List prices for Claude and OpenAI match to the cent in four tiers, so the cheaper API depends on caching, batch jobs, prompt length and reasoning tokens. This breakdown prices real workloads on both so you can pick the lower bill with numbers, not guesses.
Two API pricing pages can look almost identical and still produce bills that differ by thousands of dollars. That is where Claude and OpenAI sit right now. At list price the two vendors match to the cent in four tiers, so the cheaper option depends on how you call the model, not on whose logo is on the invoice. This breakdown of Claude API vs OpenAI API pricing puts the per-million-token rates side by side, then works out what caching, batch jobs, long prompts and reasoning tokens do to a real monthly bill.
💡 All prices are in USD per million tokens (MTok), taken from each vendor's official pricing page at the time of writing. Rates change often, so check the live pages before you lock in a budget.
Neither vendor is cheaper across the board. Tier for tier, the list prices are either identical or close, and the real gap shows up mostly in three places: cached input, very long prompts and the cheapest small models. If you only compare the headline input rate, you miss most of the story.
Four tiers line up to the cent:
OpenAI's table lists both a GPT 5.6 and a GPT-6 generation, and the Luna and Sol names appear in each with different prices. Check the generation number before you compare two rows. When list prices match this tightly, the savings hide in the details below.
Claude pulls ahead on:
OpenAI pulls ahead on:
The tables group models by role rather than by vendor, so you can read across a tier. Every figure is per million tokens.
| Model | Input | Cached input | Output |
|---|---|---|---|
| Claude Haiku 5.5 (prompts up to 100K) | $0.10 | $0.01 | $0.50 |
| GPT-6 Luna | $0.10 | $0.01 | $0.50 |
| GPT 5.6 Luna | $0.20 | $0.02 | $1.20 |
| GPT 5 Mini | $0.25 | $0.025 | $2.00 |
| GPT 5 Nano | $0.05 | $0.005 | $0.40 |
| Claude 4.5 Haiku | $1.00 | $0.10 | $5.00 |
Claude Haiku 5.5 is the only Claude model with length-based pricing: prompts above 100K tokens bill at $0.50 input and $2.50 output. The older Claude 4.5 Haiku costs ten times more than Haiku 5.5 on both input and output.
For a high-volume classifier or router, the contest is GPT 5 Nano against Haiku 5.5 and GPT-6 Luna. Nano is cheapest per token, but the other two are newer generations, so run a quality check before you assume the lowest rate wins.
| Model | Input | Cached input | Output |
|---|---|---|---|
| Claude Sonnet 5.5 | $2.00 | $0.10 | $10.00 |
| Claude Sonnet 5 | $2.00 | $0.20 | $10.00 |
| GPT-6 Sol | $2.00 | $0.20 | $10.00 |
| GPT 5.6 Terra | $2.00 | $0.20 | $12.00 |
| GPT 5.4 | $2.50 | $0.25 | $15.00 |
| Claude Sonnet 4.6 | $3.00 | $0.30 | $15.00 |
Output is the number that decides this tier, because chat and writing tasks produce long answers. GPT 5.6 Terra charges 20% more per output token than Claude Sonnet 5.5 for the same $2 input rate. GPT 5.4 costs 25% more than Sonnet 5.5 on input and 50% more on output.
Moving from Claude Sonnet 4.6 to Sonnet 5.5 cuts the rate by a third on paper. The tokenizer section below shows why the real saving is smaller.
| Model | Input | Cached input | Output |
|---|---|---|---|
| Claude Opus 5.5 | $4.00 | $0.20 | $20.00 |
| GPT 5.6 Sol | $4.00 | $0.40 | $20.00 |
| Claude Opus 4.7 | $5.00 | $0.50 | $25.00 |
| GPT-5.5 | $5.00 | $0.50 | $30.00 |
| Claude Fable 5.1 | $10.00 | $0.25 | $50.00 |
| GPT-6 Astra | $10.00 | $1.00 | $50.00 |
Claude Opus 5.5 and GPT 5.6 Sol match on input and output. The difference sits in the cached input column: $0.20 against $0.40. GPT-5.5 costs 25% more than Opus 5.5 on input and 50% more on output. Claude Opus 4.7 and Claude Opus 4.6 both list at $5 and $25, so the newest Opus is also the cheapest one.
At the very top, Claude Fable 5.1 and GPT-6 Astra are identical on list price. Claude's cache hit rate of $0.25 against $1.00 makes it four times cheaper for repeated context. Claude Fable 5 shares the $10 and $50 list price but bills cache hits at $1, four times the Fable 5.1 rate.
