LLM API Prices Compared: Claude, GPT, Gemini, DeepSeek

Per million tokens, the four vendors we checked land in three bands. Flagships run $10 in and $50 out (Claude Fable 5.1, GPT-6 Astra). Workhorses run about $2 and $10 (Claude Sonnet 5.5, GPT-6.1 Sol, Gemini 4 Argon until 1 December). Small models run $0.10 to $0.30 in, and DeepSeek sits below all of them. Prices checked 2026-10-31.

Prices checked 2026-10-31

All prices are USD per million tokens at standard speed, from each vendor's own pricing page, fetched on the check date. "Short prompt" means under the vendor's long-prompt threshold. Where a price has a stated end date, the later price is shown too.

Flagship tier

Workhorse tier

Small and cheap tier

One vendor page per row of this table is a better source than this table. Go to Claude API pricing, OpenAI API pricing or Gemini API pricing for the full lists. For the models, see Claude models, GPT models, Gemini models and DeepSeek models.

How to read it

Tiers are the story

The vendors have converged on the same shape. A top model priced around $10 and $50, a mid model near $2 and $10, and a small model at $0.10 and $0.50. Claude Sonnet 5.5 and GPT-6.1 Sol are identical on input ($2), cached input ($0.10) and output ($10). That is not proof they perform alike. It means price alone will not separate them, so use a task test.

Cache discounts differ more than rates

For agent loops that re-send a long conversation every turn, cache pricing matters more than the headline input rate. DeepSeek's cache-hit rates are the cheapest on the list.

Long-prompt thresholds

If your prompts are routinely long, OpenAI's 272K rule and Haiku 5.5's 100K rule can double your effective price.

Time-of-day pricing, DeepSeek only

DeepSeek charges peak and off-peak rates, and off-peak is half. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays. Everything else, weekends included, is off-peak. A batch job scheduled at 12:00 UTC pays half the rate of one run at 08:00. It is the only vendor in this comparison that discounts for when you call it.

Batch and flex

Use batch for anything a person is not waiting on: evals, backfills, bulk classification and nightly summaries.

What the table leaves out

• Thinking and reasoning tokens. Gemini states that thinking is billed as output, and OpenAI lists reasoning token support on its GPT-6 models with effort settings up to max. Price per visible token understates cost on reasoning models.
• Tokenizers. Claude says its 4.7-and-later tokenizer yields about 30 percent more tokens than earlier models on the same text. Other vendors do not publish a comparable figure, so identical rates do not mean identical bills.
• Tool charges. Web search is $10 per 1,000 searches on Claude and $10 per 1,000 calls on OpenAI. On Gemini 3 and newer it is $14 per 1,000 after 5,000 free a month. DeepSeek's Search and Fetch are $0.75 per 1,000 successful requests, free until 1 November 2026 when used through certain tools.
• Free tiers. Gemini has a free tier on Flash models, with the caveat that free-tier content is used to improve Google's products.
• Rate limits and spend caps. Claude's Start tier caps spend at $500 a month. Check each vendor's tiers before planning a launch.
• Self-hosted open models. Prices for Llama, Qwen, Mistral and others depend on the host. See hosting, Replicate, Hugging Face Inference and AWS Bedrock.

One request, four vendors

Our arithmetic from the tables, for 50,000 input tokens and 2,000 output tokens, with no caching and standard prices.

Rate cards are not outcomes. A model that takes two attempts costs twice, and a reasoning model that thinks for ten thousand tokens costs more than the sum above.

What changed recently

Prices moved in several directions this autumn, which is why last quarter's spreadsheet is probably wrong.

