HopBaseThe short answer: on official list prices checked on 2026-10-31, the cheapest capable models per million tokens are GPT-6 Luna ($0.10 input / $0.50 output), Claude Haiku 5.5 ($0.10 / $0.50 for prompts up to 100K tokens) and GLM-5.3 Flash ($0.15 / $0.50). Mid-tier flagships cluster around $2 input and $10 output: Claude Sonnet 5.5, GPT-6 Sol, GPT-6.1 Sol and Gemini 4 Argon all list at exactly $2 / $10. The top tier, Claude Fable 5.1 and GPT-6 Astra, lists at $10 / $50. Headline prices are only half the story. Cached input is 4 to 50 times cheaper than standard input, depending on the model. GPT, Grok and Haiku 5.5 charge more once a prompt crosses a length threshold, and DeepSeek halves its prices off-peak. The tables below link every price to its source page, and the pricing page lists every model.
How to read these tables
- Official list prices only, in the currency the source page uses, per 1 million tokens. They are not HopBase rates and include no discounts. Your own HopBase rate is shown in the console.
- Checked on 2026-10-31. Vendors change prices often, and two models below have announced increases. Follow the source link before you budget.
- Three columns: standard input, cached input (a prompt-cache read), and output. Thinking tokens bill as output (Gemini and Kimi K3 state this explicitly).
- Cache writes and storage fees are covered in the prompt caching section below.
Western frontier models
| Model | Input | Cached input | Output | Notes | Source |
|---|---|---|---|---|---|
| Claude Fable 5.1 | $10 | $0.25 | $50 | Cache write $12.50 | claude.com/pricing |
| Claude Opus 5.5 | $4 | $0.20 | $20 | Cache write $5 | claude.com/pricing |
| Claude Sonnet 5.5 | $2 | $0.10 | $10 | Cache write $2.50 | claude.com/pricing |
| Claude Haiku 5.5 | $0.10 | $0.01 | $0.50 | Prompts over 100K tokens: $0.50 / $0.05 / $2.50 | claude.com/pricing |
| GPT-6 Astra | $10 | $1 | $50 | Over 272K input: $20 / $2 / $75 | OpenAI pricing |
| GPT-6.1 Sol | $2 | $0.10 | $10 | Over 272K input: $4 / $0.20 / $15 | OpenAI pricing |
| GPT-6 Sol | $2 | $0.20 | $10 | Over 272K input: $4 / $0.40 / $15 | OpenAI pricing |
| GPT-6.1 Terra | $1 | $0.05 | $5 | Over 272K input: $2 / $0.10 / $7.50 | OpenAI pricing |
| GPT-6 Luna | $0.10 | $0.01 | $0.50 | Over 272K input: $0.20 / $0.02 / $0.75 | OpenAI pricing |
| Gemini 4 Argon | $2 | $0.10 | $10 | Rises to $4 / $0.20 / $20 on 2026-12-01 | Gemini API pricing |
| Gemini 3.8 Flash | $0.75 | $0.075 | $3.75 | Rises to $1.50 / $0.15 / $7.50 on 2027-01-01; cache storage $0.50 per 1M tokens per hour | Gemini API pricing |
| Grok 4.7 | $2 | $0.50 | $6 | Prompts of 200K tokens or more: $4 / $1 / $12 for the whole request; 500K context | xAI models |
All prices are in USD per 1M tokens, at the standard tier. OpenAI also lists Batch, Flex and priority tiers, which are not shown here.
Chinese models
| Model | Input | Cached input | Output | Notes | Source |
|---|---|---|---|---|---|
| DeepSeek V4.1 Flash (1017) | $0.30 | $0.006 | $1.20 | Peak rate; off-peak is half | DeepSeek pricing |
| DeepSeek V4 Pro (0813) | $1.32 | $0.044 | $3.96 | Peak rate on DeepSeek's page | DeepSeek pricing |
| GLM-5.3 | $1.40 | $0.26 | $4.40 | 1M context | Z.ai pricing |
| GLM-5.3 Flash | $0.15 | $0.03 | $0.50 | 1M context, natively multimodal | Z.ai pricing |
| Kimi K3 | $3 | $0.30 | $15 | 1M context | hop-base.com/pricing (Moonshot official list) |
| Qwen 3.8 Max | ¥12 | ¥1.50 | ¥36 | Single tier, 1M context | hop-base.com/pricing (Alibaba official list, CNY) |
| Qwen 3.8 Flash | ¥1 | ¥0.10 | ¥3 | Single tier, 1M context | hop-base.com/pricing (Alibaba official list, CNY) |
DeepSeek's peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays (09:00 to 12:00 and 14:00 to 18:00 Beijing time). Other hours and weekends bill at half the peak rate, according to DeepSeek's pricing page. Qwen prices are the standard-band list rates. Alibaba's limited-time and off-peak promotions are not reflected. For Doubao Seed, see the Doubao guide and the pricing page.