Four levers move the total more than the rate card does: caching, batching, prompt length and reasoning. Each one favors a different vendor in a different situation.
Caching is where the two vendors diverge most. Claude lets you pay a small premium to write a prompt prefix into a cache, then read it back at a steep discount.
| Cache event | Claude | OpenAI |
|---|---|---|
| Write, 5 minute cache | 1.25x input price | No separate write price listed |
| Write, 1 hour cache | 2x input price | No separate write price listed |
| Read | 0.1x input price (0.05x on Opus 5.5 and Sonnet 5.5, 0.025x on Fable 5.1) | 0.1x on most listed models |
A 5-minute write on Claude pays for itself after a single cache read. If your app sends the same system prompt, tool definitions or reference document with every request, caching cuts that part of the bill by 90% to 97.5% on Claude. OpenAI's 90% discount on cached input sits close behind, but Opus 5.5 and Sonnet 5.5 halve the remaining cost again.
Think of it as mise en place in a restaurant kitchen. Ingredients prepped once serve every order faster and cheaper, and a cached prompt prefix works the same way.
Both vendors take 50% off for work that can wait. Anthropic's Batch API halves input and output rates, so Claude Sonnet 5.5 drops to $1 input and $5 output, and Claude Opus 5.5 drops to $2 and $10. OpenAI's pricing page shows the same 50% reduction on its batch tier and also lists a Flex tier priced at batch rates for requests that tolerate slower responses.
Nightly classification, bulk summaries, dataset labeling and evaluation runs all qualify. Anthropic states that batch and caching discounts can be combined, because the multipliers stack.
💡 Fast mode runs in the opposite direction. Claude Opus 5.5 in fast mode costs $8 input and $40 output, exactly double the standard rate, in return for faster output. If latency is not a problem for your product, do not pay for it.
OpenAI splits its pricing at 272K input tokens. Below that line you pay the short-context rate. Above it the long-context rate applies: $8 input and $30 output for GPT 5.6 Sol, against $4 and $20 below the line. Claude's models from the 4.6 generation onward include the full 1M token window at standard pricing, so a 900K token request bills at the same per-token rate as a 9K one. The exception is Haiku 5.5, which steps up above 100K tokens.
Here is what that means for a request that sends 400K tokens of source code and gets 2,000 tokens back:
Run that 1,000 times and the gap between Opus 5.5 and Sol is $1,620. One caveat: the two vendors count tokens differently, so the same codebase will not measure exactly 400K on both sides.
Rate cards are abstract. Two workloads make them concrete.
Take a support assistant that handles one million requests a month. Each request sends 1,500 input tokens (system prompt plus the customer's message) and returns 500 output tokens. That is 1,500 MTok in and 500 MTok out per month, with no caching and no batching.
| Model | Math | Monthly cost |
|---|---|---|
| GPT 5 Nano | 1,500 × $0.05 + 500 × $0.40 | $275 |
| Claude Haiku 5.5 | 1,500 × $0.10 + 500 × $0.50 | $400 |
| GPT-6 Luna | 1,500 × $0.10 + 500 × $0.50 | $400 |
| GPT 5.6 Luna | 1,500 × $0.20 + 500 × $1.20 | $900 |
| Claude Sonnet 5.5 | 1,500 × $2 + 500 × $10 | $8,000 |
| GPT-6 Sol | 1,500 × $2 + 500 × $10 | $8,000 |
| GPT 5.6 Terra | 1,500 × $2 + 500 × $12 | $9,000 |
| GPT 5.4 | 1,500 × $2.50 + 500 × $15 | $11,250 |
| Claude Opus 5.5 | 1,500 × $4 + 500 × $20 | $16,000 |
| GPT 5.6 Sol | 1,500 × $4 + 500 × $20 | $16,000 |
| GPT-5.5 | 1,500 × $5 + 500 × $30 | $22,500 |
Without caching the two vendors tie almost everywhere. The one clear loser is GPT-5.5, which costs $6,500 a month more than Opus 5.5 for the same traffic. GPT 5.6 Terra adds $1,000 over Sonnet 5.5, all of it from the output rate.
Now suppose 1,200 of those 1,500 input tokens are a fixed system prompt that hits the cache on nearly every request. Claude Sonnet 5.5 lands at $5,720, GPT-6 Sol at $5,840 and GPT 5.6 Terra at $6,840. That puts Sonnet $120 ahead of Sol and $1,120 ahead of Terra, before the small cache write charge. Output tokens still dominate, since $5,000 of Sonnet's $5,720 is output, which is why the gap to Sol stays small.
