Claude vs GPT vs Gemini: API Pricing Compared (2026)

The three big model vendors all price by the million tokens (MTok), and on a straight per-token comparison the tiering lines up more closely than the marketing around each launch suggests. As of August 2026, the comparable footprints look like this:

Rates are per million tokens, input over output, US dollars. Two figures carry expiry dates: Sonnet 5's $2/$10 intro rate runs through August 31, 2026, and Gemini 3.7 Flash's $0.75/$3.75 intro rate runs through December 31, 2026. Both jump to higher standard rates after those dates, which is why you should never budget off the promotional number.

The verdict for most teams: OpenAI and Anthropic price their flagship tiers within a dollar of each other, and Google's Pro model undercuts both at $2/$12. But the raw rate does not decide the bill. Context windows and prompt caching move the real cost by more than the list-price gap, and the three vendors handle both differently. That is what this guide walks through.

One rule up front. Model names and rates change often. OpenAI alone cycled through GPT-5.4, GPT-5.5, then the GPT-5.6 Sol, Terra, and Luna family across the sources published this year, and Anthropic and Google ship new model names on a similar cadence. The tier structure below holds; the exact model name might not by the time you read this. Treat every figure here as current for August 2026, and point to the official price pages for live rates: Anthropic, OpenAI, and Google.

How to read these tables

All three vendors bill you for two kinds of tokens: input, the text and context you send, and output, the text the model generates. Input almost always costs less than output, and output is where production bills grow. A useful mental model for 2026: you are paying for metered compute shaped by four levers, not a fixed per-request price.

The four levers are the base per-million rate, the context window (and where the rate changes as context grows), prompt caching (and what a cache hit costs), and any service tier such as batch or fast mode. Budget surprises come almost entirely from the interaction of those four, not from the single number on a landing page.

One token is roughly three-quarters of an English word, so a million tokens is about 750,000 words. When you see $5/MTok input, read it as $5 for roughly 750,000 words of prompt.

Anthropic Claude: the per-token card

Anthropic publishes four current self-serve tiers. The names change across releases, but the tier ladder is stable and spans a ten-to-one spread on both input and output.

The first thing anyone looking at Claude notices is the fixed ratio: output costs five times input on every tier. That ratio is the single most predictable driver of a Claude bill, and it is why capping how much the model writes usually saves more money than trimming the prompt.

Prompt caching on Claude is generous. A cache read, meaning you resend content the API has already processed, costs one-tenth of the base input rate, so $0.50/MTok on Opus instead of $5. There are two cache write rates, 1.25x base input for a five-minute cache and 2x for a one-hour cache, paid only when you first store the content. If the same system prompt, tool definitions, or document prefix comes back in later requests, the read discount pays for the write quickly.

Anthropic's Batch API cuts both input and output in half for jobs that can wait up to a day. The batch discount stacks with the cache-read discount. For agentic and long-context work on current Claude models you also have the option of US-only inference, which applies a 1.1x multiplier to everything, and Claude's fast mode, which runs Opus 5 at $10/$50 for faster output.

The Sonnet caveat matters for planning. Anthropic launched Sonnet 5 at a temporary $2/$10 rate, and that rate expires August 31, 2026, after which you bill at $3/$15. Teams that built budgets on the intro number face a 50% increase on Sonnet on September 1. Forecast on the durable $3/$15 figure, not the promotional one.

OpenAI GPT: the service-tier card

OpenAI now splits its flagship family into three named tiers: Sol, Terra, and Luna. The current card reads as follows for standard processing.

OpenAI's model naming moved fast this year, and the tiers above supersede the GPT-5.4 ($2.50/$15) and GPT-5.5 ($5/$30) figures that earlier 2026 sources published. The name churn is exactly why you should confirm the current lineup on OpenAI's pricing page before you commit.

Output ratios run higher on the OpenAI side than on Claude. Across the current family output commonly costs four to eight times input, and the premium end of the lineup is steeper: Sol sells output at $30, six times its input price, and fast mode pushes that to $60.

Two mechanics stand out. First, a long-context schedule kicks in once an input crosses the short-context window, around 272K tokens. Above that threshold OpenAI charges roughly 2x input and 1.5x output for the whole session, not just the overflow. A request that fits under the line costs one price; the same request that crosses it costs a different one. That makes the threshold a budgeting event, not a footnote.

Second, OpenAI sells access to its latency and cost controls as named service tiers. Standard is the default synchronous rate. Batch prices the same models at 50% for up to 24 hours of turnaround. Flex follows the batch cost logic but trades it for slower, lower-priority delivery. Fast mode (formerly priority) protects latency at a premium, roughly 2x standard. The dangerous mistake is memorizing one GPT number and applying it everywhere, because a single model can appear at four different price points depending on the tier you request.

