GPT-6 Astra
−85%
Below the model's base public API rates.
- Input
- $1.50$10.00
- Output
- $7.50$50.00
USD per 1 million tokens
No long-context surcharge.
One endpoint. Leading AI models.
Ship AI products without provider friction.
Undercut the price, not the model.
An OpenAI- and Anthropic-compatible gateway for Claude Code and production apps. Choose the right model, pay only for tokens used, and track every request.
terminal
$ export ANTHROPIC_BASE_URL=https://undercut.pro
$ export ANTHROPIC_API_KEY=gw_••••••••••••
$ claude
✓ Connected through UndercutAIModel routing
The right intelligence for every request.
Switch models with one parameter. Context windows up to 1 million tokens; Grok supports up to 600,000 tokens. No new SDK or billing account.
✦Available
Claude Opus
Deep reasoning
✦Available
Claude Sonnet
Balanced & fast
✦Available
Claude Fable
Advanced reasoning
✦Available
GPT-6 Astra
Frontier intelligence
Standard UndercutAI rates
At least 55% below reference API token prices.
Our standard discounts are 55% or more off the listed base public API rates: 85% for GPT-6 Astra and 80% for Claude Fable 5.1. No subscription or limited-time offer. Compare matching token categories; the prices below are our applied rates.
Up to 1M context · Grok up to 600KCache usage billed separatelyNo model substitutionUSD per 1M tokens
Special rates for Astra and Fable 5.1
Permanent prices for short and long context.
Claude Fable 5.1
−80%
Below the model's base public API rates.
- Input
- $2.00$10.00
- Output
- $10.00$50.00
USD per 1 million tokens
No long-context surcharge.
| Model | Standard tokens | Cached tokens | Savings | ||
|---|---|---|---|---|---|
| Input | Output | Cache read | Cache write | ||
Claude Opus 5 claude-opus-5 | $5.00$2.25 | $25.00$11.25 | $0.50$0.225−55% vs the provider's cache-read rate | $6.25$2.8125Reference: 5-minute cache write | −55% |
Claude Sonnet 5 claude-sonnet-5 | $2.00$0.90 | $10.00$4.50 | $0.20$0.09−55% vs the provider's cache-read rate | $2.50$1.125Reference: 5-minute cache write | −55% |
Claude Fable 5 claude-fable-5 | $10.00$4.50 | $50.00$22.50 | $1.00$0.45−55% vs the provider's cache-read rate | $12.50$5.625Reference: 5-minute cache write | −55% |
Claude Fable 5.1Special rate claude-fable-5.1 | $10.00$2.00 | $50.00$10.00 | $0.25$0.05−80% vs the provider's cache-read rate | $12.50$2.50Reference: 5-minute cache write | −80% |
GPT-6 AstraSpecial rate gpt-6-astra | $10.00$1.50 | $50.00$7.50 | $1.00$0.15−85% vs the provider's cache-read rate | $12.50$1.875 | −85% |
GPT-5.6 Sol gpt-5.6-sol | $4.00$1.80 | $20.00$9.00 | $0.40$0.18−55% vs the provider's cache-read rate | $5.00$2.25 | −55% |
GPT-5.6 Terra gpt-5.6-terra | $2.00$0.90 | $12.00$5.40 | $0.20$0.09−55% vs the provider's cache-read rate | $2.50$1.125 | −55% |
GPT-5.6 Luna gpt-5.6-luna | $0.20$0.09 | $1.20$0.54 | $0.02$0.009−55% vs the provider's cache-read rate | $0.25$0.1125 | −55% |
Gemini 3.1 Pro gemini-3.1-pro-preview | $2.00$0.90 | $12.00$5.40 | $0.20$0.09−55% vs the provider's cache-read rate | No separate public write rate$0.90UndercutAI: at the input rate | −55% |
Gemini 3.7 Flash gemini-3.7-flash | $0.75$0.3375 | $3.75$1.6875 | $0.075$0.03375−55% vs the provider's cache-read rate | No separate public write rate$0.3375UndercutAI: at the input rate | −55% |
Gemini 3.8 Flash gemini-3.8-flash | $0.75$0.3375 | $3.75$1.6875 | $0.075$0.03375−55% vs the provider's cache-read rate | No separate public write rate$0.3375UndercutAI: at the input rate | −55% |
Grok 4.6 grok-4.6 | $2.00$0.90 | $6.00$2.70 | $0.50$0.225−55% vs the provider's cache-read rate | No separate public write rate$0.90UndercutAI: at the input rate | −55% |
Kimi K3 kimi-k3 | $3.00$1.35 | $15.00$6.75 | $0.30$0.135−55% vs the provider's cache-read rate | No separate public write rate$1.35UndercutAI: at the input rate | −55% |
Every price is USD per 1 million tokens. Struck-through values are standard public API reference rates, not prepaid credit amounts.Cache reads and writes use model-specific rates. Cached tokens are not also billed as ordinary input. Output always uses the output rate.Cache hits depend on the model, route and matching context; they are not guaranteed. Reference comparisons exclude provider surcharges for extended cache retention, storage, long context, media and tools.
