Prompt Caching ComparedOpus 5.5 vs GPT-6.1 Sol vs Gemini 4 Argon

Why 95% cache discounts still produce 2x to 5x bills

1 Sol, and Gemini 4 Argon all advertise a ninety five percent discount on cached input tokens, dropping two to four dollars per million tokens down to ten or twenty cents. Yet in real coding agents, teams turn prompt caching on and still watch their API bills jump two times or even five times higher than expected.

Read discounts are only one line on the invoice, while write multipliers, minimum token floors, long-context cliffs, and silent cache invalidators decide what you actually pay. Here is what prompt caching stores under the hood, how the math compares across Anthropic, OpenAI, and Google, the four traps that wipe out your savings, and which API wins for each workload.

What prompt caching stores under the hood (the KV prefix)

Start with what the API is actually caching. Every time you send a prompt, the server computes attention key and value tensors, the K V cache, from the very first token forward. Without caching, carrying a 100K token repository snapshot across ten agent turns means recomputing and paying full input rates for that same 100K prefix ten times. Prompt caching saves those K V tensors on the provider side so matching requests skip the prefill math.

Base input vs cache read rates: Anthropic, OpenAI & Google

Because K V states depend on every earlier token, matching is strictly prefix-based from token zero. Change one word near the top, like injecting a live clock timestamp into your system prompt, and every token after it misses the cache. Look at the headline read prices side by side.

5 costs four dollars per million base input tokens and twenty dollars output across a one million token window with 128K max output, and cache hits drop to zero point zero five times base, twenty cents per million. 5 is two dollars input and ten dollars output, with ten-cent cache hits.

1 is ten dollars input and fifty dollars output, cutting cache hits to zero point zero two five times base, twenty five cents per million, ninety seven point five percent off.

Trap 1: Cache write multipliers and TTL windows (5m, 30m, 1h, 12h)

1 Terra is one dollar input and five dollars output with five-cent reads, both ninety five percent off, while GPT-6 Astra is ten dollars input with one-dollar reads and GPT-6 Luna is ten cents input with one-cent reads.

On Google's paid tier, Gemini 4 Argon offers a one million token window at two dollars input, ten dollars output including thinking tokens, and ten-cent cached reads through November thirty, 2026, doubling on December one, 2026 to four dollars input, twenty dollars output, and twenty cents cached. But here is the first trap most developers miss, the cost of writing the cache. 5.

A five-minute write loses money unless at least one follow-up call hits it in time. 1 Sol, Terra, GPT-6 Astra, and Luna, except OpenAI's default minimum lifetime is thirty minutes, six times longer for the same markup, plus a twelve-hour minimum T T L at two point five times input. OpenAI's explicit mode also lets you mix twelve-hour and thirty-minute breakpoints in one request while leaving the trailing user message uncached with zero write fee.

Google takes the opposite route on Gemini 4 Argon.

Trap 2: Minimum token floors and Gemini 4 Argon's 2,050-token turn rule

8 Flash, which supports explicit cache objects at fifty cents per million tokens per hour through December thirty one, 2026. Trap number two is the minimum token floor, plus an agent-loop rule unique to Gemini. 5. Shorter prompts marked with cache control run uncached with no error, silently returning zero in cache creation and cache read tokens.

OpenAI requires one thousand twenty four tokens, excluding hidden system tokens. 8 Flash. And in multi-turn loops, Gemini 4 Argon caches the full previous request only when the new turn adds at least two thousand fifty tokens, counting model output, thoughts, and function results.

Trap 3: Haiku 5.5's 100K cliff and OpenAI's 272K context cliff

Smaller turns re-process the last two thousand fifty tokens at standard input rates. Check usage total cached tokens to verify real hits. Trap number three hits when a coding session crosses 100K or 272K tokens. 1 keep flat per-token pricing across the entire one million token window.

5. 5 jumps five times to fifty cents input, two dollars fifty cents output, and five cents cached. 1 Sol from two dollars to four dollars input, raises output from ten to fifteen dollars, and doubles cached reads from ten cents to twenty cents, while GPT-6 Astra doubles input from ten to twenty dollars.

Trap 4: Silent cache invalidators and parallel subagents

Trap number four is busting your own cache without touching your prompt text. Anthropic caches in strict order: tools, then system, then messages, with automatic ephemeral mode or up to four breakpoints and a twenty-block lookback window. Edit one tool schema and you wipe the whole cache. Toggle web search, citations, or fast speed and you wipe system and messages.

Change tool choice or output effort and you wipe messages. Also, a new Claude cache entry only goes live after the first response starts streaming, so wait for stream start before firing parallel subagents.

Real 10-turn agent math on a 100K repo prefix + 50% Batch API

On OpenAI, changing model, tools, parallel tool calls, structured outputs format, reasoning effort, text verbosity, or context management alters the prefix, which you can inspect in the Prompt Cache Diagnostics tool. Run the ten-turn math on a 100K repository prefix inside five minutes. 5. 1 Sol pays twenty five cents to write turn one plus nine one-cent reads, thirty four cents total, eighty three percent off.

5 drops from four dollars to sixty eight cents. When turns add over two thousand fifty tokens, Gemini 4 Argon pays twenty cents on turn one plus nine one-cent hits, twenty nine cents through November thirty, 2026. 1 Sol and Gemini 4 Argon.

Which API cache to pick for your agent

So here is the rule. 5 for flat pricing, and warm one stream before parallel fan-out. 1 Sol with mixed twelve-hour and thirty-minute breakpoints. For zero-config implicit caching on chunky multi-turn loops above four thousand ninety six tokens, pick Gemini 4 Argon.

So, recap. Lock static tools, instructions, and repo context at token zero and push dynamic text to the end.

Four-point checklist recap

Watch write fees, 512 to four thousand ninety six token floors, Haiku's 100K cliff, OpenAI's 272K cliff, and Gemini's two thousand fifty token turn rule, and check usage metrics every run. If this helped, subscribe. Build with AI. For real.