What the numbers say
- The $2 / $10 tier is crowded. Claude Sonnet 5.5, GPT-6 Sol, GPT-6.1 Sol and Gemini 4 Argon list at exactly the same input and output price. They differ on cache reads ($0.10 to $0.20), context limits and long-prompt surcharges, so the cheapest choice depends on how your prompts are shaped.
- Small models are almost free on input. GPT-6 Luna and Claude Haiku 5.5 list at $0.10 per million input tokens, and DeepSeek V4.1 Flash at $0.30 peak and $0.15 off-peak.
- Output dominates agent workloads. Coding agents and reasoning models produce long outputs, and thinking tokens bill as output. Compare the output column first for those workloads.
- Watch the scheduled increases. Gemini 4 Argon's introductory price ends on 2026-11-30, and Gemini 3.8 Flash doubles on 2027-01-01.
How cache pricing works
Prompt caching bills the repeated prefix of a prompt (system prompt, tool definitions, documents, earlier turns) at a much lower rate when it is read again.
- Read price. This is the "cached input" column above: from 1/4 of standard input (Grok 4.7) down to 1/50 (DeepSeek V4.1 Flash, $0.006 against $0.30).
- Write price. Anthropic charges for writing the cache: $2.50 per million tokens on Sonnet 5.5 for a 5-minute cache, against $2 standard input. Kimi K3 bills cache writes separately by lifetime (5 minutes or 1 hour), according to its pricing page. Gemini's context caching adds a storage fee per hour. OpenAI, DeepSeek and Grok list only the read price.
- Worked example. An agent sends 10M input tokens a day on Claude Sonnet 5.5. With no cache, that is 10 × $2 = $20. If 8M of those tokens are cache reads, the cost is 2 × $2 + 8 × $0.10 = $4.80, before cache-write costs.
- Hit rate is the lever. Put stable content first and changing content last, and keep timestamps out of the prefix. See How much can prompt cache save you?.
On HopBase, cache-affinity scheduling keeps a session on the same routing target for longer than the official cache lifetime, so scaling out concurrency does not reset your hit rate. The console's usage details split each request into standard input, cache write and cache read tokens.
How to check the price of every request
List prices tell you the rate. Your bill depends on usage. On HopBase:
- Every billed non-streaming response carries
X-HopBase-Cost-USDandusage.cost_usd, the exact amount deducted. - The public catalog
GET https://api.hop-base.com/api/v1/models/pricingreturns every model's official USD list price, with no key needed. See Public model catalog. - The Usage page exports a CSV where each row is official cost × discount ÷ exchange rate = charged amount. See usage reconciliation.
FAQ
Which AI model API is cheapest in October 2026?
On list price per million tokens, GPT-6 Luna and Claude Haiku 5.5 ($0.10 input / $0.50 output) and GLM-5.3 Flash ($0.15 / $0.50) are the lowest in this comparison. DeepSeek V4.1 Flash costs $0.15 / $0.60 off-peak.
Are Claude Sonnet 5.5 and GPT-6 Sol the same price?
Both list at $2 input and $10 output per million tokens. Sonnet 5.5 cache reads are $0.10 and GPT-6 Sol cache reads are $0.20. GPT-6 Sol also charges $4 / $15 once input exceeds 272K tokens.
What is a cached input price?
It is the rate for prompt tokens that are read from the provider's prompt cache instead of being processed again. It applies to repeated prefixes such as system prompts and tool definitions, and in this comparison it is 4 to 50 times cheaper than standard input.
Why are Qwen prices shown in CNY?
Alibaba's list prices for these Qwen models are published in CNY, and the HopBase pricing page shows them as published. HopBase converts CNY list prices to USD and charges your USD balance.
Do these prices include HopBase discounts?
No. Every number here is an official list price. Your own HopBase rate depends on your groups and is shown in the console. Each response also reports the exact USD charge.
How often do these prices change?
Often. Gemini 4 Argon and Gemini 3.8 Flash already have announced increases. The model changelog records list-price changes for models on HopBase, newest first.