Straitly compass logostraitly

GUIDE · FORMULA · CALCULATOR

Reduce LLM inference costs with prompt caching

To reduce LLM inference costs, cache the part of the prompt you resend: it is billed at 0.1x the normal input price at Anthropic and OpenAI.[2][5] A 50-turn coding agent session pays 83.5% less for input, and on Straitly the same cache reads cost a median 40% below list.

Last updated . Every number links to its source.

ONE CODING AGENT SESSION: 50 TURNS, 20,000-TOKEN PREFIX

  • No cache$2.10

  • 5-minute cache$0.346

  • Input cost, Claude Sonnet 5: one 1.25x write, 49 reads at 0.1x. [2] Anthropic docs, Checked 9 November 2026
83.5%
less input cost
40%
below list on Straitly cache reads, median of 51 models [14]

How much does prompt caching save?

Agents and chatbots resend the same long prefix every turn, so most of their tokens can be cache reads billed at a fraction of the normal input price.

  • 41-80%

    Lower API costs from prompt caching, measured on 500+ agent sessions (2026 study).[11] Lumer et al., 9 Jan 2026

  • 95%+

    Share of an agent's tokens that can be cache reads (Menlo Ventures, 2026).[1] Menlo Ventures, Nov 2026, survey fielded Sep 2026

  • >90%

    Cache-hit rate of an example multi-turn agent in OpenAI's guide (illustrative).[5] OpenAI docs, Checked 9 November 2026

  • 13-31%

    Faster time to first token, same study.[11] Lumer et al., 9 Jan 2026

One prefix, ten requests: what you pay

One write + 9 reads of a shared prefix, as a multiple of its input price.

  1. No cache10x
  2. DeepSeek, disk cache (off-peak) · DeepSeek V4.1 Flash1.18x
  3. Google, implicit cache · Gemini 4 Argon1.45x
  4. Anthropic, Opus 5.5, 5-minute · Claude Opus 5.51.7x
  5. OpenAI, GPT-6.1 Sol, 30-minute · GPT-6.1 Sol1.7x
  6. Anthropic, 5-minute cache · Claude Sonnet 52.15x
  7. OpenAI, 30-minute cache · GPT-6 Astra2.15x
  8. Anthropic, 1-hour cache · Claude Sonnet 52.9x
  9. OpenAI, 12-hour cache · GPT-6 Astra3.4x

[2] Anthropic docs, Checked 9 November 2026; [5] OpenAI docs, Checked 9 November 2026; [8] Google, Checked 9 November 2026; [10] DeepSeek API docs, Checked 9 November 2026

How does each provider price prompt caching?

The first request pays a write to store the prefix; later requests pay a fraction to read it. Output is never discounted.

Prompt caching rules by provider
PROVIDERCACHE WRITECACHE READLIFETIMEMINIMUM PREFIX
Anthropic (Claude)1.25x for a 5-minute cache, 2x for a 1-hour cache0.1x. Opus 5.5 and Sonnet 5.5: 0.05x. Fable 5.1 and Mythos 5.1: 0.025x5 minutes by default, refreshed at no cost on each hit512 tokens on Fable 5.1, Opus 5.5, Opus 5, Sonnet 5.5 and Haiku 5.5
OpenAI (GPT-5.6 and later)1.25x for the 30-minute default, 2.5x for 12 hours (Responses API)0.1x. GPT-6.1 Sol and Terra: 0.05x30 minutes by default, 12 hours on request1,024 tokens
Google (Gemini 4 Argon)No write fee. Implicit caching is automatic$0.10 against $2.00 input, which is 0.05xImplicit, managed by Google. Explicit caching is not offered from Argon on4,096 tokens
DeepSeek (V4.1 Flash)No write fee. A miss is billed at the normal input price$0.003 against $0.15 off-peak, which is 0.02xDisk cache, best effort, cleared within a few hours to a few days of last useNot listed

[2] Anthropic docs, Checked 9 November 2026; [5] OpenAI docs, Checked 9 November 2026; [7] Google Gemini API docs, Checked 9 November 2026; [9] DeepSeek API docs, Checked 9 November 2026

Batch discounts, rate limits and Gemini agent loops

Anthropic cache multipliers also stack with its 50% Batch discount, and cache hits do not count against Anthropic rate limits.[2]

On Gemini 4 agent loops, the previous request is cached only if the new request adds at least 2,050 tokens.[7]

Anthropic at launch: “reducing costs by up to 90% and latency by up to 85% for long prompts.”[4]

What does a cache read cost per model?

From $0.003 to $1.00 per million tokens, following each model's input price.

