Haiku 5.5 vs Gemini 3.8 Flash
For 20,000 uncached input tokens and 2,000 billable output tokens, Haiku 5.5 costs $0.003 per request. Gemini 3.8 Flash costs $0.0225 at its promotional Standard rate through December 31, 2026. That is an equal-token budget, not a benchmark of completed tasks.
Compare this workload →Standard API prices
USD per million tokens. Claude uses global routing; Gemini uses the paid Developer API. Sources: Anthropic pricing and Google pricing.
| Category | Haiku ≤100K prompt | Haiku >100K prompt | Gemini through Dec 31 | Gemini from Jan 1, 2027 |
|---|---|---|---|---|
| Fresh input | $0.10 | $0.50 | $0.75 | $1.50 |
| Output, including thinking | $0.50 | $2.50 | $3.75 | $7.50 |
| Cache reads | $0.01 | $0.05 | $0.075 | $0.15 |
| 5-minute cache writes | $0.125 | $0.625 | Google uses a separate cache-storage charge | |
| 1-hour cache writes | $0.20 | $1.00 | See storage example below | |
Gemini's promotional input and output prices are 7.5 times Haiku's short-prompt prices, but only 1.5 times its long-prompt prices. Each provider tokenizes text differently. These ratios do not establish savings on the same document or task.
Choose a model for a checkable task
For classification, extraction or short summaries, start by testing Haiku against your required labels and fields. Anthropic positions it for high-volume, bounded work. Google positions Gemini 3.8 Flash for software engineering and agent workflows; its documented inputs include text, images, video, audio and PDFs. Those descriptions come from the providers and are not comparative accuracy results.
A support-routing trial could use a fixed set of messages with known destinations. Count wrong routes and cases sent to a human, then compare spend per acceptable result. For repository work, require the same tests and review checks for both models. Extra retries and growing conversation history belong in the budget.
Haiku has a 1M context window and up to 128K output tokens. Gemini allows 1,048,576 input tokens and 65,536 output tokens. Capacity tells you whether a request fits; it does not tell you whether the answer will meet your requirements. Gemini 3.8 Flash returns text, so image input support should not be mistaken for image generation.
Budget for 10,000 requests
Assume every request uses 20K fresh input and 2K output, with no cache or paid tools. Haiku totals $30. Gemini totals $225 through December 31 and $450 from January 1, 2027. Haiku's calculation is (20,000 × $0.10 + 2,000 × $0.50) ÷ 1,000,000 × 10,000.
The preselected calculator reproduces the current Standard totals. It includes billable output and Haiku's prompt band. Cache writes, storage, Batch/Flex/Priority, paid tools and taxes are excluded. Its displayed catalog is dated; use the future-price column above for a 2027 budget.
Suppose 10% of initial Haiku calls require a Gemini retry with the same token counts. Under that hypothetical routing plan, the promotional bill is $30 + 1,000 × $0.0225 = $52.50. The 10% rate is an assumption, not a measured success rate. A retry that includes the earlier answer can consume more tokens.
The 100K boundary and caching
At 100,000 input and 2,000 output tokens, Haiku costs $0.011. At 100,001 input, it costs $0.0550005: the whole request moves to the higher band, including output. Count fresh input, cache reads and writes together. At 120K input and 2K output, Haiku costs $0.065 versus Gemini's promotional $0.0975.
If 18K of the 20K prompt tokens are cache reads and 2K remain fresh, the original request costs $0.00138 on Haiku and $0.01035 on Gemini. These figures assume an existing valid cache. Claude charges to create it; Google charges for storage as well as reads.
For example, keeping 100,000 Gemini cached tokens for one hour adds $0.05 at the promotional storage rate of $0.50 per million tokens per hour. From January 1, that storage example costs $0.10. Storage is charged over its lifetime, so multiply by hours retained, not by the number of hits. Google's Batch and Flex discounts do not halve this storage rate.
Batch, Flex and Priority
Haiku Batch halves eligible token rates in either prompt band. For Gemini during the promotion, Batch and Flex each use $0.375 input, $1.875 output and $0.0375 cache reads per million. Priority uses $1.35, $6.75 and $0.135 respectively. All these Google token prices double on January 1, 2027; the published storage rate also doubles.
The 10K-request uncached example is $15 on Haiku Batch, $112.50 on Gemini Batch or Flex, or $405 on Gemini Priority during the promotion. Choose processing mode for your response-time needs. Google Search grounding and other tools can add charges; Claude US-only inference adds 10% to token prices.
Why a subscription test can show zero cost
In a community comparison, the author says Haiku ran through a subscription and therefore displayed zero incremental cost. That does not make the API free. The reported synthetic tasks and effort settings are useful context for designing a trial, but cannot establish your invoice or a universal winner.
Record actual usage for every attempt, including thinking output, and divide total spend by accepted results. Check the Haiku pricing guide for detailed cache-write accounting, or compare the Gemini catalog entry. API plans, free-tier limits and consumer subscriptions are separate products.