AI model prices moved faster in the last two months than in the previous two years. Since late September, Anthropic cut its flagship price, OpenAI shipped three GPT-6 variants at three very different price points, DeepSeek started billing by the clock, and Mistral put a trillion-parameter model on a launch discount.
If you are building with these APIs, the sticker price is the least interesting number on the page. What follows is the full October 2026 price list, every major model in one table, plus the four pricing tricks that actually change your bill.
The full price list: October 2026 API rates
All prices are in US dollars per million tokens, input and output, at standard on-demand rates. Where the provider adds conditions, they are in the notes.
| Model | Input / 1M | Output / 1M | Notes |
|---|---|---|---|
| Anthropic Claude Haiku 5.5 | $0.10 | $0.50 | Prompts up to 100K tokens; $0.50/$2.50 above that |
| OpenAI GPT-6 Luna | $0.10 | $0.50 | Budget tier, launched Sept 22 |
| DeepSeek V4.1 Flash (off-peak) | $0.15 | $0.60 | Cache hits $0.003; peak hours double all rates |
| DeepSeek V4.1 Flash (peak) | $0.30 | $1.20 | Peak: 01:00-04:00 and 06:00-10:00 UTC, weekdays |
| Mistral Large 4 (launch discount) | $0.68 | $2.09 | 50% off; list price $1.36/$4.18 when it ends |
| DeepSeek V4 Pro (off-peak) | $0.66 | $1.98 | Doubles at peak: $1.32/$3.96 |
| Anthropic Claude Sonnet 5.5 | $2.00 | $10.00 | Released Sept 28; same rate as Sonnet 5 |
| OpenAI GPT-6 Sol | $2.00 | $10.00 | Launched Sept 22 |
| OpenAI GPT-6.1 Sol | $2.00 | $10.00 | Released Sept 29; near-Astra scores at Sol pricing |
| Google Gemini 4 Argon | $2.00 | $10.00 | Doubles to $4/$20 after November 30 |
| Anthropic Claude Opus 5.5 | $4.00 | $20.00 | 20% cut from Opus 5’s $5/$25 on Sept 22 |
| Meta Muse Wonder | $4.00 | Not disclosed | 816B-parameter open-weight model |
| Anthropic Claude Opus 5.5 fast mode | $8.00 | $40.00 | Up to 2.5x faster; via Claude Code and platform |
| Anthropic Claude Fable 5.1 | $10.00 | $50.00 | Most capable tier in the Claude lineup |
| OpenAI GPT-6 Astra | ~$10.00 | ~$50.00 | Estimated: OpenAI says 6.1 Sol costs “one-fifth” of Astra; exact rate not published |
For Indian readers, at roughly ₹95 to the dollar: a million Haiku 5.5 input tokens costs about ₹9.50, Opus 5.5 input runs about ₹380, and DeepSeek Flash off-peak input is barely ₹14. These are API rates before any caching or batch discounts.
Haiku 5.5 has a pricing cliff at 100,000 tokens
Read the fine print on Anthropic’s cheapest model, released October 7. Haiku 5.5 costs $0.10 input and $0.50 output per million tokens only while the prompt stays at or under 100,000 tokens. Cross that line and every rate on the request jumps five times, to $0.50 and $2.50.
At maximum effort, Haiku 5.5 scores 39.2% on Terminal-Bench 4.0, which makes it a genuine worker for classification, extraction, and agent subtasks. But if your prompts creep past 100K, you are suddenly paying Sonnet-class money for Haiku-class output. Keep prompts under the line or budget for the cliff.
DeepSeek bills you by the clock
DeepSeek’s pricing, in effect since August 16, is the strangest and most exploitable in the industry. Off-peak, V4.1 Flash costs $0.15 input and $0.60 output per million tokens. During peak windows, 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, every rate doubles. Weekends and Chinese public holidays are off-peak all day.
For batch jobs, nightly re-indexing, or anything that does not need to run the instant a user clicks, scheduling outside those windows halves the bill for the cost of a cron change. The cache line matters even more: repeated input costs $0.003 per million tokens off-peak, about 2% of the fresh rate, so agent loops that replay the same context should be engineered to maximize cache hits.
Cheaper per token is not cheaper per task
Here is the number that breaks naive comparisons. Testing by Artificial Analysis found that Claude Opus 5.5 consumes about 119,000 tokens on a standard test set versus roughly 17,000 for GPT-6 Astra. Opus 5.5 is cheaper per token by a wide margin, yet the task can cost more.
OpenAI’s own framing makes the same point from the other side. GPT-6.1 Sol, at $2/$10, reportedly matches GPT-6 Astra on agentic coding at roughly one-fifth the cost per task: about $5 per science task versus $23 for Astra in the company’s reported numbers. Benchmark your own workload before switching, because the token meter and the invoice do not always agree.
There is one more hidden variable. Models from Claude 4.7 onward use a newer tokenizer that produces roughly 30% more tokens for the same text. A rate cut can still raise your bill if the token count rises with it.
Four levers that actually cut the bill
First, batch processing. Anthropic’s Batch API halves both input and output rates across all models, and the tradeoff is only latency. Second, prompt caching: cached input costs 10% of the base rate on most Claude models, 5% on Opus 5.5 and Sonnet 5.5, and 2.5% on Fable 5.1. Third, right-size the model: Sonnet 5.5 reportedly needs fewer tokens for the same work than Sonnet 5, so most jobs cost up to 30% less at the identical $2/$10 rate. Fourth, watch the calendar: Gemini 4 Argon’s $2/$10 rate doubles after November 30, and Mistral’s 50% Large 4 discount has no announced end date.
The pattern of the last two months is clear. Frontier capability keeps getting cheaper, but the savings go to whoever reads the pricing page instead of the headline. Our breakdown of the Opus 5.5 launch has the full benchmark context, our DeepSeek V4.1-Flash piece covers the coding numbers behind the cheapest row in the table, and the Mistral Large 4 launch story explains Europe’s trillion-parameter bet. For the cloud angle, OpenAI’s models are now on Google Cloud too, which is starting to make the clouds compete on model pricing itself.