• Claude Opus fell from $5 and $25 (Opus 5 and the 4.x line) to $4 and $20 on Opus 5.5, and Opus 5.5 cache reads are priced at 0.05x input.
• Claude Sonnet 5 stayed at $2 and $10. The planned rise to $3 and $15 on 1 September 2026 did not happen, and Sonnet 5.5 holds the same price.
• Gemini 4 Argon launched at $2 and $10 with a promotional end date. From 1 December 2026 the page lists $4 and $20.
• Gemini 3.8 Flash and 3.6 Flash carry $0.75 and $3.75 until 31 December 2026, then double.
• DeepSeek now prices by time of day, with off-peak at half the peak rate.
• DeepSeek Search and Fetch become billable at $0.75 per 1,000 successful requests from 00:00 Beijing time on 1 November 2026, when used through tools such as Claude Code or Codex. Until then they are free.

Monthly bill scenarios

Using the per-request figures above (50,000 input tokens, 2,000 output tokens, no caching), at 100,000 requests a month. This is arithmetic, and your own token counts will differ.

Now apply what the table cannot show. If 90 percent of each prompt is a cache hit, the input half of a Sonnet 5.5 bill falls from $10,000 to about $1,450 (5,000 uncached tokens at $2 plus 45,000 cached at $0.10, times 100,000 requests, ignoring cache writes), so the monthly bill drops from $12,000 to about $3,450. Send the same traffic through batch and the figure halves again. Caching and batch together can move a bill more than switching vendors does.

Open-weight models and hosted prices

The four vendors above all sell their own models. Open-weight models such as Llama, Qwen, Mistral and DeepSeek's own weights can also be run by third-party hosts, and the price varies by host and by hardware. Hosts we track include RunPod, Modal, Lambda and Cloudflare Workers AI. Our open-source LLM page lists the models themselves. We have not compiled a per-host token price table, so this page does not claim one.

Who should pick which

Our take

The market has settled into a pattern, and in each tier the vendors are close enough that price is the wrong first question. Pick two models per tier, run twenty of your real tasks through each, and compare cost per finished task, including retries. That test costs a few dollars and settles what no rate card can.

Two warnings. First, build your forecast on announced future prices, which for Gemini means the December and January numbers. Second, read the footnotes: long-prompt thresholds, tokenizer changes and peak-hour pricing move the real price by 2x or more. We have not run our own quality comparison across these vendors and make no claim about it.

For the tools that sit on top of these APIs, see AI coding agent pricing compared. To switch between models behind one key, see OpenRouter alternatives. For a head-to-head on two vendors, see Anthropic vs OpenAI API.

Checklist

• Choose a tier first, then compare two models inside it.
• Forecast at post-promotion prices.
• Measure cost per finished task on your own data.
• Turn on caching for repeated prefixes and batch for async work.
• Check the long-prompt threshold before sending big contexts.
• Re-read each vendor page before every budget review.

How to run a fair price test

A fair test takes an afternoon. Pick twenty real tasks from your product, including some hard ones. Run each through two models per tier with the same prompt, the same tools and the same reasoning effort settings. Record the input tokens, output tokens including reasoning, the number of attempts, and whether the answer was acceptable. Multiply tokens by each vendor's price, including cached rates for the repeated prefix, and divide by the number of acceptable answers. That is your cost per finished task.

Do it twice, once at current promotional prices and once at the post-promotion prices announced for Gemini. Then check the long-prompt thresholds: if any of your prompts cross 100K tokens on Haiku 5.5, 200K on Gemini Pro models or 272K on GPT-6 models, price those separately.

Finally decide what failure costs. A model that is half the price but needs a human fix a fifth of the time may be dearer overall. The rate card cannot show this, and no vendor page tries. Keep the test set and rerun it whenever a vendor announces a price or model change, since this autumn has had several.

Why the tiers keep converging

The three-band structure is no accident. Vendors price against each other's closest model, and a buyer who sees $2 and $10 at one vendor will test it against $2 and $10 at another. The practical result is that price stops being the deciding variable inside a tier, and decisions move to tool support, latency, rate limits, data terms and how a model behaves on your tasks. Spend your evaluation time there. The one place price still differs by a wide margin is DeepSeek, which sits well below the other three at peak and half that off-peak, so it is worth a test if cost dominates and its terms suit your data.