Gemini API vs OpenAI vs Claude: Complete 2026 Cost Decision Guide

As of October 8, 2026, Google Gemini is no longer the cheapest API at every tier. At the budget end, Gemini 2.5 Flash-Lite, OpenAI's GPT-6 Luna and Anthropic's Claude Haiku 5.5 all charge $0.10 per million input tokens, and Flash-Lite only wins on output ($0.40 vs $0.50). In the $2-input band, GPT-6.1 Sol and Claude Sonnet 5.5 both list $2 / $10, while Gemini 3.1 Pro Preview lists $2 / $12. Gemini's clear price lead is in the middle: Gemini 3.8 Flash costs $0.75 / $3.75 through December 31, 2026, then doubles on January 1, 2027. Flash-Lite's budget slot comes with a condition: Google now limits the 2.5 models to users who have actively used them before and points new projects to 3.5 Flash-Lite or 3.8 Flash (Gemini deprecations page).

For a typical workload of 10,000 requests a month at 1,500 input and 500 output tokens each, that works out to $3.50 on 2.5 Flash-Lite (if your project has 2.5 access), $4 on Luna or Haiku 5.5, $80 on Sol or Sonnet 5.5, and $400 on GPT-6 Astra or Claude Fable 5.1. A new project on Gemini 3.1 Flash-Lite pays $11.25: 10,000 × (1,500 × $0.25 + 500 × $1.50) ÷ 1,000,000. The bill you actually get depends on four things the list price hides: thinking tokens billed as output, long-context thresholds (100K, 200K or 272K tokens depending on the model), caching, and Batch discounts. The workload rule near the end turns all of this into a pick.

Gemini API vs OpenAI vs Claude pricing by tier (October 2026)

The table matches models by price band and each vendor's positioning, not by measured quality. All prices are Standard (on-demand) rates in US dollars per million tokens, as of October 8, 2026, from Google's Gemini API pricing page, OpenAI's API pricing page and Anthropic's Claude pricing page.

Gemini's cached-input prices are context-caching rates and also carry an hourly storage charge, covered in the hidden-cost section below. Gemini 4 Argon, Google's flagship, became generally available in the paid Gemini API on October 13, 2026. Its introductory $2 / $10 runs through November 30, 2026; from December 1 Google's pricing page lists $4 / $20, the same as Opus 5.5 (Argon prices checked October 14). Argon has no free tier and supports implicit caching only, so the cached rate applies only when Google's automatic cache hits. Access by plan and the Argon-versus-Claude cost math are in Gemini 4 Argon vs Claude Opus 5.5 vs Sonnet 5.5.

Google Gemini API prices

Google sells the widest spread of models. Every model below except 3.1 Pro Preview also has a free tier; its rate limits change, so check them in the Gemini API free tier guide before you plan around them. The three 2.5 models are open only to past 2.5 users; Google tells new projects to use its latest models.

Batch and Flex cost half the Standard rate. Priority processing costs 1.8 times Standard.

Gemini 3.5 Flash (last listed at $1.50 / $9.00) and 3.7 Flash are gone from the price list. Google deprecated both on October 8, 2026 and routes their requests to 3.6 Flash and 3.8 Flash (release notes). It doesn't say which rate applies to calls that still use the old IDs; billing most likely follows the model that serves them, but check the SKU in Cloud Billing.

OpenAI API prices

OpenAI's long-context rule applies to the whole request. Once input passes 272K tokens, every token in that call is billed at the long rate, not only the part above the threshold.

Batch and Flex are half price. Fast processing costs 2 times Standard, and Ultrafast costs 6 times on both Astra and GPT-6.1 Sol (Sol Ultrafast, live since October 8, lists $12 / $60). GPT-6.1 Sol and GPT-6 Sol have the same list price, but 6.1 Sol's cached input is half as much. That matters if you reuse long system prompts. The GPT-6.1 Sol vs Claude Opus 5.5 comparison goes deeper on the mid-to-flagship gap.

Claude API prices

Writing to the cache costs 1.25 times the input rate for the 5-minute cache and 2 times for the 1-hour cache. The Batch API is half price. Opus 5.5 fast mode lists $8 / $40. For choosing between the four Claude models on capability rather than price, see Claude Sonnet vs Opus vs Haiku vs Fable.

Is Gemini still the cheapest API at every tier? Not in October 2026

Gemini is cheapest in the mid tier, ties on input in the budget tier only if you have 2.5 access, and at the flagship tier has the lowest list price until November 30, 2026 with Gemini 4 Argon, then ties Opus 5.5. Here is how each band compares at list price.

