Most LLM API pricing pages show you one vendor’s rates. Comparing three means opening three tabs and doing arithmetic, and the arithmetic is where the surprises are, because the headline rate is not what you pay.
This page has current LLM API pricing for all three providers, and then what those rates come to on three real workloads.
Current LLM API pricing
Published price per one million tokens, taken from each provider’s official pricing page in September 2026.
| Provider | Model | Input | Output |
|---|---|---|---|
| OpenAI | gpt-6-astra | $10.00 | $50.00 |
| OpenAI | gpt-5.6-sol | $4.00 | $20.00 |
| OpenAI | gpt-5.4 | $2.50 | $15.00 |
| OpenAI | gpt-5.6-terra | $2.00 | $12.00 |
| Anthropic | Claude Opus 5 | $5.00 | $25.00 |
| Anthropic | Claude Sonnet 5 | $2.00 | $10.00 |
| Anthropic | Claude Haiku 4.5 | $1.00 | $5.00 |
| Gemini 3.5 Flash | $1.50 | $9.00 | |
| Gemini 3.8 Flash | $0.75 | $3.75 | |
| Gemini 2.5 Pro | $1.25 | $10.00 |
Verify before you budget. Several of these carry promotional rates with published expiry dates, and this market reprices often. When we checked, third-party pricing aggregators disagreed with the official pages on multiple models. Use the vendor’s own page as the source.
What that costs on real work
Rates per million tokens tell you very little until you multiply them by your own volume. Here are three workloads, costed across every model above.
The token estimates are deliberately conservative and stated openly so you can substitute your own.
A support assistant
10,000 conversations a month. Roughly 1,500 input tokens each, covering the system prompt and whatever context gets retrieved, and 300 output tokens. That comes to 15M input and 3M output tokens a month.
| Model | Monthly cost |
|---|---|
| Gemini 3.8 Flash | $22.50 |
| Claude Haiku 4.5 | $30.00 |
| Gemini 2.5 Pro | $48.75 |
| Gemini 3.5 Flash | $49.50 |
| Claude Sonnet 5 | $60.00 |
| gpt-5.6-terra | $66.00 |
| gpt-5.4 | $82.50 |
| gpt-5.6-sol | $120.00 |
| Claude Opus 5 | $150.00 |
| gpt-6-astra | $300.00 |
Document analysis
500 documents a month at roughly 80,000 input tokens each, which is a long contract or a dense report, producing a 2,000 token summary. That is 40M input and 1M output.
| Model | Monthly cost |
|---|---|
| Gemini 3.8 Flash | $33.75 |
| Claude Haiku 4.5 | $45.00 |
| Gemini 2.5 Pro | $60.00 |
| Gemini 3.5 Flash | $69.00 |
| Claude Sonnet 5 | $90.00 |
| gpt-5.6-terra | $92.00 |
| gpt-5.4 | $115.00 |
| gpt-5.6-sol | $180.00 |
| Claude Opus 5 | $225.00 |
| gpt-6-astra | $450.00 |
Content generation
2,000 pieces a month, 1,000 tokens of brief and context in, 1,500 tokens out. That is 2M input and 3M output.
| Model | Monthly cost |
|---|---|
| Gemini 3.8 Flash | $12.75 |
| Claude Haiku 4.5 | $17.00 |
| Gemini 3.5 Flash | $30.00 |
| Gemini 2.5 Pro | $32.50 |
| Claude Sonnet 5 | $34.00 |
| gpt-5.6-terra | $40.00 |
| gpt-5.4 | $50.00 |
| gpt-5.6-sol | $68.00 |
| Claude Opus 5 | $85.00 |
| gpt-6-astra | $170.00 |
Three things the tables show
Tier choice. Cheapest to most expensive is a factor of about thirteen on every workload here, whether you are reading long documents or writing short ones. Choosing a tier badly costs more than choosing a provider badly.
Output length. Output tokens cost three to five times input tokens at every provider without exception. A system that writes long answers costs far more than one that reads long documents. Capping output length is usually the fastest saving available, and it is a one-line change.
Absolute scale. The most expensive option in the support scenario is $300 a month. For most businesses the model bill is not the expensive part of an AI project, and teams that spend weeks optimising it are optimising the wrong line. The engineering time to connect the model to your business costs more, and that is covered further down.
Where each provider is genuinely different
Three things separate them in practice. Everything else is close enough that it rarely decides a project.
