Claude and Gemini are the two APIs where prompt caching changes the math most. Anthropic charges 1.25 times the input price to write a cache and then 0.05 times input to read it on Sonnet 5.5. Google charges a flat cached-token rate plus hourly storage, $0.075 per million tokens and $0.50 per million tokens per hour on Gemini 3.8 Flash. So the cheaper one depends on how long you keep the cache alive, not on the headline list price.
The other split is at the bottom of the ladder. Claude Haiku 5.5 costs $0.10 per million input and $0.50 output, close to Gemini 2.5 Flash-Lite at $0.10 and $0.40, but Google also gives away daily usage on its Flash models while Anthropic gives away nothing on the API. Google has two price increases on the calendar that Anthropic doesn’t. Prices here were pulled from both providers’ own pages on November 2, 2026.
What you can try before paying
Anthropic has no free tier on the Claude Developer Platform. The Free plan exists only for the chat apps, and API billing is prepaid credits with auto-reload and one-year expiry, no signup grant. New organizations also start in an Evaluation tier with rate limits below the published Start tier until they build usage history. The practical minimum to try Claude is a small credit purchase, and we’d treat the first one as the cost of a benchmark.
Google gives away the Flash tier. Gemini 3.8 Flash, 3.6 Flash, 3.5 Flash-Lite, 3.1 Flash-Lite, 2.5 Flash-Lite, Gemini Embedding 2 and most Live and TTS models run on the free tier, with Gemini 4 Argon and 3.1 Pro Preview paid only. The per-model free limits now live only on the AI Studio rate limit screen, and they apply per Google Cloud project with daily resets at midnight Pacific. That’s enough to build and demo a prototype without a card.
How caching changes what you pay
Anthropic bills caching as a multiplier. Writing a 5 minute cache costs 1.25 times the input price and a 1 hour cache costs 2 times, then reads cost 0.1 times input on most models, 0.05 times on Sonnet 5.5 and Opus 5.5, and 0.025 times on Fable 5.1. Minimum cacheable length is 512 tokens on the 5.5 models. There’s no separate storage charge, because the TTL is baked into the write price. Pay once, read cheap.
Google bills caching as a storage product. On Gemini 3.8 Flash, cached tokens cost $0.075 per million on read and the cache itself costs $0.50 per million tokens per hour to keep. On Gemini 3.1 Pro Preview storage is $4.50 per million tokens per hour. Gemini 4 Argon only lists implicit caching at $0.10. A 200,000 token document held for an eight hour workday on 3.8 Flash costs $0.80 in storage before any reads, which adds up across a fleet of documents.
The rate limit side favors Anthropic. Cache read tokens don’t count toward input tokens per minute on current models, so a 2 million ITPM limit with an 80% hit rate behaves like 10 million. Google doesn’t publish its tier limits per model, so you can’t do that arithmetic in advance.
List prices today and after December 1
On raw rates Google is cheaper at every rung today. Gemini 3.8 Flash at $0.75 and $3.75 undercuts Sonnet 5.5 at $2 and $10, and Gemini 4 Argon at $2 and $10 undercuts Opus 5.5 at $4 and $20. Anthropic’s cheapest, Haiku 5.5, matches Flash-Lite on input at $0.10 but costs $0.50 output against $0.40, and Haiku jumps to $0.50 and $2.50 once a prompt passes 100,000 tokens.
The calendar narrows the gap. Argon moves to $4 and $20 on December 1, 2026, which is exactly Opus 5.5 pricing, and 3.8 Flash and 3.6 Flash move to $1.50 and $7.50 in 2027. Gemini 2.5 Pro and 2.5 Flash are deprecated and will route to newer models on December 1, 2026. Anthropic has no scheduled changes. Both give a 50% batch discount. A budget built on today’s Gemini rates is wrong by New Year.
Who keeps your prompts and for how long
Anthropic deletes API inputs and outputs within 30 days by default and won’t use commercial inputs or outputs for training unless you give feedback or otherwise allow it. Flagged policy violations can be kept up to 2 years. A zero data retention agreement replaces the 30 day window for customers who sign one.
Google splits by tier. Unpaid service content may be used to improve products and read by human reviewers, and Google warns you not to submit confidential information there. Paid service prompts aren’t used to improve products and are logged only for a limited period to detect abuse. Grounding with Google Search stores prompts and output for 30 days on both tiers. Link billing before the first customer prompt, not after.
Startup money on each side
Claude Startups is between cohorts, with the current one full and a newsletter signup for the next. Listed benefits are partner discounts and credits worth up to $45,000, redeemed by code in the Console, and up to $100,000 in extra API credits for founders backed by a VC in Anthropic’s partner network. The credits work only on the first-party Claude API.
Google’s offer is bigger and open. The Google for Startups Cloud Program covers 100% of spend up to $250,000 in year one for AI startups and 20% up to $100,000 in year two, usable on Gemini and Gemma but not third-party models. You need Pre-Seed to Series A funding, a company founded within the last three to five years, a working MVP and no prior Google Cloud credits above $5,000. That first-year number alone pays for a lot of Argon.
How we would decide
Run the numbers on your own traffic. If most tokens in a request are a reread system prompt or document, Anthropic’s 0.05x cache reads and cache-aware limits usually beat Gemini’s lower list price. If requests are short, unique and numerous, Gemini Flash-Lite at $0.10 and $0.40 with a free tier for testing is hard to argue with. Paid tiers on both keep your prompts out of training.
The winner: Tie
Google wins on list price, free usage and startup credits. Anthropic wins on caching cost, cache-aware rate limits, a higher entry spend cap and no scheduled price increases. Since the December 1, 2026 Argon increase lands it on Opus 5.5 pricing, the paid gap is closing, and the right pick depends on whether your prompts repeat.
Choose Anthropic if your workload rereads long context, you need rate limits that ignore cached tokens, or you want pricing with no announced increases.
Anthropic API offers (4)Choose the Gemini API if you need a free tier for prototypes, the lowest per-token rates on short unique requests, or qualify for up to $350,000 in Google Cloud credits.
Gemini API offers (4)