  • Up to 95%

    OpenAI's maximum discount on cached input tokens.[5] OpenAI docs, Checked 9 November 2026

  • ~$0.70

    Blended cost per million tokens of an Opus 5.5 agent session that is mostly cache reads (Menlo estimate).[1] Menlo Ventures, Nov 2026, survey fielded Sep 2026

Prompt cache prices by model, USD per million tokens
MODELINPUTCACHE READCACHE WRITE
Claude Fable 5.1$10.00$0.25-
Claude Opus 5.5$4.00$0.20$5.00 (5m) / $8.00 (1h)
Claude Sonnet 5.5$2.00$0.10$2.50 (5m) / $4.00 (1h)
Claude Sonnet 5$2.00$0.20$2.50 (5m) / $4.00 (1h)
Claude Haiku 5.5Prompts up to 100K tokens$0.10$0.01$0.125 (5m) / $0.20 (1h)
GPT-6 Astra$10.00$1.00$12.50 (30m) / $25.00 (12h)
GPT-6.1 Sol$2.00$0.10$2.50 (30m) / $5.00 (12h)
GPT-6.1 Terra$1.00$0.05$1.25 (30m) / $2.50 (12h)
GPT-6 Luna$0.10$0.01$0.125 (30m) / $0.25 (12h)
Gemini 4 ArgonInput doubles on 1 Dec 2026$2.00$0.10Implicit, no fee
Gemini 3.8 FlashThrough 31 Dec 2026$0.75$0.075Storage $0.50 per MTok per hour
DeepSeek V4.1 Flash (off-peak)Peak rates are double$0.15$0.003Miss price, no fee
Show all 12 models

USD per million tokens, standard requests. [2] Anthropic docs, Checked 9 November 2026; [6] OpenAI, Checked 9 November 2026; [8] Google, Checked 9 November 2026; [10] DeepSeek API docs, Checked 9 November 2026

How do you calculate prompt caching cost?

Caching is cheaper on the prefix once w + (k − 1) × r < k.

Without a cache

cost = k × (P + U) × input price

With a cache

cost = (P × w + (k − 1) × P × r + k × U) × input price

P cached prefix tokens, k requests per write, U new tokens per request, w and r the write and read multipliers.

Requests per cache write needed to break even
CACHEWRITE / READCHEAPER FROMSAVING AT 10
Anthropic, 5-minute cache1.25x / 0.1x2 requests78.5%
Anthropic, 1-hour cache2x / 0.1x3 requests71.0%
OpenAI, 30-minute cache1.25x / 0.1x2 requests78.5%
OpenAI, GPT-6.1 Sol, 30-minute1.25x / 0.05x2 requests83.0%
OpenAI, 12-hour cache2.5x / 0.1x3 requests66.0%
Google, implicit cache1x / 0.05x2 requests85.5%
DeepSeek, disk cache (off-peak)1x / 0.02x2 requests88.2%

Saving at 10: the prefix saving over 10 requests in one window. [2] Anthropic docs, Checked 9 November 2026; [5] OpenAI docs, Checked 9 November 2026; [8] Google, Checked 9 November 2026; [10] DeepSeek API docs, Checked 9 November 2026

“writing a prefix once and fully reusing it once costs 1.35× its ordinary input cost, compared with 2× for processing it twice without caching”
[5] OpenAI docs, Checked 9 November 2026

How much would caching save you?

Three common workloads save 83.5% to 87.3% of input cost.[2][6][10]

Start from an example

Claude Sonnet 5: write 1.25x, read 0.1x of $2 per million input tokens.

Input cost saved

83.5%

  • No cache, per cache write$2.10

  • With a cache, per cache write$0.346

Month, no cache
$2,100.00
Month, cached
$346.00
The three examples, worked out line by line

Coding agent session: 83.5% saved. A 20,000-token prefix (tools, rules, repo map) reused across 50 agent turns inside the 5-minute window, with 1,000 new tokens per turn. Claude Sonnet 5, write 1.25x, read 0.1x, $2/M input.

Without a cache: 50 × (20,000 + 1,000) = 1,050,000 tokens × $2/M = $2.10. With a cache: 20,000 × 1.25 + 49 × 20,000 × 0.1 + 50 × 1,000 = 173,000 tokens × $2/M = $0.346.

Support bot: 86.8% saved. An 8,000-token system prompt and help-center excerpt, 20 questions per 30-minute window, 200 new tokens per question. GPT-6.1 Sol, write 1.25x, read 0.05x, $2/M input.

Without a cache: 20 × (8,000 + 200) = 164,000 tokens × $2/M = $0.328. With a cache: 8,000 × 1.25 + 19 × 8,000 × 0.05 + 20 × 200 = 21,600 tokens × $2/M = $0.0432.

Document Q&A: 87.3% saved. A 30,000-token report asked 10 questions, 300 new tokens per question, off-peak hours. DeepSeek V4.1 Flash, write 1x, read 0.02x, $0.15/M input.

