Output per Mtok	$25.00	$10.00
Cache read multiplier	0.10	0.25
Cache write multiplier	1.25	1.00
Context window	1,000,000	2,000,000

The cost gap is structural, Opus is positioned as the flagship reasoning model with premium pricing; Gemini Pro is positioned as price-performance leader. The 10-K body archetype is ~72,000 characters, which the engine tokenises per-model (3.5 chars/token for Anthropic, 4 for Google), so the input-token counts differ. On a one-pass 10-K extraction (output 2k tokens, 50% cache), the per-call cost decomposes:

Opus 4.7, 50% cache, 20,571 input tokens:

• Input: 20,571 × $5/Mtok with half the tokens at the 0.10 cache-read multiplier = $0.057
• Output: 2k × $25/Mtok = $0.050
• Engine returns onePassCost $0.1066; with 2-peer synthesis (input ×3 = 61,713 tokens for focal + 2 peers) synthesisCost $0.2197 — only 2.06× the one-pass cost, not 3×, because the 2,000-token output is fixed and is not tripled by adding peers

Gemini Pro, 50% cache, 18,000 input tokens:

• Input: 18,000 × $1.25/Mtok with half at the 0.25 cache-read multiplier = $0.014
• Output: 2k × $10/Mtok = $0.020
• Engine returns onePassCost $0.0341, synthesisCost $0.0622

The cost arithmetic is the dominant fact; everything else is buyer-specific.

Where Opus earns its premium: extraction quality at the 96%+ threshold

Vendor-published evaluations on document-extraction tasks consistently show Opus and the flagship-tier OpenAI / Google models within 2–4 percentage points of each other on most benchmarks. The gap widens on specific failure modes:

• Numeric precision under context pressure. At 90% of the published context window (900k+ tokens for Opus, 1.8M+ for Gemini Pro), accuracy degrades for all models; the rate of degradation is the differentiator.
• Multi-step reasoning chains. When the extraction requires intermediate inference (compute YoY growth, reconcile restated figures), models with extended-thinking modes diverge from models without.
• Adversarial format handling. Filings with non-standard table formatting, footnote-heavy disclosure, or restatement annotations differentiate the high-end models from the mid-tier.

For workloads where the extraction must be 96%+ accurate (audit-trail-bound contexts, regulatory filings going into compliance reports), the 4-point gap is the difference between auto-publishing and human review on every line. The cost of human review (5 minutes × $30/hr per flagged extraction × thousands of extractions) easily exceeds the 3× API premium.

For workloads where 92% accuracy is acceptable (research notes, internal trade memos, exploratory analysis), the 3× premium is harder to justify, Gemini Pro at $58/mo delivers the buyer's quality threshold at roughly one-third the cost.

The selector's score-98 tie reflects this ambiguity. It cannot adjudicate the cost-quality trade-off; it can only confirm both models pass the filter axes. The buyer's eval harness is the next step.

The 50-task eval threshold

A defensible procurement decision requires a 50-task eval on the buyer's own documents:

1. Build a representative 50-task set: 50 filings or filing-sections the buyer regularly extracts.
2. Define a grading rubric: accuracy on specific fields (revenue, EPS, guidance, segment splits) with binary pass/fail per task.
3. Run both candidates against the rubric, blind to model identity.
4. Compute pass rate per model with 95% confidence interval.
5. If pass-rate CIs overlap, the models are statistically indistinguishable on this workload — pick the cheaper one.

For 50 tasks the CI half-width is approximately √(p·(1−p)/50) × 1.96 ≈ 0.14 at p = 0.85. So if Opus passes 90% and Gemini Pro passes 80%, the CIs are (0.81, 0.99) and (0.69, 0.91) — overlapping, statistically indistinguishable.

A 200-task eval narrows the CI to ±0.07, distinguishing 90% from 80% reliably. For high-volume workloads where the cost gap matters, the 200-task investment is worth several months of API spend.

The cache-architecture interaction

The engine's 50% cache hit rate assumption applies to both models. The cache architectures differ:

• Anthropic prompt caching has a 5-minute TTL by default (extendable). The 50% blended hit rate assumes most calls hit the cache within the TTL.
• Gemini context caching has explicit-create semantics with longer TTLs (~1 hour). The 50% blended hit rate assumes the buyer maintains a per-filing cache key.

Both vendors' cache pricing makes the cache-hit case cheap, but the operational architecture differs. Anthropic's cache is implicit (the system extracts the cached prefix from a marked region of the prompt); Gemini's cache is explicit (the buyer creates and manages cache lifecycle).

