How Much Do AI Tokens Cost in 2026? Real Pricing Across GPT, Claude, Gemini

You probably spend between $40 and $700 a month on tokens — but nobody tells you which end of that range is yours. The answer hides inside six questions that pricing pages never quite answer. Below, six real ones, with the arithmetic shown.

How much does GPT-6 Astra cost per 1M tokens?

GPT-6 Astra costs $10 per 1M input tokens, $50 per 1M output tokens, and $1 per 1M cached input tokens at standard tier. Context window is 1.05M tokens. Long-context output (over a 272K threshold) drops to $37.50/M — the only model where long context actually lowers your bill.

OpenAI publishes three rates per model, not one. Most calculators flatten them into "input" and "output" and hide the cache column, which is why a quick quote can be off by 10x. The full table for the GPT-6 family looks like this.

The spread inside OpenAI is 62x between gpt-5.6-luna and gpt-5.6-cyber, same company, same API. Most teams reach for the flagship out of habit and pay for reasoning depth their workload never uses.

算账示例：A code-generation workload on GPT-6 Astra — 200 requests/day, 4,000 tokens of context in, 1,500 tokens of code out — runs $8.00/day input + $15.00/day output = $23.00/day, $690/month. Same workload on gpt-5.6-terra: $156/month. That's a 4.4x cut for code that is good enough most of the time.

What does 1M input tokens actually cost on Claude Sonnet 5?

Claude Sonnet 5 charges $2 per 1M input tokens, $10 per 1M output tokens, and $0.20 per 1M cached input tokens. Context window is 500K. The cache hit rate is 10x cheaper than fresh input — but writing the cache costs $2.50/M (5-minute TTL) or $4.00/M (1-hour TTL), both above the fresh rate. Cache a prompt you use once and you paid a premium for nothing.

Anthropic's pricing is the cleanest in the industry because the cache contract is honest: they tell you the write cost up front. Most providers don't. OpenAI's automatic caching doesn't charge a write premium, but the prefix has to match within ~512 tokens for a hit. Anthropic's explicit caching gives you control at the cost of an up-front fee.

For a customer-support chatbot — 1,000 conversations/day, 1,200-token system prompt (identical every time, cacheable), 300 tokens of question + history, 500 tokens out — the math is:

• Cached input: 1,200 × 1,000 × $0.20 ÷ 1M = $0.24/day
• Fresh input: 300 × 1,000 × $2.00 ÷ 1M = $0.60/day
• Output: 500 × 1,000 × $10.00 ÷ 1M = $5.00/day
• Total: $175/month

算账示例：Skip caching entirely on the same workload and the 1,200 system tokens bill at $2.00 instead of $0.20, adding $2.16/day. Total climbs to $240/month. Caching saves 27% here, not the 10x the discount implies, because output dominates the bill and output never caches.

Is DeepSeek really cheaper than GPT-6, or is the cache-hit rate misleading?

DeepSeek V4-Flash is $0.22 per 1M input and $0.66 per 1M output off-peak, with cache hits at $0.007/M — that's 31x cheaper than GPT-6 Astra's cache rate. The catch: peak hours double the rates, to $0.44/$1.32. Peak is 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday — which lands on European business hours. If your traffic is European, budget the doubled rate as your baseline.

The cache-hit number looks too good to be true, so let me show when it holds and when it doesn't. DeepSeek's cache is automatic and prefix-based, like OpenAI's. The $0.007 rate applies only to re-reads of identical context, and only off-peak. A prompt with even one timestamp shifting each call won't hit cache, and your bill reverts to the standard $0.22.

算账示例：Document summarization at 500 docs/day, 8,000 input tokens, 400-token summaries on V4-Flash off-peak: $0.88/day input + $0.13/day output = $1.01/day, $30/month. The same workload on Gemini 2.5 Flash at standard rates is $51/month. DeepSeek wins by 41% — but only if your workload runs off-peak.

How much do AI tokens cost for a typical chatbot in 2026?

A typical production chatbot — 1,000 conversations/day, ~1,500 input tokens and ~500 output tokens per turn — costs $40 to $700 a month depending on model tier. The 17x spread is real and it's not noise. Here's the same workload costed on four tiers with caching enabled.

Most teams default to the strongest model. That's a $515/month mistake if their tickets are mostly password resets. Routing simple requests to a cheap tier and hard ones to the flagship typically cuts spend 40-60% at similar output quality — the table above is the proof, not a hypothetical.

算账示例：A support workload where 70% of tickets are tier-1 (password resets, order lookups, FAQ lookups) and 30% are tier-2 (refund disputes, account troubleshooting). Route tier-1 to Haiku 4.5 ($88/month × 0.7 = $62) and tier-2 to Sonnet 5 ($175/month × 0.3 = $53). Total: $115/month instead of $175. The flagship stays on call for the 30% that needs it.