Without a cache: 10 × (30,000 + 300) = 303,000 tokens × $0.15/M = $0.0455. With a cache: 30,000 × 1 + 9 × 30,000 × 0.02 + 10 × 300 = 38,400 tokens × $0.15/M = $0.0058.

Assumptions: input tokens only, since output is billed the same with or without a cache. Every request after the write lands inside the cache lifetime and hits the cache. The Anthropic example uses Claude Sonnet 5 with a 1.25x write and 0.1x read. Prices from the tables above, checked 9 November 2026.

How do you get more cache hits?

Caches match an exact prefix, so one changed token near the top turns everything after it into a miss.

  1. 1Fixed parts first. Tools, system prompt, documents, history, then the new message. Timestamps go last.[11]
  2. 2Clear the minimum. A prefix under the minimum is silently not cached.[2][5][7]
  3. 3Check the usage fields. cache_read_input_tokens at Anthropic, cached_tokens at OpenAI, prompt_cache_hit_tokens at DeepSeek. A study found 32 caching mistakes in 571 agent projects.[12]

Minimum prefix that can be cached

Tokens. Shorter prompts are processed without caching and without an error.

  1. Claude Fable 5.1, Opus 5.5, Sonnet 5.5, Haiku 5.5512
  2. Claude Sonnet 5, Opus 4.81,024
  3. OpenAI GPT-5.6 and later1,024
  4. Claude Opus 4.72,048
  5. Gemini 2.5 Pro and Flash2,048
  6. Gemini 4 Argon, 3.8 Flash4,096
  7. Claude Opus 4.6, Haiku 4.54,096

[2] Anthropic docs, Checked 9 November 2026; [5] OpenAI docs, Checked 9 November 2026; [7] Google Gemini API docs, Checked 9 November 2026

All 14 more prompt caching facts
More prompt caching facts
~70%Token cache-hit rate reported by an example OpenAI deployment of an LLM judge. OpenAI notes this illustrates a possible outcome, not a guarantee.[5] OpenAI docs, Checked 9 November 2026
0.025xCache read on Claude Fable 5.1: $0.25 against $10 input.[2] Anthropic docs, Checked 9 November 2026
15 / minRequests per minute above which traffic to one cached prefix at OpenAI can overflow to machines without the cache.[5] OpenAI docs, Checked 9 November 2026
221Tokens: across 10 requests, padding an OpenAI prefix of at least 221 tokens up to the 1,024 minimum is cheaper than not caching it.[5] OpenAI docs, Checked 9 November 2026
128On models before GPT-5.6, OpenAI rounds reported cached_tokens down to a multiple of 128. GPT-5.6 and later report the exact cached boundary.[5] OpenAI docs, Checked 9 November 2026
4Cache breakpoints allowed per request at Anthropic. OpenAI also allows up to four cache writes per request.[2] Anthropic docs, Checked 9 November 2026
20 blocksAnthropic's lookback window: each breakpoint checks at most 20 earlier positions for a cached entry.[2] Anthropic docs, Checked 9 November 2026
$0Cost of a cache breakpoint itself at Anthropic. You pay only for writes, reads and uncached input.[2] Anthropic docs, Checked 9 November 2026
512Minimum cacheable prompt in tokens on Claude Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5. Shorter prompts are processed without caching.[2] Anthropic docs, Checked 9 November 2026
4,096Minimum cacheable prompt in tokens on Claude Opus 4.6, Opus 4.5 and Haiku 4.5, 8x the newest models.[2] Anthropic docs, Checked 9 November 2026
4,096Minimum tokens for an implicit cache hit on Gemini 4 Argon, 3.8 Flash, 3.6 Flash and 3.1 Pro. Gemini 2.5 models need 2,048.[7] Google Gemini API docs, Checked 9 November 2026
OnImplicit caching is enabled by default on all Gemini 2.5 and newer models, with no code change.[7] Google Gemini API docs, Checked 9 November 2026
SecondsTime DeepSeek's disk cache takes to build. It is best effort, with no 100% hit guarantee, and is cleared within hours to days of last use.[9] DeepSeek API docs, Checked 9 November 2026
18Upstream fixes, including one in Anthropic's official SDK, after a study of 120 caching pull requests found 32 mis-integrations in 571 agent projects (0.86 precision).[12] Zhang et al., 26 Oct 2026

Are cache reads cheaper on Straitly?

Yes for Claude, GPT, Gemini and Grok. Straitly bills 40% below the published rate with no fee on top, and caching stacks on that.[13]

Every model: LLM API price index. Setup: caching docs. Agent bills: coding agent cost.