Context window. How much text the model can consider at once, which binds on long contracts, large codebases, or a year of support tickets. Claude’s current lineup runs to 1M tokens. OpenAI’s GPT-5.2 runs to 400,000. Gemini’s larger models also reach 1M, but Gemini prices differently past 200,000 tokens, which changes the arithmetic on long-document work specifically.
Price per token. Covered above. The tier gap inside one provider is wider than the gap between providers.
Ecosystem fit. If your company runs on Google Workspace, Gemini reaches your documents natively. If your engineers already build against OpenAI’s libraries, that familiarity is worth real delivery time.
Raw capability on general business tasks has narrowed to the point where it rarely decides anything. All three write competent prose, handle structured data, and call external tools reliably.
For a fuller comparison of how the three providers differ on capability and fit rather than price, see OpenAI vs Claude vs Gemini.
Choosing by workload
Customer support. Volume is high and each interaction is short, so cost per interaction dominates. Start on a cheap tier. The quality gap on routine replies is smaller than teams expect and the cost gap is not. Escalate to a larger model only for cases the small one fails.
Document analysis. Context window is the binding constraint, not price. Test with your longest realistic document before committing, and check how your provider prices beyond its base context tier.
Coding assistance. All three are competent. The deciding factor is which one integrates with your existing editor and CI setup.
Business planning and analysis. Reasoning quality matters and volume is low, so frontier pricing is justified here. It is also where you should test on your own data rather than trust any published comparison, this one included.
Content generation at scale. Output tokens dominate the bill, as the third table shows. Model the cost at your real volume before choosing.
The cost most pricing pages leave out
Choosing a model takes an afternoon. Making it useful takes considerably longer, and that is where AI projects fail.
A model on its own knows nothing about your business. It has not seen your customer records, your pricing rules, your inventory, or your internal documentation. Connecting it to those is the work, and it is where the real budget goes.
Retrieval. Your business data has to reach the model at the moment it answers, which means indexing your documents and records so the relevant pieces get pulled into each request. Retrieval quality determines answer quality more than model choice does.
System integration. Reading data is half of it. An AI system that can look up an order but not update one is a search box. Useful systems write back into your CRM, ticketing system or database, which brings authentication, permissions and error handling.
Guardrails. A model that answers every question confidently, including ones it should decline, is a liability in customer-facing use. Production systems need boundaries on what the model may say and do, and a defined handoff to a person.
Evaluation. Without a way to measure whether outputs are correct, you cannot tell whether a change improved anything. Teams that skip this tune prompts by feel and cannot say whether the system is getting better.
Cost controls. Caching repeated context, routing simple requests to smaller models, and capping output length are the standard levers, and they matter more than the rate you negotiated.
This is why provider choice matters less than it appears. The retrieval layer, the integrations, the guardrails and the evaluation harness are the expensive parts, and they nearly all survive a change of model. It is also why comparing LLM API pricing alone will not tell you what a project costs, and why our LLM development work starts with the connective layer rather than the model.
Working out your own number
Published comparisons cannot tell you which model suits your work, because they were not tested on your work. Run your own check. It takes about a week.
- Collect 20 to 50 real examples of the task you want automated. Real inputs, not invented ones.
- Write down what a good answer looks like for each, before testing anything. Doing it afterwards means grading against whatever the model produced.
- Run all three providers on the same set, at the cheap tier first.
- Score against your criteria. Where the cheap tier fails, retest that subset on a larger model.
- Multiply by your real monthly volume, using the method in the tables above.
The usual result is that several models are good enough and the decision comes down to cost, integration effort, or a vendor relationship you already have. That is a far easier decision than the one you started with.
Where this leaves you
LLM API pricing has no single winner in 2026, and any page claiming otherwise is guessing on your behalf. The three providers are close on general capability, meaningfully different on context window and price, and largely interchangeable once your integration layer exists.
Pick on the constraint that binds you, whether that is document length, cost per interaction, or the systems you already run. Then spend your effort on the connective work, because that is what decides whether the project delivers anything.
Ethersofts builds the integration layer. Our AI integration work connects language models from any provider to the systems a business already runs: retrieval over your own data, writes back into your CRM or ERP, guardrails for customer-facing use, and evaluation so you can tell whether it works.
If you are deciding between providers, or you have chosen one and need it connected to your business, talk to an engineer.