Reference prices reviewed: .
The discount compares our rates with providers' base public text-token rates, excluding their long-context surcharges. This defines the price comparison, not our context limits. Claude cache writes use the 5-minute reference. UndercutAI applies the listed flat token rates without a long-context surcharge or a separate time-based cache-storage charge; managed cache retention is not offered by this table.
The reference includes OpenAI Sol promotional pricing available at least through November 21, 2026, and Gemini 3.7/3.8 Flash promotional pricing through December 31, 2026. These provider promotions are separate from UndercutAI's permanent discounts of at least 55%.
Built for developers
Familiar APIs. Full visibility.
Keep your OpenAI or Anthropic client. Change the base URL and API key, then monitor model, input tokens, output tokens, and derived cost from one dashboard.
Usage-based billing
USD balance with request-level spend history.
Keys you control
Create and revoke credentials instantly.
Transparent pricing
Configured per-million input and output rates.
Claude Code ready
Use a custom Anthropic-compatible endpoint.
Current balance
$48.27
claude-sonnet-5
12,480 tokens
$0.084
gpt-5.6-sol
7,234 tokens
$0.037
claude-opus-5
3,902 tokens
$0.126
Frequently asked questions
Questions worth asking.
Clear answers about pricing, model authenticity, verification, and what stays private to protect the service.
A practical trust policy
We do not treat a model's self-reported name as proof. Controlled, repeatable comparisons are more meaningful than screenshots or unverifiable claims.
01What is UndercutAI?
UndercutAI is a prepaid AI API gateway. One key and one balance give you access to supported frontier models through familiar OpenAI- and Anthropic-compatible endpoints. You choose the model in every request and see token-level usage and cost in one dashboard.
02Why is it cheaper than buying from each provider directly?
UndercutAI sets its own retail rates. Our standard discounts start at 55% below the listed public reference rates and vary by model; they do not describe our upstream costs. Cache hits reduce the cost of repeated input separately, not the cost of output. A lower price does not permit us to substitute the model you request.
03Is my balance the same as credit at a model provider?
No. Your USD balance pays for usage within UndercutAI at our published rates; it is not transferred to an Anthropic or OpenAI account. Each model's discount is already included in token prices, not added as a credit multiplier on top-up. The dashboard shows the amount credited and the available payment methods; currency conversion and payment fees may affect what you pay.
04Are these discounts a limited-time offer?
No. Discounts of at least 55% are our permanent standard pricing policy relative to the listed base public API rates, with no scheduled expiry. GPT-6 Astra has an 85% discount and Claude Fable 5.1 has an 80% discount. Dollar token prices are not fixed forever: they may change when providers change their reference rates. The pricing table shows the rates currently applied.
05Are requested models replaced with cheaper models?
No. UndercutAI does not intentionally substitute a requested model with an unrelated lower-cost model. That is the whole point of "Undercut the price, not the model." — the model identifier in your request is the product you are buying. Routing details may remain confidential, but a lower price is not permission to silently downgrade the model.
06How can I verify that a model is genuine?
A model's self-description and the model field in an API response are not independent proof of its identity. Controlled side-by-side comparisons can help assess consistency: use the same prompts and supported settings through UndercutAI and the official API, then compare capabilities, tool calls, long-context recall and token usage over several runs. These comparisons are useful evidence, not cryptographic proof.
07Can I just ask the model what it is?
No. Self-identification is weak evidence: a system prompt can change the answer, and different models can imitate the same response. Capability tests with fixed inputs, low temperature, repeat runs, and an official baseline are much harder to fake and produce a more useful comparison.
08Why don't you publish upstream accounts or raw provider headers?
Publishing credentials, account identifiers, raw upstream headers, or routing topology would make the infrastructure easy to abuse or copy and could reduce availability for every customer. We keep those operational details private while exposing the information customers need to audit their own usage: requested model, token categories, applied rate, cost, and history.
09Why might two answers differ from the official console?
Model output is probabilistic. Temperature, system instructions, tool definitions, context, provider-side model updates, and even repeated runs can change an answer. For a fair comparison, keep every input and parameter identical, run several samples, and judge capability and consistency rather than exact wording.
10What do I need to change in my application?
Usually only the base URL and API key. Keep your existing OpenAI or Anthropic client, select a supported model, and monitor usage from the dashboard. The quickstart page includes ready-to-copy examples for popular SDKs and developer tools.
Standard rates · save at least 55%
Your next model is one API call away.
Create an account, add funds, and issue your first key in minutes.
Create free account