48 of 51
models with a cache read below the official price
40%
median gap below the official cache-read price
Cache read price per million tokens: official and Straitly
MODELOFFICIALSTRAITLYDIFFERENCE
Claude Opus 5.5$0.20$0.1240% below
Claude Fable 5.1$0.25$0.1540% below
GPT-6 Astra$1.00$0.6040% below
GPT-6.1 Sol$0.10$0.06040% below
Gemini 4 Argon$0.10$0.06040% below
Grok 4.7$0.50$0.3040% below
Claude Haiku 5.5$0.010$0.006040% below
GPT-6 Luna$0.010$0.006040% below
Gemini 3.8 Flash$0.075$0.04540% below
DeepSeek V4.1 Flash$0.0030$0.001840% below
Kimi K3$0.30$0.1549% below
GLM-5.3$0.26$0.06973% below
Claude Sonnet 5.5$0.10$0.1220% above
DeepSeek V4 Pro$0.022$0.02620% above
Show all 14 models

Official: Claude [2], GPT [6], Gemini [8], DeepSeek V4.1 Flash [10], others as listed in the price snapshot [15]. Straitly: [14] Straitly, Snapshot 9 November 2026

What do people ask about prompt caching?

Short answers, each drawn from the sources below.

How do I reduce LLM inference costs?

Cache the part of every request you resend. Cache reads cost 0.1x normal input at Anthropic and OpenAI and 0.05x or less on the newest models, and a 2026 study measured 41-80% lower API costs on agent sessions. Then lower the price you pay per token: on Straitly, cache reads are a median 40% below the official price across 51 models.

How much does prompt caching save?

On the cached part of the prompt, up to 90% at Anthropic and OpenAI (reads at 0.1x input), 95% on models with 0.05x reads, and about 98% on DeepSeek. Total savings depend on how much of each request is cached and how often it is reused. A 2026 study measured 41-80% lower API costs on agent sessions.

When does prompt caching cost more than it saves?

When the prefix is written but never read. A 5-minute Anthropic write or 30-minute OpenAI write costs 1.25x, so a single request costs 25% more on the cached part. From the second request in the window, caching is cheaper. A 1-hour Anthropic or 12-hour OpenAI write needs three requests.

Do Gemini and DeepSeek charge for cache writes?

No. Gemini 4 Argon uses implicit caching with no write fee, billing hits at $0.10 per million tokens against $2.00 input. DeepSeek bills a cache miss at the normal input price and a hit at $0.003 per million tokens off-peak on V4.1 Flash.

How long does a prompt cache last?

Anthropic: 5 minutes by default, refreshed free on each hit, or 1 hour at a higher write price. OpenAI GPT-5.6 and later: 30 minutes by default, or 12 hours in the Responses API. DeepSeek clears unused caches within a few hours to a few days. Gemini manages implicit caches itself.

Cache the prefix, pay 40% below list

One OpenAI-compatible API for Claude, GPT, Gemini and Grok. No fee, no subscription, no minimum.

Sources and method

How these numbers were checked

Every price and rule was read from the provider's own documentation on 9 November 2026. Prices are list prices in USD per million tokens for standard (non-batch) requests. Worked examples count input tokens only and assume every request after the write hits the cache. Straitly cache-read prices come from the catalog snapshot of 9 November 2026.

  1. [1]2026: The State of Enterprise AI Menlo Ventures, Nov 2026, survey fielded Sep 2026
  2. [2]Prompt caching Anthropic docs, Checked 9 November 2026
  3. [3]Claude API pricing Anthropic, Checked 9 November 2026
  4. [4]Prompt caching with Claude Anthropic, Page dated 14 Aug 2025
  5. [5]Prompt caching guide OpenAI docs, Checked 9 November 2026
  6. [6]API pricing OpenAI, Checked 9 November 2026
  7. [7]Context caching Google Gemini API docs, Checked 9 November 2026
  8. [8]Gemini Developer API pricing Google, Checked 9 November 2026
  9. [9]Context caching DeepSeek API docs, Checked 9 November 2026
  10. [10]Models and pricing DeepSeek API docs, Checked 9 November 2026
  11. [11]Don't Break the Cache: An Evaluation of Prompt Caching for Long-Horizon Agentic Tasks Lumer et al., arXiv 2601.06007, 9 Jan 2026
  12. [12]Silent and Costly: Characterizing and Detecting Prompt-Caching Mis-Integration in LLM Agents Zhang et al., arXiv 2610.34817, 26 Oct 2026
  13. [13]Straitly pricing Straitly, Checked 9 November 2026
  14. [14]Straitly live model catalog Straitly, Snapshot 9 November 2026
  15. [15]Straitly price snapshot vs official list and OpenRouter (164 models) Straitly, Snapshot 9 November 2026
Show all 15 sources