LLM API Pricing Comparison 2026: Every Major Model, Priced

LLM API Pricing Comparison 2026: Every Major Model, Priced

An LLM API pricing comparison sounds simple until you try to build one: prices moved twice in the first half of 2026, every provider prices input and output differently, and the "headline" number rarely matches your actual bill. So we pulled every figure below straight from the official pricing page on July 14, 2026, and did the math for a real workload at the end.

The Complete LLM API Pricing Table (2026)

Here's the whole field on one screen, priced per 1 million tokens in USD. This is the table people screenshot, so it's first.

Two footnotes that matter. Claude Sonnet 5 is on introductory pricing of $2.00 input / $10.00 output per million through August 31, 2026, then it reverts to the $3.00 / $15.00 shown above. And Gemini 2.5 Pro jumps to $2.50 input / $15.00 output once a single prompt crosses 200K tokens, so its "list" price is really a floor.

Every model here also offers a Batch API at a flat 50% discount on both input and output (Anthropic, OpenAI, Google, and Alibaba), which we'll fold into the worked example below.

What You're Actually Paying For

Three numbers drive your bill, and most pricing pages bury two of them.

• Input tokens are everything you send: system prompt, conversation history, retrieved context, the user's question. On chat and RAG apps this is usually the bigger half.
• Output tokens are what the model generates, and they cost 3-6x more than input on almost every provider. GPT-5.6 Terra charges $2.50 in and $15.00 out, a 6x multiplier. Verbose responses hurt.
• Cached input is the sleeper. If you send the same system prompt on every request, prompt caching bills those repeated tokens at roughly 10% of the normal rate. Look at the "Cached Input" column above: Claude Opus 4.8 drops from $5.00 to $0.50. On a high-traffic app, that one column decides whether your bill is $10K or $3K. Our prompt caching guide covers the setup for each provider.

The takeaway: a raw "input price" comparison is misleading. The model with the lowest sticker can lose to a slightly pricier one that caches better, and the model with the scariest output price can be cheapest if your task returns two-word answers.

The Cheapest Models for High-Volume Work

If you're processing millions of tokens on tasks like classification, extraction, summarization, or moderation, the bottom of the table is where the money is.

• DeepSeek-V4 is the price floor at $0.14 input / $0.28 output. Its cache-hit input rate of roughly $0.003 per million is almost free. On a 1M-token context with 384K max output, it comfortably handles most non-frontier work.
• Gemini 2.5 Flash-Lite at $0.10 / $0.40 is Google's answer, with the same 1M context and a mature ecosystem.
• Zhipu GLM-4.6 at $0.43 / $1.74 sits in the middle and is a strong coding model. If you're using it heavily for development, GLM's coding-plan subscription can beat per-token pricing outright.
• Mistral Small 4 at $0.15 / $0.60 is the cheapest option with EU data residency, which matters for regulated workloads.

The gap between this tier and the flagships is not subtle. DeepSeek-V4's output price is over 50x cheaper than GPT-5.6 Sol's. For any task where a mid-tier open-weights model passes your quality bar, that spread is the single biggest lever on your bill.

The Flagship Tier, Compared

When the task genuinely needs frontier reasoning (complex agents, hard code, deep analysis), you're choosing between four models. Here's how they line up on price for the same intelligence class:

On paper Gemini 2.5 Pro is the bargain of the group, roughly a quarter of Sol's output price. The catch is its long-context penalty: cross 200K tokens in a single prompt and it reprices the whole request at $2.50 / $15.00. For agents that stuff large contexts, that erases much of the advantage, so price your actual prompt sizes before deciding.

Claude Fable 5 is the outlier on cost. It's positioned above the Opus tier for the most demanding long-horizon reasoning, so it's a deliberate "reach for it when correctness beats cost" model, not a default.

The Hidden Costs Nobody Puts in the Table

A per-token comparison misses four things that move real bills:

1. Long-context tiers. Gemini 2.5 Pro (above 200K) and GPT-5.6 (above roughly 272K) charge 1.5-2x once prompts get large. If you do long-document work, your effective rate is the tiered one.
2. Cache write premiums. Anthropic charges 1.25x to write the cache (2x for the 1-hour TTL) before you get the 0.1x read rate. It pays off after two requests, but a low-repetition workload can pay the write premium and never cash in.
3. Batch vs real-time. The 50% batch discount is free money if your workload can wait 24 hours. Most teams have more batch-eligible traffic than they think (nightly jobs, backfills, moderation).
4. The provider that isn't on the list. For very high volume, self-hosting an open model on rented GPUs can undercut every API rate here. Our guide to running LLMs locally covers when that math flips.

Price Per Task, Not Per Token: A Real Job

Per-token prices are abstract. So we ran a real one. In June 2026 we pushed a 50-document summarization batch (about 1.2M input tokens and 120K output tokens) through five models and logged the actual bill:

Same 50 documents, same prompt. The bill ranged from $4.80 to $0.20, a 24x spread on identical work, before batching. Once we moved the batch-eligible providers to their overnight queue, Gemini 2.5 Flash landed at 33 cents. For summarization, where the quality difference between these models was inside our margin of error, that decision is worth thousands of dollars a month at scale.

