On October 21, 2026, OpenAI API pricing lists GPT-6 Luna at $0.10 per million input tokens and GPT-6 Astra at $10. Some per-token prices have fallen since July and one rose. What you pay depends far more on which tier you call and how many tokens you send.
What happened: some prices fell, one rose
Since July, Anthropic's Opus tier and OpenAI's GPT-5.6 Sol got cheaper, while DeepSeek's budget model got dearer. Anthropic pricing shows Claude Opus 5.5 at $4 input and $20 output per million tokens, 20% below Opus 5's $5 and $25. GPT-5.6 Sol, listed in OpenAI's Cyber models table, fell from $5 input and $30 output in July to a promotional $4 and $20. DeepSeek V4.1 Flash, which replaced V4 Flash, costs $0.15 input and $0.6 output off-peak, versus V4 Flash's $0.14 and $0.28.
- Key number: OpenAI's flagship ladder runs from $0.50 to $50 per million output tokens, a 100x gap.
- Who's affected: founders and PMs paying API bills, especially for agents and background features.
- What to do: recheck your model per task, and move simple work to a cheap tier.
Over time the decline is steep. Epoch AI's price trends find that the price of reaching a given performance level falls "from 9x to 900x per year", depending on the milestone.
The same task costs up to 100x more depending on tier
GPT-6 Astra's $50 per million output tokens is 100x GPT-6 Luna's $0.50, and Claude Opus 5.5's $20 is 40x Haiku 5.5's $0.50.
Each row shows the vendor's standard list price per million tokens, checked October 21, 2026, next to the July price where research confirms a change. Use the LLM pricing tracker for models not listed here.
| Model | July 2026 | Now (input / output per 1M tokens) | Tier |
|---|---|---|---|
| GPT-6 Astra | not recorded | $10 / $50 | OpenAI top flagship |
| GPT-6.1 Sol | not recorded | $2 / $10 | OpenAI middle |
| GPT-6 Luna | not recorded | $0.10 / $0.50 | OpenAI cheapest |
| Claude Opus 5.5 | Opus 5: $5 / $25 | $4 / $20 | Anthropic Opus tier |
| Claude Sonnet 5.5 | not recorded | $2 / $10 | Anthropic middle |
| Claude Haiku 5.5 | not recorded | $0.10 / $0.50 | Anthropic cheapest, prompts up to 100,000 tokens |
| Kimi K3 | $3 / $15 (unchanged) | $3 / $15 | Kimi flagship |
| Grok 4.7 | Grok 4.5 (price not confirmed) | $2 / $6 (labels unconfirmed) | xAI |
| DeepSeek V4.1 Flash | V4 Flash: $0.14 / $0.28 | $0.15 / $0.6 off-peak | DeepSeek budget |
DeepSeek charges double during peak hours, which run 01:00–04:00 and 06:00–10:00 UTC on weekdays.
Why bills still rise when tokens get cheaper
Bills can rise even as tokens get cheaper, because the bill is price per token times tokens used, and usage can outrun price cuts. If the price per token falls 4x while your feature's token use grows 10x, the bill ends up 2.5x larger.
Tier choice alone can move the bill 100x. Twenty million output tokens a day cost $10 on GPT-6 Luna and $1,000 on GPT-6 Astra. Over 30 days that is $300 versus $30,000 for the same volume.
CloudZero, which sells cloud cost tools, takes the vendor view: "Prices keep falling. Bills keep rising." It also claims routing saves 60-90%, a vendor figure we haven't checked.
Route by task: a decision rule
Match the tier to the job, then cut repeat input with caching. On OpenAI's page, cached input is 90% cheaper than fresh input on both Luna and Astra.
- Extraction, tagging, short replies: start on GPT-6 Luna, Claude Haiku 5.5 or DeepSeek V4.1 Flash.
- Drafting and routine coding: use a middle tier such as GPT-6.1 Sol or Claude Sonnet 5.5.
- Hard reasoning a person will act on: pay for Astra or Opus 5.5, and only there.
- Batch jobs on DeepSeek: schedule them outside peak hours to pay half.
- By November 21, 2026: if you use GPT-5.6 Sol (now $4 input / $20 output on OpenAI's Cyber models table), recheck its price; OpenAI says the promotional pricing runs "at least through" that date.
Before you switch, run your own prompts through the price check tool and compare LLM inference tools.
Related
What's still unconfirmed
- GPT-5.6 Terra and Luna prices. The earlier figures no longer appear on OpenAI's current pricing page.
- Grok 4.7's input and output split. xAI's page lists $2 and $6 per million tokens, but our extract didn't label them.
- Routing savings. The 60-90% figure is CloudZero's claim, not an independent measurement.