Cached input on OpenAI runs at about one-tenth of standard input, with a separate cache write charge. OpenAI also prices tools separately: web search runs $10 per thousand calls plus search content tokens billed at the model rate, and a tool call costs $2.50 per thousand.

Google Gemini: the intro-rate card

Google prices Gemini per million tokens, and it bills output including thinking tokens, which matters for reasoning-heavy workloads because the thinking phase is not free. The current lineup:

The Gemini 3.7 and 3.6 Flash models run on introductory pricing of $0.75/$3.75 through December 31, 2026, then step up to $1.50/$7.50 on January 1, 2027. Like the Sonnet intro window, that is a schedule you must fold into cost forecasts.

Google's Flash line is the price leader on pure input cost most quarters, and the Pro model undercuts the OpenAI and Anthropic flagships. The trade is in the context mechanics. Gemini Pro rates are tiered at a 200K-token boundary: prompts under 200K bill at the standard rate, and prompts over 200K jump to 2x input and 1.5x output. The same tiered design appears across Google's larger models, so long-document and retrieval-heavy workloads cross a real pricing line.

Gemini's context caching reads at roughly one-tenth of input, and Google adds a storage charge on top, billed per million tokens per hour while the cache sits in memory. Standard Gemini 3.x Flash storage runs $0.50/MTok/hour through 2026 and $1.00/MTok/hour after that, and the Pro model charges $4.50/MTok/hour. The storage line is the piece most multi-vendor comparisons miss, and it favors short-lived caches.

Google also offers a free tier on many Gemini models, free tier plus prepaid then pay-as-you-go scaling, and a batch and flex mode at 50% of standard. Priority runs at a premium, roughly 1.8x standard, mirroring OpenAI's fast mode.

Side by side: the honest comparison

Put the three vendors on the same scale and the pattern is clear. Entry tiers span from $0.20 to $1 per million input tokens. Mid tiers cluster around $2 to $3 input. Flagship tiers cluster at $5 input on the OpenAI and Anthropic sides, with Google's Pro at $2. On the output side, flagships run $25 to $30, and Google's Pro again breaks the pattern at $12.

That last line is worth saying plainly because it is the counterintuitive finding of the comparison. Google's strongest current Pro model bills at $2/$12, under the $5/$25 Opus and the $5/$30 Sol. If you compare solely on list price, Gemini's flagship looks like a bargain relative to the other two. If you compare something more complex, such as a 300K-token agent loop, the tiered Google pricing erases part of that advantage, and the OpenAI long-context schedule adds cost its way too.

Nobody can hand you a single winner from the rate card, and the sources that try are oversimplifying. The honest take: on list price, Google tends to be the least expensive on both input and output at equivalent tiers, OpenAI and Anthropic land close to each other at the flagship level, and Anthropic's cache-read discount is the most aggressive of the three. Those are the durable patterns. The exact spread moves with every model launch.

Context windows change the real cost

The headline rate assumes a short, normal-sized prompt. Production workloads rarely fit that shape. Here is where the three vendors diverge.

Anthropic prices most of its current large models with a 1M-token context and, for current generations, no premium for long input under standard routing. Prompt caching, batch, and data-residency multipliers apply across the full window. Haiku 4.5 tops out at 200K. The practical effect: a big Claude document dump costs the same per token as a small one, but you pay five times for whatever the model writes back.

OpenAI splits at the short-context boundary around 272K tokens. Under that line you get the standard rate; over it the whole session bills 2x input and 1.5x output. For retrieval-heavy assistants that routinely push 300K or more, that threshold is a real cost inflection, not an edge case. The GPT-5.6 lineup lists short and long context columns precisely because OpenAI prices the two lengths differently.

Gemini tiers its Pro models at 200K tokens, with the same 2x input and 1.5x output jump above the line. Its Flash models hold standard rates across the full 1M window, which is why Flash is the safer pick for long-context work on the Google side. Google also documents a 2M-token ceiling on some model configurations in third-party coverage, though current developer-API docs list 1M for most current models.

The takeaway: compare context engineering as carefully as you compare tokens. A vendor with a lower headline rate can be the more expensive choice once your prompts cross the threshold where its pricing steps up.

Prompt caching changes the real cost

Repeated context is where the bill either compounds or collapses, depending on how each vendor bills a cache hit.

Anthropic gives the clearest mechanics: a cache read costs 0.1x the base input rate, so $0.50/MTok on Opus against a $5/MTok fresh read. Five-minute writes cost 1.25x and one-hour writes 2x, paid once. Because the discount is structured as a percentage of input, the saving scales with how expensive your model is. For agentic coding, where the same codebase context reloads on every turn, that single mechanic frequently dominates all other input savings.

OpenAI bills cached input at about one-tenth of standard input too, with a separate cache write charge, so the discount is comparable in shape. The caveat is service tier: cached-input rates move with the tier you select, and the long-context schedule applies to cached input as well.

