LLM API Pricing Comparison: Claude, Gemini, Grok, and OpenAI

This is an LLM API pricing comparison of the list prices I could read on the vendors' own pages on October 12, 2026: Claude API pricing, Gemini API pricing, Grok API pricing, and the OpenAI API price table. The number that matters is not the flagship's input rate. It is what one real reply costs, then what 10,000 of those replies cost in a month.

How to read a per-token price

Every vendor bills a base input price and a higher output price, per million tokens. Cache hits are cheaper than fresh input. Long prompts cross a threshold and the whole request reprices, not just the extra tokens. Batch and regional endpoints change the rate again. If a comparison table hides those, it is advertising.

The figures below are standard (not batch, not priority) list prices for the short-context or global tier, in US dollars per million tokens. I did not send paid API calls to confirm an invoice. Where a page lists a cached-input rate, I mention it in the notes and leave it out of the worked example, so the example is the pessimistic case: nothing cached.

Sources, all read on October 12, 2026:

• Claude API pricing
• Gemini API pricing
• Grok API pricing on the xAI docs
• OpenAI API pricing

Claude API pricing

Anthropic's pricing page lists standard global rates. These are the three current models I would actually put in a budget. Older Opus and Sonnet rows are still on the page; I am not repeating the retired ones.

Two footnotes from the same page that change a real invoice. For Claude 4.6 and later, inference_geo: "us" is a 1.1x multiplier on input, output, and cache. And Claude 4.7 and later models use a newer tokenizer that the page says produces about 30% more tokens for the same text, depending on the workload. A lower sticker price can still cost more if the tokenizer is chattier. I did not measure that 30% on a sample of my own prompts.

The Batch API is a 50% discount on input and output for these models (Haiku 5.5 batch is $0.05 / $0.25 under the 100k threshold, Sonnet 5.5 batch is $1 / $5, Opus 5.5 batch is $2 / $10). Batch is for work nobody is waiting on.

Gemini API pricing

Google's Gemini Developer API page prices the paid standard tier per million tokens. Output prices include thinking tokens. That matters: a model that "thinks" for 2,000 tokens and then writes 500 is billed for 2,500 on the output side, not 500. The worked example later assumes a 500-token output and says so.

The paid tier also advertises a Batch API at a 50% cost reduction. I used the standard column for the comparison so batch does not make Gemini look cheaper than a synchronous Claude or Grok call. Free tier exists for some of these models and is not a production plan: the page says free-tier content can be used to improve Google's products.

I am not listing image, live-audio, or transcribe SKUs here. They are on the same page and they are priced differently, often per minute. If your product is voice, read those rows yourself. The voice agent guide is about latency, not these token rates.

Grok API pricing

xAI's docs price Grok per million tokens, with a long-context rate once the prompt reaches 200,000 tokens. The higher rate applies to all tokens in that request. grok-4.7 is the flagship on the page. Cached input is a separate column, and I kept it out of the worked example.

The US regional endpoint (https://us.api.x.ai/v1) is 1.1x the global token rates. For grok-4.7 under 200k that is $2.20 input, $0.55 cached, $6.60 output. The page says that endpoint currently serves grok-4.7 and grok-4.6 only.

Batch is not a blanket discount. The docs list a 20% batch discount for grok-4.3 and the grok-4.20 variants, and say models not on that list have no batch discount. grok-4.7 is not on the list. Priority processing is 2x standard rates, and only when the response confirms service_tier: "priority". I did not find a batch price for grok-4.7 because the page says there isn't one.

OpenAI, then the comparison

OpenAI's API pricing page is a long table. I pulled the current cheap tier and the current general-purpose tiers, standard short-context column only. Short context on that page means 272,000 input tokens or fewer. Above that, models that have a long-context column reprice the whole session.

The same page says GPT-5.6 Sol's promotional pricing is available at least through November 21, 2026. Treat that row as provisional. It also says regional processing is a 10% uplift for models released on or after March 5, 2026. I did not verify each model's release date against that cutoff, so I am not applying the 10% in the table.

Put next to each other, standard short-prompt rates look like this. Output is what dominates once a reply is longer than a sentence, because every vendor charges several times more for it.

Claude Sonnet 5.5 and gpt-6.1-sol are the same sticker: $2 and $10. grok-4.7 matches them on input and is cheaper on output ($6 vs $10). Gemini 3.8 Flash is cheaper than those flagships on both sides at the 2026 introductory rate, and Gemini 3.1 Pro Preview matches the $2 input but charges $12 out. None of that tells you which one answers your tickets. It tells you what you pay if they use the same number of tokens. They will not. Tokenizers differ, and thinking tokens show up on the Gemini output line.

What one conversation costs

Here is the worked example I use when a founder asks "what will this bot cost?" One support reply sends 4,000 input tokens (system prompt, a few prior turns, the new message) and returns 500 output tokens. No cache. No batch. Prompt is under every long-context threshold above. Gemini's published output price includes thinking tokens, so if the model thinks, add those tokens to the 500. This example does not.

Cost = (4,000 / 1,000,000) × input price + (500 / 1,000,000) × output price.

Read the last column as a floor, not a quote. A real thread is several replies. The input grows as history grows, unless you trim it. Tool calls add more round trips. An agent that loops can spend the monthly number on a single bad ticket. The nine fixes for that, in the order I apply them, are in how to reduce LLM costs. Routing the easy replies to the cheap column and the hard ones to the expensive column is the biggest lever, and it is written up separately as LLM routing.

On a production system the gap is larger than this table. In the cost-optimization case study, semantic caching, prompt compression, and model routing cut a SaaS platform's OpenAI spend by 65% with no drop in CSAT. The playbook with the before-and-after for a support agent is the 60% LLM cost cut. Those are measured bills, not list-price arithmetic. If you want that pass run on your own traffic, it is the LLM cost optimization engagement. Build and run-rate ranges for the agent itself are in what it costs to build an AI agent.

What I could not price from the pages

• I did not confirm invoices. A list price and a line item can differ once cache writes, thinking tokens, and failed retries land.
• I did not measure tokenizer differences. Anthropic's own note of roughly 30% more tokens on Claude 4.7 and later is a warning, not a number I reproduced.
• Gemini 3.1 Pro is still labeled Preview. The page can change when it leaves preview.
• I did not apply OpenAI's 10% regional uplift, because I did not check each model's release date against the March 5, 2026 cutoff on that page.
• Image, video, voice, and tool-call fees (xAI web search is $5 per 1,000 calls on the same docs page) are outside this token table.
• Anthropic's page names models ("Claude Sonnet 5.5") rather than the API slug. I did not guess slugs.

Frequently Asked Questions

Conclusion

Claude API pricing, Gemini API pricing, and Grok API pricing are close enough on the flagship input rate ($2 per million for Sonnet 5.5, Gemini 3.1 Pro Preview, grok-4.7, and gpt-6.1-sol) that the output rate and the tokenizer decide the bill. The cheap tier is where the money is if most of your traffic is easy. Measure tokens per route, then route. The list prices above are a snapshot of October 12, 2026, from the four official pages linked in the first section.

If the bill is already larger than this arithmetic and you want it taken apart against your own logs, that is what the LLM cost optimization work is for. The concepts behind the agent those calls sit inside are in what AI agents are.