Key findings
- Sonnet 5.5 and GPT-6.1 Sol both list at US$2 input and US$10 output per million tokens. Haiku 5.5 and GPT-6 Luna both list at US$0.10 and US$0.50 for shorter prompts.
- Cache reads cost 5% of the input price on Opus 5.5 and Sonnet 5.5, 2.5% on Fable 5.1, and 10% on most other Claude models. OpenAI charges 10% on most models and 5% on GPT-6.1 Sol and GPT-6.1 Terra.
- Claude 4.7 and later models produce about 30% more tokens for the same text. One builder's classification job used 11.9M input tokens on Haiku 5.5 and 7.9M on GPT-6 Luna.
- Prices step up on long prompts and over time. Haiku 5.5 costs 5x more above 100,000 prompt tokens, GPT-6 models move to long-context rates above 272,000, and Gemini 3.8 Flash's introductory price ends on December 31, 2026.
- Teams that redesigned prompts for the cache saved from 36% (Neradot) to about 10x on input (kern). The largest losses came from caches that missed without any error.
Compare model APIs on cost per completed task, because Anthropic and OpenAI now list the same price per million tokens at matching tiers. Cache hit rate, batch share, and the tokens each model spends on the same job decide the bill.
List prices at matching tiers
Standard paid-tier prices in US dollars per million input and output tokens, checked October 12, 2026. Compare within a tier only. A smaller model that fails your task more often costs more per completed task.
- Flagship: Fable 5.1 US$10 and US$50; GPT-6 Astra US$10 and US$50.
- Default agent tier: Opus 5.5 US$4 and US$20.
- Mid tier: Sonnet 5.5 US$2 and US$10; GPT-6.1 Sol US$2 and US$10.
- Low-cost tier: Haiku 5.5 US$0.10 and US$0.50 up to 100,000 prompt tokens; GPT-6 Luna US$0.10 and US$0.50; Gemini 3.8 Flash US$0.75 and US$3.75 through December 31, 2026.
Caching saves the most and breaks without warning
On the Claude API, a 5-minute cache write costs 1.25x the input price and a 1-hour write costs 2x. A cache hit costs 0.1x on most models, 0.05x on Opus 5.5 and Sonnet 5.5, and 0.025x on Fable 5.1. On most models cache reads also stop counting toward input rate limits, so a high hit rate buys throughput as well as savings.
Prompt order decides the savings. Revefi found a timestamp near the top of its agent's system prompt kept the conversation history out of the cache. Moving the timestamp to the end lifted cache reads to 99% of input tokens. Neradot cut a production support agent's cost by 36% on Claude via Bedrock, and the kern project cut input cost about 10x with three stable breakpoints.
A broken cache raises no error. One builder's agent ran through LiteLLM to Claude Opus 4.6 on Bedrock and billed US$37,901.73 gross, about US$35.6k of it uncached input. Another team found Bedrock double counting cache writes with thinking enabled.
Batch pays off for background work
All three providers take 50% off asynchronous work. Claude's Message Batches API runs on the Claude API and Claude Platform on AWS. Most batches finish within an hour, and every batch ends within 24 hours. OpenAI offers a 24-hour batch window and Flex processing at the same rate, and Gemini discounts batch on its paid tier.
Eran Sandler ran an agent loop through Anthropic's Batch API and measured 90 to 120 seconds per turn. That is too slow for interactive work and a real saving for evaluation suites and fleets of background agents.
Judge models by what a finished task costs
Plotly's DataAnalyticsBench put Haiku 5.5 at US$0.38 for 40 analytics questions, against US$15 for Opus 5.5, and two letter grades better than Haiku 4.5. OutcomeOps found Haiku 4.5 gave the most complete answer of eight Bedrock models on its retrieval workload. Run your own evaluation set before acting on either result.
- Run the same evaluation set on two adjacent models per provider, with caching on.
- Record input, cached input, and output tokens per completed task, summed across its requests.
- Price the batch-eligible share at the batch rate and the rest at the standard rate.
- Re-run the comparison when a model or tokenizer changes.
Methodology
GILD compared the official pricing, caching, and batch pages of Anthropic, OpenAI, and Google, checked October 12, 2026. Prices are standard paid-tier list prices in US dollars per million tokens. Latency and long-context figures come from the independent benchmarks in GILD's Claude API record, including Epoch AI's time-to-first-token measurements on prompts up to one million tokens. Every other benchmark, token count, and savings figure is the measuring team's own, published by Plotly, OutcomeOps, Revefi, Neradot, kern, or the named builders and linked in Sources.
Products in this piece
Sources
- Official documentation
- 10
- Public source
- 9
- Claude API docs: Pricing
- Claude API docs: Prompt caching
- Claude API docs: Batch processing
- Claude API docs: Rate limits
- OpenAI API docs: Pricing
- OpenAI API docs: Prompt caching
- OpenAI API docs: Batch API
- Gemini API docs: Pricing
- Gemini API docs: Context caching
- Gemini API docs: Batch API
- Revefi: How Revefi Reduced Its Data Agent's Spend by 90% (Sep 23, 2026)
- Neradot: Case Study: Cutting LLM Cost by 36% with Prompt Caching (Aug 31, 2026)
- kern: How we cut our agent's API costs by 10x with prompt caching (Apr 6, 2026)
- Hacker News: US$38k AWS Bedrock bill caused by a simple prompt caching miss (Apr 28, 2026)
- Hacker News comment on Bedrock cache-write double counting (Apr 8, 2026)
- Eran Sandler: Batch API is terrible for one agent. It might be great for a fleet. (Apr 27, 2026)
- Hacker News comment comparing Haiku 5.5 and GPT-6 Luna token counts (Oct 8, 2026)
- Hacker News comment on Plotly's DataAnalyticsBench run of Haiku 5.5 (Oct 7, 2026)
- OutcomeOps: You're Probably Using the Wrong Bedrock Model (Apr 1, 2026)