Note, added November 2026: This post was written before we fixed how benchmark statements are grouped. Until 21 October 2026, statements that read alike were merged even when they were about different models, tasks or snapshots, so a figure here may belong to a different model, task or snapshot than the one named beside it, and some sources may come from other snapshots. We have left the post as it was published. Current figures are on the benchmark.
Batch versus sync: cost is material, but routing remains task-specific
Batch can reduce the price of LLM work substantially. It does not make the wrong model right, preserve quality by definition, or satisfy a user who needs an immediate answer.
Of 68 tasks, 50 are batch-eligible and 18 are not. Batch handled 410,019 of 1,109,798 successful calls, or 37%.[^1][^49] Eligibility belongs to a particular model-task configuration; a sync-only route does not become cheaper just because the surrounding pipeline uses batch.
The discounts are real
Matched price pairs repeatedly favor batch. Gemini 3.5 Flash costs $0.07 rather than $0.12 for taxonomy matching and $0.05 rather than $0.09 for content-set relevance.[^5][^12][^17][^19][^40][^47][^51][^52]
GPT-5.6 Sol falls from $0.64 to $0.18 for reference-preserving analysis, while GPT-5.5 falls from $0.08 to $0.03 for image-prompt generation.[^3][^9][^25][^31][^45][^53] Claude Sonnet 5 also has lower listed batch prices for cluster labeling, factual-claim extraction, relevance scoring, profile matching, decision synthesis, and metadata rewriting.[^10][^13][^14][^18][^20][^21][^23][^29][^33][^39][^42][^44][^50][^52]
These comparisons establish a price direction. They do not establish that batch and sync outputs have equal quality.
Batch can change the economic winner
Image-prompt generation provides the clearest example. GPT-5.6 Terra scores 9.08 at $0.01 synchronously; GPT-5.6 Luna scores 8.90 at $0.0006 in batch. Their confidence intervals overlap, so the measured quality is a tie.[^3][^31]
For a deferrable workload, Luna’s much lower listed cost can decide the route. GPT-5.4 Nano is cheaper than Terra but scores only 7.38, so price alone does not make it interchangeable with the tied higher-quality options.[^3][^31]
Other batch results remain useful without proving cross-mode equivalence. Gemini 3.5 Flash scores 9.78 for batch structured extraction at $0.05, and Claude Sonnet 5 scores 9.32 for batch newsletter copy at $0.0041.[^2][^11][^36][^43] Without same-model synchronous quality results, neither comparison proves that moving the task to batch is quality-neutral.
The price card is not total cost
Gemini 3.1 Pro Preview required a separately billed repair call on 1.7% of 15,389 attempts. MiniMax M2.5 required one on 17.2% of 40,876 attempts, in addition to substantial local repair.[^1][^49] Shadow traffic to models whose answers will not be selected can add more cost.
A production comparison should therefore include repair calls, local processing, parse failures, redispatch, and comparison traffic—not just the displayed per-task price.
Routing conclusion
Use batch when all three conditions hold:
The work can wait.
The exact model-task configuration supports batch and clears the quality bar.
Total completed-task cost remains lower after repairs and comparison traffic.
Keep synchronous service for immediate-response paths. Do not compare prices across different tasks or assume that a batch discount changes the quality ranking.
