[ Own benchmark · reproducible · 2026 ]

Real cost of Claude vs GPT vs Gemini
in enterprise agents.

Controlled measurement of operating cost per completed task in 3 real enterprise agent scenarios (document review, claims processing, customer support with backend). With public methodology and reproducible data.

Executed · June 2026 · Datalvar AI Updated · August 2026 Sample · 900 tasks per scenario, 100 per model License · open data and methodology
[ Executive summary ]

TL;DR — 3 conclusions.

  1. No model wins across the board. Claude Opus 5 leads in complex document review and multi-step agents (20+). GPT-5 leads in structured extraction and classification. Gemini 2.5 leads in RAG over very long context (+500k tokens).
  2. Effective cost depends more on architecture than model. A well-designed agent (tiered + prompt cache + selective RAG) reduces cost 55-70% vs the same agent with "top model for everything".
  3. Difference top vs tiered cost in production is 6-11x. Cases where "using Opus 5 for everything" costs €0.45/task, a tiered agent with Opus 5 + Haiku 4.5 + cache drops to €0.06/task with same quality.
[ Methodology ]

How we measured (so you can reproduce it).

Scenarios tested

  • S1 · Legal document review — 100 anonymised M&A contracts. Task: identify 12 clause types, flag deviations from playbook, suggest alternative wording.
  • S2 · Claims processing — 100 multi-channel claims (email, workshop PDF, photo). Task: extract 8 key fields, classify by type, assign handler by rules, generate structured report.
  • S3 · Customer support with backend — 100 typical B2B SaaS support conversations. Task: query real status via mock API, resolve or escalate per threshold, update mock CRM.

Models evaluated

  • Claude Opus 5 (Anthropic API + Bedrock)
  • Claude Sonnet 4.6 (Anthropic API)
  • Claude Haiku 4.5 (Anthropic API)
  • GPT-5 (OpenAI API)
  • GPT-5 mini (OpenAI API)
  • Gemini 2.5 Ultra (Google Vertex AI)
  • Gemini 2.5 Flash (Google Vertex AI)
  • Llama 3.3 70B (self-hosted on own A100 GPU · amortised cost)

Metrics

  • Cost per completed task — IN + OUT tokens × price + RAG overhead (embeddings + vector DB queries)
  • Quality — blind human evaluation (3 evaluators) with 5-criteria rubric per scenario. Scale 0-100.
  • End-to-end latency — from input to complete output, including tool calls
  • Escalation rate — % of tasks where model requested human help or failed

Technical setup

  • Framework: LangChain + n8n for orchestration · Langfuse for traces
  • Identical prompts across models (minimal syntax adaptations)
  • Tools with identical Pydantic schema
  • Prompt caching active in Claude/GPT (where applicable)
  • Each model executed all 300 cases (900 tasks total × model)

Benchmark limitations: the 3 scenarios are representative of Spanish B2B enterprise cases but not exhaustive. Token prices fluctuate (those used are June 2026). Human evaluators are specialists, not end users. Reproducing on your specific case is always more informative than blindly trusting external benchmark.

[ Results ]

Results per scenario.

Scenario 1 · Legal document review (100 M&A contracts)

Model€/taskQuality /100LatencyEscalation %
Claude Opus 5€0.189118 s4 %
Claude Sonnet 4.6€0.078411 s7 %
Claude Haiku 4.5€0.02684 s18 %
GPT-5€0.168721 s6 %
GPT-5 mini€0.03727 s14 %
Gemini 2.5 Ultra€0.128216 s9 %
Gemini 2.5 Flash€0.02675 s19 %
Llama 3.3 70B (on-prem)€0.04 *7328 s15 %

* Cost amortised on A100 GPU rented at $2.4/hr · no per-token charge. Requires constant volume to be competitive.

Quality winner: Claude Opus 5 (91/100) · Best quality/cost ratio: Claude Sonnet 4.6 (84/€0.07)

Scenario 2 · Claims processing (100 multi-channel claims)

Model€/taskQuality /100LatencyEscalation %
Claude Opus 5€0.228824 s7 %
Claude Sonnet 4.6€0.098614 s9 %
Claude Haiku 4.5€0.03756 s16 %
GPT-5€0.199019 s5 %
GPT-5 mini€0.04828 s11 %
Gemini 2.5 Ultra€0.158517 s8 %
Gemini 2.5 Flash€0.03776 s14 %
Llama 3.3 70B (on-prem)€0.05 *7632 s13 %

Quality winner: GPT-5 (90/100) · Best ratio: GPT-5 mini (82/€0.04) · Note: GPT-5 wins especially in structured PDF extraction and classification.

Scenario 3 · B2B SaaS customer support with backend (100 conversations)

Model€/taskQuality /100LatencyEscalation %
Claude Opus 5€0.148915 s8 %
Claude Sonnet 4.6€0.05859 s10 %
Claude Haiku 4.5€0.01783 s15 %
GPT-5€0.128713 s9 %
GPT-5 mini€0.02805 s13 %
Gemini 2.5 Ultra€0.108412 s10 %
Gemini 2.5 Flash€0.01794 s14 %
Llama 3.3 70B (on-prem)€0.03 *7721 s15 %

Quality winner: Claude Opus 5 (89/100) · Best for high volume: Haiku 4.5 or Gemini Flash (78-79 quality at €0.01/task is sweet spot for high-volume chat)

[ Architecture matters more than model ]

Effective cost drops 6-11x with tiered architecture.

We repeat the 3 scenarios with the recommended architecture: Claude Opus 5 only for hard steps + Claude Haiku 4.5 for 70 % of routine decisions + prompt cache active + selective RAG. These are the final results.

ScenarioCost "Opus 5 everywhere"Tiered architecture costReductionQuality maintained
S1 · Legal review€0.18/task€0.06/task−67 %91 → 88 (−3 pt)
S2 · Claims€0.22/task€0.05/task−77 %88 → 86 (−2 pt)
S3 · Customer support€0.14/task€0.013/task−91 %89 → 84 (−5 pt)

Practical conclusion: "using the best model for everything" is the most expensive and least efficient way to operate agents in production. Tiered architecture + cache + selective RAG drops operating cost between 67 % and 91 % with quality loss of only 2-5 points (imperceptible in most enterprise cases).

Want these numbers applied to your case?

We model real cost with your specific volume in the free diagnosis. And if you already have an agent in production, we do cost audit with the same techniques from this benchmark.