What about long-context pricing? Does it double the bill?

Yes — long-context pricing typically doubles your bill when you cross the threshold, but the threshold and the multiplier vary by provider. Gemini 3.1 Pro is $2.00/$12.00 up to 200K tokens and $4.00/$18.00 above it. Gemini 2.5 Pro doubles from $1.25 to $2.50 past 200K. GPT-6 Astra is the odd one out — long-context output drops to $37.50/M, a 25% cut from its standard $50/M.

The trap is silent: a RAG pipeline that grows its retrieval window from 50K to 250K can cross the line without anyone changing a config value. No error, no warning. Your next invoice just doubled.

算账示例：A RAG pipeline serving 50K-token retrievals at 200 calls/day on Gemini 3.1 Pro: 50,000 × 200 × $2.00 ÷ 1M × 30 = $600/month. Bump retrieval to 250K tokens and you cross 200K on every call: 250,000 × 200 × $4.00 ÷ 1M × 30 = $1,200/month. The retrieval quality probably improved by 5%. Your bill doubled.

The mitigation is mechanical: cap your retrieval window below the threshold, or summarize retrieved chunks before they hit the prompt. Staying under 200K on Gemini 3.1 Pro keeps you at $2.00 instead of $4.00.

How do I estimate my monthly token bill before signing up?

Use the formula (monthly_tokens ÷ 1,000,000) × rate, multiplied by 30 for daily volumes. Add a 1.3x buffer for output tokens (output is 3-6x input cost, and most people under-estimate it). Then verify with the provider's own tokenizer or counting endpoint before you commit a budget.

The arithmetic:

1. Count your daily input tokens. For a chatbot: (system_prompt_tokens + user_message_tokens) × conversations_per_day. Use the provider's tokenizer for exact counts (OpenAI: tiktoken locally, Anthropic: count_tokens() endpoint, Google: countTokens ahead of time).
2. Count your daily output tokens. (avg_response_tokens) × conversations_per_day. Estimate high — output is the expensive direction.
3. Multiply by 30 to get monthly volume.
4. Apply the rates. (monthly_input ÷ 1M) × input_rate + (monthly_output ÷ 1M) × output_rate.
5. Subtract cache savings if your prefix is stable: replace input_rate with cached_rate for the cacheable portion.

算账示例：A new chatbot product, 500 conversations/day estimated at launch, scaling to 2,000 by month three. At 2,000 conversations: 1,500 input + 500 output = 2,000 tokens/turn × 2,000 turns = 4M input + 1M output per day = 120M input + 30M output per month. On Sonnet 5 with caching: 120 × $0.20 + 30 × $10.00 = $24 + $300 = $324/month. Plan a 1.3x buffer: $420/month. If actual traffic overshoots to 3,000 conversations, you're at $486/month — still predictable.

The 1.3x buffer is the part most forecasts skip. Output tokens tend to grow as users ask harder questions, and a single power user with long context can dominate your bill. Budget the buffer and you won't be the team that gets surprised by a 3x invoice.

Sources

All rates verified 2026-09-06 against each provider's official pricing page. The current page was re-verified on 2026-09-07.

1. OpenAI — platform.openai.com/docs/pricing. GPT-6 Astra $10/$50/$1.00 cached, 1.05M context, long-context output $37.50/M. GPT-5.6 family verified 2026-09-02.
2. Anthropic — docs.anthropic.com. Sonnet 5 $2/$10, cache hit $0.20/M, 5-minute cache write $2.50/M, 1-hour write $4.00/M. count_tokens() endpoint for exact pre-send counts.
3. Google — ai.google.dev/gemini-api/docs/pricing. Gemini 3.1 Pro $2/$12 up to 200K, $4/$18 above; 2.5 Flash $0.30/$2.50; batch 50% off.
4. DeepSeek — api-docs.deepseek.com/quick_start/pricing. V4-Pro $0.66/$1.98 off-peak, V4-Flash $0.22/$0.66 off-peak, both doubling at peak (01:00–04:00 and 06:00–10:00 UTC, Mon–Fri).
5. xAI — docs.x.ai. Grok 4.6 $2/$6, 500K context; Grok 4.6 Fast $0.20/$0.50.
6. Qwen (Alibaba Cloud) — bailian.console.aliyun.com. Published in RMB; Qwen3.8-Flash ¥0.8/¥2.7 per million, converted at 0.1389. Qwen3.5-Flash is the current floor at $0.028/$0.278.

Rates move quarterly. If you're reading this well after the verified date, check the source pages before you commit a budget. And if a single rate card ever tells you AI tokens cost "about X dollars" without naming the model, the tier, and the cache contract — assume it doesn't.