Budget tier: a tie on input, a 20% gap on output. For past 2.5 users, Gemini 2.5 Flash-Lite, Luna and Haiku 5.5 all charge $0.10 per million input tokens, and 2.5 Flash-Lite's $0.40 output is 20% below the $0.50 the other two charge. On an output-light workload the difference is small: the Starter workload below costs $3.50 on 2.5 Flash-Lite and $4 on the other two. Without 2.5 access, Gemini's budget model is 3.1 Flash-Lite at $0.25 / $1.50, and the gap runs the other way: Starter costs $11.25 on it, about 2.8 times Luna or Haiku 5.5. Luna and Haiku 5.5 cost the same until a prompt passes 100K tokens. Above that, Haiku 5.5 bills every rate at 5 times; Claude Haiku 5.5 vs GPT-6 Luna works through that split.

Mid tier: Gemini 3.8 Flash is the cheapest by a wide margin, for now. At $0.75 / $3.75 it costs well under half of Sol or Sonnet 5.5. From January 1, 2027 it moves to $1.50 / $7.50, which is still below the $2 / $10 band at list price. Any budget that runs past December 2026 should use the 2027 rate.

$2-input band: Gemini 3.1 Pro Preview is the most expensive. Sol, Sonnet 5.5 and 3.1 Pro Preview share a $2 input price. Gemini's $12 output is 20% above the $10 that OpenAI and Anthropic charge. Above 200K tokens, 3.1 Pro Preview moves to $4 / $18. Sonnet 5.5 keeps $2 / $10 up to its 1M context.

Flagship tier: Gemini 4 Argon is cheapest until November 30, then ties Opus 5.5. Argon lists $2 / $10 through November 30, 2026 and $4 / $20 from December 1, the same as Opus 5.5. GPT-6 Astra and Claude Fable 5.1 both list $10 / $50, so Opus 5.5 still costs less than half of either.

Price bands are not capability bands. A model that is twice as expensive per token can still be cheaper per finished task if it needs fewer attempts or shorter outputs. No public cross-vendor quality benchmark settles that for your prompts, so run your own evaluation set before you switch for price alone.

Monthly API cost for three workloads: $3.50 to $32,000

Per-token prices only mean something once they become a monthly bill. The three scenarios below use Standard prices as of October 8, 2026, with no caching and no Batch discount, so they are a ceiling rather than a forecast.

The formula for each cell is:

monthly cost = requests × (input tokens × input price + output tokens × output price) ÷ 1,000,000

• Starter: 10,000 requests a month, 1,500 input + 500 output tokens each. A small app or internal tool.
• Growth: 500,000 requests a month with the same request shape. A production chatbot or extraction pipeline.
• Long documents: 20,000 requests a month, 150,000 input + 2,000 output tokens each. Contract review, long-PDF Q&A, or codebase questions that send a lot of context.

To check one cell, take Claude Sonnet 5.5 on Starter: 10,000 × (1,500 × $2 + 500 × $10) ÷ 1,000,000 = 10,000 × $0.008 = $80.

Three things stand out:

• The long-document column is where the budget tier splits. At 150,000 input tokens, Haiku 5.5 is past its 100K threshold and bills $0.50 / $2.50. That makes it $1,600, five times Luna's $320. Luna stays under its 272K threshold, and Flash-Lite has none.
• Gemini 3.8 Flash's 2027 price doubles every cell. The Growth bill goes from $1,500 to $3,000 a month with no change on your side.
• Ties at list price are real ties only at list price. Sol and Sonnet 5.5 cost the same in all three columns here. Their actual bills will differ by how many thinking tokens each produces and how each tokenizer counts your text.

Hidden costs that change the bill: thinking tokens, long context, caching, Batch

The scenario table assumes every token is visible and billed at Standard rates. Five things break that assumption, and two of them can multiply your bill rather than add to it.

Thinking tokens are billed as output on all three APIs

OpenAI bills reasoning tokens as output tokens. Claude bills extended thinking as output. Google's pricing page states that Gemini output prices include thinking tokens. None of these tokens shows up in the visible answer, but all of them are billed at the output rate, which is the expensive side of every price pair.

Here is an illustration with assumed numbers, not a measurement. Suppose a reasoning task on GPT-6.1 Sol produces 1,500 thinking tokens on top of a 500-token answer. Each request now bills 2,000 output tokens: (1,500 × $2 + 2,000 × $10) ÷ 1,000,000 = $0.023, against $0.008 without thinking. That is almost 3 times the cost. The real ratio depends on the model, the reasoning effort setting and the task, so measure it from the usage figures each response returns. The reasoning token billing guide covers how each vendor reports and limits them.

