Gemini

42.1 Two Caching Mechanisms #

Gemini distinguishes implicit from explicit caching, and the two carry different cost structures.⁵⁰

Implicit caching is on by default for Gemini 2.5 and newer. When a request reuses a prefix, the discount is passed on automatically. No storage charge, no configuration, no guarantee. Minimum token counts are model-specific rather than tier-specific: 2,048 on Gemini 2.5 Flash and 2.5 Pro, and 4,096 on the 3.x models, Gemini 3.1 Pro included.

Explicit caching is a managed object. You create a CachedContent, reference it by resource name, set a TTL, and delete it when done. It guarantees the discount, 90 percent on Gemini 2.5 and later, and it adds a storage charge.

from google import genai
from google.genai import types

client = genai.Client()

cache = client.caches.create(
    model="gemini-3.1-pro-preview",
    config=types.CreateCachedContentConfig(
        display_name="codebase-context-v3",
        system_instruction="You are a code reviewer for this repository.",
        contents=[large_document],
        ttl="3600s",                        # storage meter runs for this whole hour
    ),
)

response = client.models.generate_content(
    model="gemini-3.1-pro-preview",
    contents="Review the diff below.",
    config=types.GenerateContentConfig(cached_content=cache.name),
)

client.caches.delete(name=cache.name)       # stop the meter — do not skip this

That final line is what separates Gemini from the other providers. Elsewhere a cache is an optimization you can forget about; here it is an allocated resource with a lifecycle you own. The next subsection explains why forgetting it costs money.

42.2 The Storage Meter Changes Everything #

Explicit caching bills per token per hour that the cache exists, whether or not it is read. The rate is per model and per service tier rather than per family: $0.50 per million tokens per hour on Gemini 3.6, 3.7 and 3.8 Flash through December 31, 2026 and $1.00 from January 1, 2027; $1.00 already on 3.5 Flash and the Flash-Lite models at Standard, though 3.1 Flash-Lite bills $0.50 under Batch; and $4.50 on Gemini 3.1 Pro.⁵¹

Worked, on Gemini 3.1 Pro at $2.00/MTok input and $4.50/MTok/hour storage:

100K-token context held for one hour:
  storage:  0.1M × $4.50 = $0.45/hour
  saving per read: 0.1M × ($2.00 − $0.20) = $0.18

Break-even: 0.45 / 0.18 = 2.5 reads per hour.

Both terms scale with the prefix, so the crossover is the storage rate over the per-token saving and holds at any size inside the tier. Size still decides which tier you are in: above the 200K threshold in §42.4, input and cached input both double while the storage meter does not, and the crossover halves to 1.25 reads an hour. Keep the prefix clear of the line because a 200K cached context plus a question is already over it.

Below roughly three reads an hour, explicit caching loses money. On every other provider a cold cache is a sunk cost; on Gemini explicit caching it is a running meter.

The operating rules follow:

• Default to implicit caching. No meter, no management, no downside.
• Use explicit caching for large, hot contexts where the read rate clearly exceeds break-even and you want the discount guaranteed.
• Delete caches when the job finishes. A forgotten cache is a standing charge. Set short TTLs and treat cache lifecycle as owned resource management.

42.3 Hit Reliability #

Implicit caching is best-effort and controlled upstream. Practitioner reporting suggests hit rates trail the explicit-control providers, which means cost estimates for Gemini workloads should be built at the uncached price with implicit caching treated as upside rather than as budget.

To improve implicit hit rates: put large common content at the beginning of the prompt, and send requests with similar prefixes close together in time.⁵⁰ Both are the same advice as everywhere else in this document; the difference is that you have no lever beyond it.

42.4 Pricing Notes #

Cached input costs 10 percent of fresh input on Gemini 2.5 and later; the discount was 75 percent on 2.0-generation models. Gemini 3.1 Pro carries a context threshold at 200K tokens, above which input and cached input double and output rises by half—the same multiplier structure as OpenAI’s long-context tier, which begins at 272K (§21.3). Batch halves input and output. Whether it also discounts cached input is model-specific, and the pricing page is the only way to settle it: on Gemini 3.1 Pro the Batch cached-input row reads “same as Standard,” so batching and explicit caching do not compound there, while the 3.6, 3.7 and 3.8 Flash Batch tables halve cached input alongside everything else, and 3.1 Flash-Lite halves the storage meter too.⁵¹ Read the row for the model you are billing, not the pattern from another one.

42.5 What Is Distinctive #

The explicit-cache-as-a-managed-object model is simultaneously the most controllable and the most dangerous. A cache write that is never read costs more than ordinary input on Anthropic and on GPT-5.6 and later as well (§21.2), and there it is a one-time premium; here it is a standing charge that accrues until the object is deleted. Gemini is the only provider that meters a cache by the hour, and the only one that requires cache lifecycle management as an operational concern. Teams migrating from Anthropic or OpenAI frequently carry over the assumption that caching is free upside and are surprised.

The 1M-token context on the Pro tier is large and useful for document-scale work—subject to every caveat in §29.5 about effective versus advertised context.