The lesson: the right comparison is always cost per task, computed on your real input/output ratio, not the sticker price of a single token. A model with pricey output is cheap for extraction; a model with cheap input is pricey for long generation.

Which Model Is Cheapest for Your Use Case?

Short version, priced from the table above:

• Chatbots and RAG (long input, short output): input price and caching dominate. Gemini 2.5 Flash and DeepSeek-V4 win; add prompt caching on whichever you pick.
• Long-form generation (short input, long output): output price dominates. Gemini 2.5 Flash-Lite and DeepSeek-V4 are cheapest; avoid GPT-5.6 Sol unless you need the reasoning.
• Agents and tool use (large context, many turns): watch the long-context tiers. Claude Opus 4.8 and GPT-5.6 Terra are predictable; Gemini 2.5 Pro is cheapest only if you stay under 200K.
• Coding: GLM-4.6 and its coding-plan subscription, or Claude Opus 4.8 for the hardest problems.
• Frontier reasoning where correctness beats cost: Claude Opus 4.8 or Claude Fable 5.

Whatever you land on, route it through a gateway so switching providers is a config change, not a rewrite. Our best LLM gateway tools comparison covers the options, and if you want the full playbook for driving the bill down, see our guide on reducing LLM API costs. For a Gemini-only deep dive with the multimodal and audio rates this table doesn't cover, see our Gemini API pricing breakdown.

How Techsy Chooses Models for Clients

We build production AI systems for B2B clients, and model selection is one of the first decisions we make, usually before a line of code. Our approach is boring on purpose: we profile the real input/output ratio of the workload, price the top three candidates on that ratio (not the sticker), run them against an eval set to confirm quality parity, then wire the cheapest passing model behind a gateway so we can re-price it every quarter as new models launch. That last step matters more than picking the "best" model today, because the price you see above will be wrong in three months.

If you're choosing a model or staring at a bill that keeps climbing, that's the kind of decision we help with. Explore our AI integration services or book a free architecture review.

Frequently Asked Questions

What is the cheapest LLM API in 2026?

DeepSeek-V4 is the cheapest capable model at $0.14 input / $0.28 output per million tokens as of July 2026, with cache-hit input near $0.003. Gemini 2.5 Flash-Lite ($0.10 / $0.40) and Mistral Small 4 ($0.15 / $0.60) are close behind. For frontier-quality work, Gemini 2.5 Pro is the cheapest flagship.

Is GPT-5.6 or Claude Opus 4.8 more expensive?

They match on input ($5.00/M) but Claude Opus 4.8 is cheaper on output ($25.00/M vs GPT-5.6 Sol's $30.00/M). For output-heavy workloads, Opus 4.8 wins on price; for input-heavy ones, they're effectively tied before caching.

Why is output more expensive than input?

Generating tokens is more compute-intensive than reading them, so nearly every provider charges 3-6x more for output. This is why trimming response length (and using structured outputs) saves more per token than trimming your prompt.

How much does prompt caching actually save?

Cached input tokens bill at roughly 10% of the standard rate across Anthropic, OpenAI, and Google, a 90% discount on the repeated portion of your prompt. For apps that send the same system prompt on every request, caching often cuts the total bill 30-50%.

Do these prices include the Batch API discount?

No. The main table shows standard real-time pricing. Anthropic, OpenAI, Google, and Alibaba all offer a Batch API at a flat 50% discount on both input and output for non-urgent workloads that can tolerate up to 24 hours of latency.

Is DeepSeek really that much cheaper than US models?

Yes, on paper. DeepSeek-V4's output price ($0.28/M) is over 50x cheaper than GPT-5.6 Sol's ($30/M). The tradeoff is that frontier US models still lead on the hardest reasoning tasks, so most teams arbitrage the tasks where quality parity holds and keep a flagship for the rest.

What's the catch with Gemini 2.5 Pro's low price?

Its long-context tier. Gemini 2.5 Pro lists at $1.25 / $10.00, but any single prompt over 200K tokens reprices the entire request at $2.50 / $15.00. If your prompts are large, budget the tiered rate, not the headline.

How often does LLM API pricing change?

Frequently. Prices shifted twice in the first half of 2026, and new model families (like the GPT-5.6 tier that launched July 9, 2026) reset the field each time. Always confirm against the provider's official pricing page before committing a workload, and re-price quarterly.

Which model has the largest context window for the price?

DeepSeek-V4 and the Gemini 2.5 family offer a 1M-token context at the low end of the price range. GPT-5.6 models edge slightly higher at 1.05M, and GPT-5.4 nano reaches 1.1M at just $0.20 input, making it the cheapest large-context option for simple tasks.

Sources

• Anthropic Claude API Pricing (accessed 2026-07-14)
• OpenAI API Pricing (accessed 2026-07-14)
• Google Gemini API Pricing (accessed 2026-07-14)
• DeepSeek API Pricing (accessed 2026-07-14)
• Alibaba Cloud Model Studio Pricing (accessed 2026-07-14)
• Mistral API Pricing (accessed 2026-07-14)
• Z.AI (Zhipu GLM) Pricing (accessed 2026-07-14)