Gemini's cache read also lands at roughly one-tenth of input, but Google adds the per-hour storage fee on top while the cache is retained. That storage line makes Gemini caching a different economic object: it rewards caches that are read repeatedly and then dropped, and it punishes caches you keep warm across long gaps. On the Pro model the storage price is high enough that a cache should only live while traffic is active.

Across all three, a silent cache miss re-bills the content at the full input rate and nothing notifies you. The discipline is the same on every vendor: keep stable content such as the system prompt and tool definitions at the front of the request, verify the cache hit rate with the provider's usage metadata, and only cache content you will actually read again.

The levers that move the bill

The rate card is the starting line, not the finish. These levers work across all three vendors and are the difference between a steady bill and a hostile one.

Route each request to the smallest model that can handle it. The spread from entry tier to flagship is a ten-to-one gap on most vendors. Classification, extraction, and routine transforms belong on the entry tier; keep the strongest model for reasoning that needs it. This is the biggest lever and it costs nothing to try.

Cache aggressively, then verify the hits. Repeated context should be read at the discounted rate on every vendor. Watch the hit rate, because a silent miss bills you the full input price.

Batch whatever can wait. All three vendors cut rates by about half for asynchronous batch work. Nightly analysis, backfills, and evaluations belong in the queue.

Cap output length. Output is the expensive direction on every vendor, and on Claude it is five times input on every tier. Constrain answers and many bills follow.

Watch the thresholds and the intro windows. Track the 200K and 272K context boundaries, and plan around Sonnet's August 31 expiring rate and Gemini Flash's December 31 one. Budget on durable rates, not promotional ones.

Measure token shape, not just spend. Log input versus output, cache reuse, and context size by feature so you know which workload actually drives the cost.

Which one should you choose

Start from the workload, not the vendor. If you want the least expensive flagship on list price, Google's Pro is the standout. If long, stable agentic context dominates, Anthropic's cache-read discount and flat 1M pricing reward you the most. If you need OpenAI's tooling, ecosystem, and controllable latency tiers, the GPT-5.6 family is the coherent choice, and its long-context threshold is the cost line to watch.

For a team picking a single default model, the tier question usually outweighs the vendor question. A Sonnet or Terra class model at $2 to $3 input covers most production work at a fraction of flagship cost, and reserving the flagship for the handful of requests that genuinely need it keeps the blended rate low. That is standard multi-model practice, and a model router or gateway that sends each request to the tier that fits it is the mechanism teams use at scale. You can learn the full routing and caching playbook in our LLM cost optimization guide.

If your cost problem is specifically Claude Code in the terminal, the per-token card here is the same one that drives agentic coding bills, and our breakdown of Claude Code cost walks through each plan, the API rate tiers, and the levers that keep that specific workload predictable.

Frequently asked questions

Which API is the least expensive? On list price, Google's entry and flagship tiers generally come in below the equivalent OpenAI and Anthropic models. Google's Flash line undercuts on input cost, and Gemini 3.1 Pro at $2/$12 is below both flagship rivals. OpenAI's gpt-5.6-luna at $0.20/$1.20 is the least expensive entry tier of the three.

Is output always more expensive than input? Yes, on every vendor. Output commonly runs four to eight times input, and on Anthropic it is exactly five times on every tier. Output is where the bill grows.

Do prices change often? They do. OpenAI shifted its flagship naming through GPT-5.4, GPT-5.5, and the GPT-5.6 family this year, and Anthropic and Google ship new model names on a similar cadence. Check the official pricing pages for live rates.

What is the long-context threshold? OpenAI prices the whole session 2x input and 1.5x output once input crosses about 272K tokens. Google tiers its Pro models at 200K. Anthropic prices large current models flat across a 1M window.

How does caching cut the cost? All three bill a cache read at about one-tenth of standard input. Anthropic adds write fees of 1.25x and 2x depending on duration. Google adds a per-hour storage fee on top of the read discount.

Is the intro rate safe to budget on? No. Sonnet 5's $2/$10 rate expires August 31, 2026, and Gemini 3.7 Flash's $0.75/$3.75 rate expires December 31, 2026. Budget on the standard rates that follow.

The takeaway

The three vendors price their API tiers closer than their launch marketing suggests. Flagship input rates cluster around $2 to $5, output around $12 to $30, and each vendor has a less expensive entry tier for the work that does not need flagship reasoning. Nobody wins a blanket comparison, because the real bill is decided by context windows, prompt caching, and which tier actually handles each request.

Build the comparison on the durable pattern, not the model name: compare your token shape against each vendor's thresholds, cache the repeated context, cap the output, and route each request to the tier that fits it. That discipline, not the rate card alone, is what keeps a multi-provider LLM bill predictable, and it is the lever every team controls once the marketing settles.