For workloads that re-query the same filing many times in a session, Anthropic's implicit cache is easier; for workloads with stable across-day caches, Gemini's explicit cache is more efficient. The engine's blended 50% rate does not distinguish; the architecture decision is parallel to the cost decision.

Context window: where Gemini Pro takes a small lead

Opus 4.7 ships with 1M context¹; Gemini 2.5 Pro with 2M². For most 2-peer 10-K extractions (41k tokens of input) the difference is irrelevant. For workloads that consume the full window (12-peer comparison across 8 years of filings = ~720k tokens), Opus is approaching its limit while Gemini Pro has 60% headroom.

The Liu et al. "Lost in the Middle" research³ documents that accuracy degrades disproportionately as input approaches the context-window limit. For workloads that routinely consume > 80% of the published window, Gemini Pro's 2M ceiling is a material edge. The eval harness should specifically include some long-context tasks (>700k tokens) to probe this dimension.

The decision flowchart

For an LLM-driven retail finance workflow, the procurement decision tree is:

1. What is your extraction-quality tolerance? Below 92%: Gemini Pro and stop. Above 96%: probably Opus, after eval. Between: run the 50-task eval and decide on CI.
2. What is your context-window utilization? Below 50% of Opus's 1M: either model. Above 70%: Gemini Pro.
3. Do you have stable cache patterns? If yes (per-filing caches with hours-long lifecycle), Gemini Pro's explicit-cache architecture wins on efficiency. If no (ad-hoc retail), Anthropic's implicit cache is simpler.
4. Is the 3× cost gap material at your volume? At 100 extractions/month the gap is about $7; at 10,000/month the gap is about $725. The eval cost has to be amortized against the lifetime savings.

For most retail solo workflows, Gemini Pro is the right starting point. For audit-trail-bound institutional workflows, Opus is the safer starting point. The eval harness is required regardless of starting point.

Connects to

• Earnings Call Summarization Eight LLMs Q2 2026 — empirical eight-model comparison on a related task.
• Finance LLM Eval-Harness Guide — the eval-harness template.
• Reading Financial Filings with LLMs 2026 — the broader workflow context.
• Financial Document Token Estimator — engine endpoint for cost calculation.
• Model Selector for Finance — engine endpoint for tier selection.
• Token Cost Optimizer — companion for the full cost-per-validated-extraction frame.

References

• Anthropic. "Pricing." anthropic.com/pricing, accessed 2026-05-21.
• Google. "Gemini API pricing." ai.google.dev/pricing, accessed 2026-05-21.
• Stanford Institute for Human-Centered AI (HAI). "AI Index Report 2026." aiindex.stanford.edu, accessed 2026-05-21. Methodology reference for cross-model eval design.

Footnotes

1. Anthropic. "Claude 4.7 model card." anthropic.com/news, accessed 2026-05-21. https://www.anthropic.com/ ↩

2. Google. "Gemini 2.5 model documentation." ai.google.dev/gemini-api/docs/models/gemini, accessed 2026-05-21. ↩

3. Liu, N. F., Lin, K., Hewitt, J., et al. (2024). "Lost in the Middle: How Language Models Use Long Contexts." Transactions of the Association for Computational Linguistics 12. https://arxiv.org/abs/2307.03172 ↩

Verified engine output

Show the recompute-verified inputs and outputs

Inputs

mode	archetype
archetype_id	10-k
output_tokens	2000
peers	2
cache_hit_rate	0.5

Result

(10 items)	[...]

Computed live at build time.

Frequently asked questions

Is Opus 4.7 always more accurate than Gemini 2.5 Pro?

No — accuracy depends on the task, prompt shape, and document corpus. Vendor benchmarks frequently disagree; the 50-task eval on the buyer's own documents is the only defensible decision input.

Why does the engine return 50% cache hit rate as the default?

Because empirical retail workflows tend to settle in the 40–60% range. Higher requires stable across-call prefixes; lower is the no-cache baseline. Sweep the input to see sensitivity.

Does the 3.1× selector cost ratio match the per-call ratio?

Yes. With Opus 4.7 at $5/$25 the selector's 3.1× monthly ratio matches the per-call one-pass ratio and sits just under the 3.5× synthesis ratio. The tools size different reference workloads, but both point to Gemini Pro as the cheaper pick.

Should I always use the synthesis cost, not the one-pass cost?

Only when peer comparison is part of the workflow. Use onePassCost for single-filing extraction; use synthesisCost for research workflows that compare focal filings against peers.

Are there other flagship models worth considering?

GPT-5.5 now lands slightly above Opus 4.7 and well above Gemini Pro on cost; it ties Opus on most reasoning-heavy tasks. A three-way eval is institutional-grade; two-way is sufficient for retail.