AI Stacks · Pricing checked November 21, 2026

LLM API Pricing Compared: Claude vs GPT vs Gemini vs Grok vs DeepSeek

Input and output token prices for the five main model APIs, copied from their own pricing pages on November 21, 2026, and the times model costs and limits bit us on the show.

Watch: Episode 22, Let's do the Math! from 03:08. Jordan's LLM Math Roaster calls Gemini, GPT, Claude and Grok through their APIs side by side and has a judge model score them.
Subscribe on YouTubeWe build something real with AI every week

AI price-change tracker and build-cost estimator: every dated price change from the vendors' own pages, starting with Gemini 4 Carbon doubling on December 1, 2026.

The short answer

Flagship models from Anthropic and OpenAI cost about the same: $8 per million input tokens and $40 per million output for Claude Fable 5.2 and gpt-6.2-astra. The mid-tier models are where most apps should live: Claude Sonnet 5.5 and gpt-6.2-sol at $2 / $10, Grok 4.7 at $2 / $6. At the cheap end, Claude Haiku 5.5 and gpt-6-luna are $0.10 / $0.50, and DeepSeek halves its prices off-peak. Every vendor except DeepSeek offers a batch discount: 50% at Anthropic, OpenAI and Google, 20% at xAI.

In our builds, the bill was driven less by the price per token than by how many tokens a setting quietly used (see below).

Pricing and limits (checked November 21, 2026)

Prices are US dollars per million tokens, input / output, at standard (non-batch) rates unless noted. Long-context surcharges apply above the thresholds shown.

Anthropic Claude API

Plan / modelPriceIncludedDocumented limitsOverage / pay-as-you-go
Claude Fable 5.2$8 in / $40 out per 1M tokensCache hits $0.20 per 1M—Batch API 50% off
Claude Opus 5.5$4 in / $20 outCache hits $0.20 per 1M—Batch 50% off
Claude Sonnet 5.5$2 in / $10 outCache hits $0.10 per 1M—Batch 50% off
Claude Haiku 5.5$0.10 in / $0.50 out (prompts up to 100K tokens)Cache hits $0.01 per 1M$0.50 / $2.50 above 100K tokensBatch 50% off

Source: Anthropic Claude API pricing, checked November 21, 2026.

OpenAI API

Plan / modelPriceIncludedDocumented limitsOverage / pay-as-you-go
gpt-6.2-astra$8 in / $40 out per 1M (up to 272K context)Cached input $0.40$16 / $60 above 272KBatch $4 / $20
gpt-6.2-sol$2 in / $10 outCached input $0.10—Batch $1 / $5
gpt-6.1-terra$1 in / $5 outCached input $0.05—Batch $0.50 / $2.50
gpt-6-luna$0.10 in / $0.50 outCached input $0.01—Batch $0.05 / $0.25

Source: OpenAI API pricing, checked November 21, 2026.

Google Gemini API

Plan / modelPriceIncludedDocumented limitsOverage / pay-as-you-go
gemini-4-carbon$2 in / $10 out per 1M through Nov 30, 2026; $4 / $20 from Dec 1, 2026Context caching $0.10 (then $0.20)Not on free tierBatch half price
gemini-3.1-pro-preview$2 in / $12 out (prompts up to 200K)Caching $0.20$4 / $18 above 200K; no free tierBatch $1 / $6
gemini-3.5-flash-lite$0.30 in / $2.50 outContext caching $0.03Free tier availableBatch $0.15 / $1.25

Source: Google Gemini API pricing, checked November 21, 2026.

xAI Grok API

Plan / modelPriceIncludedDocumented limitsOverage / pay-as-you-go
grok-4.7$2 in / $6 out per 1M (under 200K prompt)Cached input $0.50; 500K context$4 / $12 at 200K+ prompt tokensBatch 20% off
grok-4.3$1.25 in / $2.50 outCached $0.20; 1M context$2.50 / $5 at 200K+Batch 20% off
grok-build-0.1$1 in / $2 outCached $0.20; 500K context$2 / $4 at 200K+No batch discount

Source: xAI Grok API pricing, checked November 21, 2026.

DeepSeek API

Plan / modelPriceIncludedDocumented limits
deepseek-v4-pro$1.32 in / $3.96 out per 1M (peak)Off-peak $0.66 / $1.98; 1M contextConcurrency limit 500
deepseek-flash$0.30 in / $1.20 out (peak)Off-peak $0.15 / $0.60; 1M contextConcurrency limit 2,500

Source: DeepSeek API pricing, checked November 21, 2026.