Now a research assistant that answers questions about one 20,000-token reference document. It handles 100,000 questions a month and the document is identical every time, so the repeated prefix adds up to 2,000 MTok. Output is billed the same way on both sides, so compare the input side only.
| Setup | Cost of the repeated document |
|---|---|
| Claude Opus 5.5, no caching | $8,000 |
| Claude Opus 5.5, cache reads | $400 |
| GPT 5.6 Sol, cached input | $800 |
| Claude Sonnet 5.5, cache reads | $200 |
| GPT 5.6 Terra, cached input | $400 |
The Claude figures assume the cache stays warm. At about 2.3 requests a minute it does, because every read refreshes the 5-minute window. A 1.25x write charge each time the cache goes cold barely moves the total.
On the same tier, Claude's cache reads cost half of OpenAI's. For a read-heavy product that is hundreds of dollars a month at this scale and thousands at larger ones.
List prices only describe the tokens you plan to send. Three things quietly change how many tokens you pay for.
Thinking is billed as output. On both platforms the tokens a model spends reasoning before it answers are charged at the output rate, the expensive column. A hypothetical request that returns a 300-token answer after 2,000 tokens of reasoning is billed as 2,300 output tokens. On Claude Opus 5.5 that is $0.046 instead of $0.006, almost eight times more for the same visible answer.
Both vendors give you a dial. Claude Sonnet 5 exposes an effort setting where low turns thinking off for the fastest, cheapest responses, and GPT 5.6 Terra exposes a reasoning effort setting that defaults to none. Start at the bottom setting and raise it only when answers get worse.
Anthropic states that Claude models from 4.7 onward use a newer tokenizer that produces roughly 30% more tokens for the same text than the older one. That matters for upgrades inside the Claude line. Sonnet 4.6 to Sonnet 5.5 cuts the rate by a third, but if your text now splits into 30% more tokens, the net saving is closer to 13%.
The same logic applies across vendors. A per-token rate is not a per-word cost, and no public conversion exists between the two tokenizers. The usage object returned with every response lists billed input and output tokens, so measure instead of guessing.
Built-in tools carry their own charges. On Claude, web search costs $10 per 1,000 searches plus the tokens the results add to your prompt, web fetch adds no charge beyond tokens, and code execution is free when used alongside web search or fetch. Standalone code execution includes 1,550 free hours a month, then $0.05 per hour per container.
Pinning Claude inference to US-only processing multiplies every token category by 1.1x. OpenAI prices its own built-in tools separately on its pricing page, so add those lines to the same spreadsheet before you compare totals.
You now have the rates, the multipliers and two worked examples. The last step is your own traffic.
Pull 50 real requests from your logs, the messy ones included. Send each to two candidates at the same tier. Record billed input tokens, billed output tokens and whether the answer was good enough to ship. Multiply by each vendor's rates, including cached and batch rates where they apply.
The model with the lowest cost per acceptable answer is the cheaper one, even when its rate card looks worse. A model that needs two attempts to get the answer right costs double, whatever the price per million tokens says.
| Your workload | Likely cheaper | Why |
|---|---|---|
| Millions of short classification calls | GPT 5 Nano, then Haiku 5.5 or GPT-6 Luna | Lowest list rates |
| Chat with a long fixed system prompt | Claude Sonnet 5.5 or Opus 5.5 | Cache reads at 5% of input price |
| Whole repositories or contracts above 272K tokens | Claude Opus 5.5 | One rate up to 1M tokens |
| Overnight bulk jobs | Either | Both take 50% off, so pick on quality |
| Output-heavy writing in the mid-range | Claude Sonnet 5.5 or GPT-6 Sol | $10 output against $12 on Terra |
| Top-tier reasoning with repeated context | Claude Fable 5.1 | Cache reads at $0.25 against $1.00 |
Before you wire an SDK into your product, run the same prompt through both vendors' models in the browser. PicassoIA hosts both Claude Sonnet 5 and GPT 5.6 Terra in its large language model collection, so you can compare answers side by side with no code.
💡 Both models accept images, so you can attach a screenshot to test vision tasks too. Image input adds tokens: the Claude Sonnet 5 page lists it as width × height ÷ 750 input tokens.
The habit that saves money on text models also works for image generation: test small, adjust, then scale. Write one prompt, render it, change the lighting or the lens, render again. The photos in this article were all generated from text prompts and a few quick revisions.
Open PicassoIA's model library, pick a text-to-image model and describe the scene you want: subject, light, lens, mood. Then run the same prompt on a second model and compare the results. A few minutes of experimenting shows you which model fits your style, and every render is a chance to build a visual that makes your next article, product page or campaign stand out.