Long-context thresholds: 100K, 200K and 272K

Each vendor puts its long-context price step in a different place:

This changes the ranking for long prompts. Take one request with 300,000 input tokens and 2,000 output tokens in the $2 band:

• Claude Sonnet 5.5: 300,000 × $2 + 2,000 × $10 = $0.62
• GPT-6.1 Sol: 300,000 × $4 + 2,000 × $15 = $1.23
• Gemini 3.1 Pro Preview: 300,000 × $4 + 2,000 × $18 = $1.24

(Each product is divided by 1,000,000.) At list price, Sonnet 5.5 costs about half as much as the other two once prompts pass 272K. Below 200K, Sol and Sonnet 5.5 cost the same, and 3.1 Pro Preview costs more only through its $12 output price.

Caching cuts repeated input, but write and storage costs apply

If many requests share the same system prompt, documents or tool definitions, cached reads are far cheaper than fresh input:

• OpenAI: GPT-6.1 Sol cached input is $0.10, 5% of the $2 input price. Luna and Astra cache at 10% of input. Cache writes cost $2.50 on Sol, 1.25 times input.
• Claude: cache reads cost 5% of input on Sonnet 5.5 and Opus 5.5, 2.5% on Fable 5.1, and 10% on Haiku 5.5. Writes cost 1.25 times input for the 5-minute cache and 2 times for the 1-hour cache.
• Gemini: context caching costs $0.20 on 3.1 Pro Preview, $0.075 on 3.8 Flash (2026 rate) and $0.01 on 2.5 Flash-Lite. Gemini also charges storage per million cached tokens per hour: $4.50 on 3.1 Pro Preview, $0.50 on 3.8 Flash and $1.00 on 2.5 Flash-Lite.

Here is a worked example. In the Starter workload on Sonnet 5.5, suppose 1,000 of the 1,500 input tokens are a cached system prompt. Each request then costs 500 × $2 + 1,000 × $0.10 + 500 × $10, divided by 1,000,000, which is $0.0061 instead of $0.008. That brings the month to about $61 instead of $80, before cache-write charges. The saving grows with the share of input that repeats.

Batch and Flex halve the price on all three

Google (Batch and Flex), OpenAI (Batch and Flex) and Anthropic (Batch API) all charge half the Standard rate for work that can wait. For a nightly classification job on the Growth workload, Gemini 2.5 Flash-Lite drops from $175 to $87.50 a month if you have 2.5 access, 3.1 Flash-Lite from $562.50 to $281.25, and Sonnet 5.5 from $4,000 to $2,000. If a share of your traffic doesn't need an instant answer, move that share first. It is the only discount that needs no prompt or model changes.

The reverse applies too. Paying for speed raises the price: Gemini Priority costs 1.8 times Standard, OpenAI Fast 2 times and Ultrafast on Astra or 6.1 Sol 6 times, and Opus 5.5 fast mode $8 / $40.

Tokenizers and search grounding: the same job, different counts

The same text produces a different number of tokens on each vendor's tokenizer. A per-million price comparison is only exact if your token counts are measured on each model. Send the same 50–100 real requests to every candidate and compare the billed usage, not the character count.

Grounding with Google Search is a separate line item on Gemini. Gemini 3.x models share 5,000 free search requests a month, then cost $14 per 1,000. Paid-tier Gemini 2.5 models share 1,500 free grounded prompts a day, then pay $35 per 1,000 grounded prompts (the free tier gets 500 a day). An app that grounds every answer can spend more on search than on tokens.

Which API to choose for your workload: a cost decision rule

Start from the shape of your requests, not from the vendor. Apply these rules in order and stop at the first one that fits:

1. Most prompts are under 100K tokens and the task is simple (classification, extraction, routing, short chat): use the budget tier. Gemini 2.5 Flash-Lite, GPT-6 Luna and Claude Haiku 5.5 are within 15% of each other at list price. Pick the one that passes your quality checks; if two pass, Flash-Lite is the cheapest, provided your project still has 2.5 access. A new project without it lands on 3.1 Flash-Lite ($0.25 / $1.50) or 3.5 Flash-Lite ($0.30 / $2.50), and then Luna or Haiku 5.5 is cheaper: $4 against $11.25 a month for 3.1 Flash-Lite on the Starter workload.
2. Prompts regularly exceed 100K tokens on a budget model: leave Haiku 5.5 out. Luna keeps its $0.10 / $0.50 rate up to 272K, and Flash-Lite has no threshold. Without 2.5 access, Luna is the cheaper of the two: $320 a month on the long-document workload, against $810 on 3.1 Flash-Lite.
3. You need mid-tier quality and can run Gemini: Gemini 3.8 Flash is the cheapest mid-tier option through 2026. Budget at the 2027 rate of $1.50 / $7.50, which still undercuts Sol and Sonnet 5.5.
4. You need a $2-input model and prompts stay under 200K: GPT-6.1 Sol and Claude Sonnet 5.5 cost the same and both cost less than Gemini 3.1 Pro Preview on output. Choose on quality and on which one produces fewer thinking tokens for your tasks.
5. Prompts regularly pass 272K tokens: Claude Sonnet 5.5 or Opus 5.5 avoid the long-context step that applies to GPT-6 models above 272K and to Gemini Pro above 200K.
6. You need the top tier: Gemini 4 Argon at $2 / $10 is the cheapest flagship through November 30, 2026; from December 1 it costs the same $4 / $20 as Opus 5.5. Pay $10 / $50 for GPT-6 Astra or Claude Fable 5.1 only when your evaluation shows they finish tasks that Argon or Opus 5.5 cannot.