What broke when we built with it

Anthropic: great results, runaway spend on high effort

Jordan's first 36 hours with Anthropic's top model went like this:

First, I put on like the highest. And then that starts like sub agents and a bunch of other stuff. My spend went really high, so I had to like kinda turn it down a little bit.

— Jordan Metzner, Episode 46, at 05:09

I ran it once like that and like my they like immediately told me like I ran out of tokens.

— Jordan Metzner, Episode 46, at 07:05

His request to the model makers: warn before an expensive run, along the lines of "that's gonna cost you like a $100... could do this for $5 instead" (at 07:25). Earlier he'd put it more simply: "these anthropic models are so expensive" (at 32:48), and when prices dropped, "it might be getting cheaper, but we're spending more money using it" (at 17:38).

Open models: not enough context for real documents

For the solopreneur finance dashboard, Jordan started on an open model running on Groq (most likely gpt-oss-120b) and had to switch:

there was a lot of data inside these p and l's and balance sheets, and so it wasn't enough. I needed a bigger model, and so in fact, I moved to Gemini

— Jordan Metzner, Episode 14, at 06:59

Gemini: new models you can't get yet, and throttling

Building the Math Roaster, Jordan wanted the newest models in the lineup, but "these models just came out like yesterday, and I don't have API access yet" (at 03:08). That same week he hit rate limits everywhere he tried Gemini 3: "I'm being throttled everywhere" (at 18:13). Even in Google AI Studio, which he loves for quick apps, paid use means "you might have to set it up with, an API key... It's a little complicated" (at 11:12).

OpenAI: routing that picked the wrong mode

In GPT-5's first week Sam kept getting the fast answer when he wanted the reasoning one:

there's been times when I've gotten the the quick answer when I want the thinking answer, and it's just I've had to, you know, re prompt to think deeply

— Sam Nadler, Episode 8, at 23:27

For our auto-blogging build, Jordan kept a human review step because of "a little bit of risk of, you know, [ChatGPT] hallucinating" off-brand content (at 12:01).

Grok and DeepSeek

Jordan's take on Grok for coding was "It's okay. It's not better than anything else" (at 21:34), and Ryz Labs doesn't use it in production (at 24:02). We have no DeepSeek API build on the show, so it appears in the price table only.

What our guests reported

A healthcare guest calling Claude through AWS said "our last bill was about six x" what they expected, because of a change in how long their Lambda held the Anthropic connection open (at 14:33). Another guest said Gemini 3 is "so much cheaper to run compared to, like, GPT" (at 18:34).

How to choose

  • High-volume classification, extraction, routing: Haiku 5.5, gpt-6-luna or deepseek-flash, with batch or off-peak pricing where you can wait.
  • Most product features: Sonnet 5.5, gpt-6.2-sol or Grok 4.7 ($2 in). Test quality on your own prompts; price differences here are small.
  • Hard reasoning and coding: Fable 5.2 or gpt-6.2-astra. Cap effort settings and sub-agents, because that's where our spend jumped.
  • Price changes: check the source links before you commit. gemini-4-carbon's input price doubles on December 1, 2026.
  • Many models, one bill: see LLM gateways and routing.

FAQ

What is the cheapest LLM API?

Among the models we priced on November 21, 2026, the cheapest per token were Claude Haiku 5.5 and OpenAI gpt-6-luna at $0.10 per million input tokens and $0.50 per million output, then DeepSeek's deepseek-flash at $0.30 / $1.20 (half that off-peak) and Gemini 3.5 Flash-Lite at $0.30 / $2.50, which also has a free tier.

How much does the Claude API cost?

Anthropic's pricing page lists Claude Fable 5.2 at $8 per million input tokens and $40 per million output, Opus 5.5 at $4 / $20, Sonnet 5.5 at $2 / $10 and Haiku 5.5 at $0.10 / $0.50 for prompts up to 100,000 tokens. The Batch API is 50% off.

How much does the OpenAI API cost?

OpenAI lists gpt-6.2-astra at $8 / $40 per million tokens (up to 272K context), gpt-6.2-sol at $2 / $10, gpt-6.1-terra at $1 / $5 and gpt-6-luna at $0.10 / $0.50. Batch is half price.

Is the Gemini API price going up?

Yes for gemini-4-carbon. Google's pricing page shows $2 per million input tokens and $10 output through November 30, 2026, then $4 and $20 from December 1, 2026. Gemini 3.5 Flash-Lite stays at $0.30 / $2.50 and has a free tier.

Which LLM API do you use in production?

At Ryz Labs, Jordan said production runs mainly on Anthropic, OpenAI and Gemini models, and that they haven't found a high-quality production use case for Grok yet. Internal tools use API keys as a pay-for-what-you-use layer instead of per-seat subscriptions.

Sources