Then apply the discounts that fit any vendor: Batch for anything that can wait, caching for repeated prefixes, and a token budget on reasoning effort.

Model tiering usually saves more than switching vendors. Suppose you route 70% of the Growth workload to the vendor's budget model, 25% to its mid model and 5% to its flagship. These are the monthly totals at list price, with the same request shape as above:

With Gemini 4 Argon as the Gemini flagship instead of 3.1 Pro Preview, the 5% slice costs $200 a month through November 30, 2026 and $400 from December 1. The two Gemini mixes then come to $697.50 and $968.75, rising to $897.50 and $1,168.75 in December.

The split is an assumption. It only holds if the cheaper models actually pass your checks on the traffic you send them. If your budget model handles a larger share, the savings grow.

If Gemini's low prices are the reason you're considering it, Cheap Gemini API or Better Alternative? covers when switching away from Gemini actually saves money.

Frequently asked questions about Gemini, OpenAI and Claude API costs

Which AI API is most cost-effective in 2026?

At list price as of October 8, 2026, Gemini 2.5 Flash-Lite ($0.10 / $0.40 per million tokens) is the cheapest model from the three vendors if your project already has 2.5 access; for a new project, GPT-6 Luna and Claude Haiku 5.5 ($0.10 / $0.50) are cheapest, because the cheapest Gemini left is 3.1 Flash-Lite at $0.25 / $1.50. On the Starter workload that's $3.50, $4 and $11.25 a month. Cost-effectiveness also depends on whether the cheap model passes your task. If it needs retries or a bigger model's help, the effective price rises. Haiku 5.5 is only competitive for prompts under 100K tokens. Google limits the 2.5 models to users who have actively used them before but hasn't deprecated them.

Is Gemini API expensive?

Not compared with the other two at most tiers. Gemini 3.8 Flash costs $0.75 / $3.75 through 2026, well under the $2 / $10 of GPT-6.1 Sol or Claude Sonnet 5.5. The exceptions: Gemini 3.1 Pro Preview's $12 output is above Sol and Sonnet 5.5, its price doubles to $4 / $18 above 200K-token prompts, and it has no free tier.

What are the differences in price between Gemini and Claude?

At the budget tier, Gemini 2.5 Flash-Lite ($0.10 / $0.40) beats Claude Haiku 5.5 ($0.10 / $0.50) on output if your project already has 2.5 access, while a new project's cheapest Gemini, 3.1 Flash-Lite at $0.25 / $1.50, costs more than Haiku 5.5 on both. In the $2-input band, Claude Sonnet 5.5's $2 / $10 beats Gemini 3.1 Pro Preview's $2 / $12. The biggest structural difference is long context. Claude Sonnet, Opus and Fable keep one price up to 1M tokens, while Gemini Pro models step up above 200K and Haiku 5.5 steps up above 100K.

Is Claude API or Gemini API better?

On price, Gemini wins the mid tier and, for projects that already have 2.5 access, the budget tier on output; for a new project, Claude Haiku 5.5 ($0.10 / $0.50) is cheaper than Gemini 3.1 Flash-Lite ($0.25 / $1.50). Claude wins the $2-input band and long prompts above 200K. On quality, no public cross-vendor test covers your prompts. The reliable answer comes from running the same evaluation set through both and comparing cost per accepted answer, not cost per token.

How can I track AI API costs across Claude, OpenAI and Gemini?

Each vendor shows usage and spend in its own console: the OpenAI Platform, the Claude Console, and Google AI Studio or Google Cloud billing for Gemini. To see one combined number, log the token counts that every API response returns, including thinking tokens, together with the model name. Multiply them by the rate table above. That also shows which model and feature is driving the bill.

Is it worth switching from OpenAI to Gemini to save money?

It depends on which OpenAI model you would replace. Moving from GPT-6.1 Sol to Gemini 3.8 Flash cuts list price by more than half through 2026, and still by about a quarter at the 2027 rate. Moving from GPT-6 Luna to Gemini 2.5 Flash-Lite saves only 20% on output, and only if your project already has 2.5 access; for a new project, Gemini 3.1 Flash-Lite ($0.25 / $1.50) costs more than Luna. Moving from Sol to Gemini 3.1 Pro Preview costs more. Test quality on your own traffic before you move a production workload, and keep your integration modular so you can switch back.