AI Agent Task Cost: Claude, GPT and Gemini (Oct 2026)

A simple agent task on a budget-tier model like Claude Haiku 5.5, GPT-6 Luna, or Gemini 3.5 Flash-Lite costs roughly $0.001 to $0.003. A medium task, the kind that resolves a support ticket across a dozen model calls with growing context, runs from under $0.01 to $0.66 depending on the model you pick. A complex coding or research task that accumulates 100,000 input tokens can reach $1.25 on a flagship model or stay under $0.03 on a budget one. The spread across providers is wide enough to matter; the right tier to pick depends on what the task actually needs.

Why Agent Tasks Use More Tokens Than You Expect

Unlike a single chat exchange, an agent loop re-sends its full context on every step: the system prompt, tool definitions, the results of previous tool calls, and the full conversation history. A user might see one completed task, but the model processes every prior turn again on every new call.

So in a long agent loop, input usually outweighs output by a wide margin: you pay more for feeding the model what it already knows than for the answer it produces. A detailed breakdown of why those loops compound so quickly is in the article on the multi-agent cost spiral.

For the examples below, three task sizes set the range:

• Simple task: a single-call classification or short summary, 5,000 input tokens and 500 output tokens.
• Medium task: a support or research task across roughly 12 model calls with accumulating context, 36,000 input tokens and 6,000 output tokens total. That is 12 calls averaging 3,000 input and 500 output tokens each.
• Large task: a coding or deep-research task with 100,000 input tokens and 5,000 output tokens.

Current Prices for the Models You'd Actually Pick

Prices are per million tokens, taken from each provider's pricing page fetched October 26, 2026.

OpenAI (standard tier, short context):

• GPT-6 Luna: $0.10 input / $0.50 output
• GPT-5.6 Luna: $0.20 input / $1.20 output
• GPT-5.6 Terra: $2.00 input / $12.00 output
• GPT-6.1 Sol: $2.00 input / $10.00 output
• GPT-6 Astra: $10.00 input / $50.00 output

Anthropic (standard tier):

• Claude Haiku 5.5 (prompts up to 100,000 tokens): $0.10 input / $0.50 output
• Claude Sonnet 5.5: $2.00 input / $10.00 output
• Claude Opus 5.5: $4.00 input / $20.00 output
• Claude Fable 5.1: $10.00 input / $50.00 output

Google Gemini (paid tier, standard):

• Gemini 3.5 Flash-Lite: $0.30 input / $2.50 output
• Gemini 3.8 Flash: $0.75 input / $3.75 output (promotional rate through December 31, 2026; rises to $1.50/$7.50 from January 1, 2027)
• Gemini 4 Argon: $2.00 input / $10.00 output (introductory rate through November 30, 2026; rises to $4.00/$20.00 from December 1, 2026)

What Each Task Size Actually Costs

The table below applies the token counts from each task size to the standard per-token rates above. All numbers are per task.

Model	Simple (5K in / 500 out)	Medium (36K in / 6K out)	Large (100K in / 5K out)
GPT-6 Luna	$0.0008	$0.0066	$0.0125
Claude Haiku 5.5	$0.0008	$0.0066	$0.0125
GPT-5.6 Luna	$0.0016	$0.0144	$0.026
Gemini 3.5 Flash-Lite	$0.0028	$0.026	$0.043
Gemini 3.8 Flash	$0.0056	$0.050	$0.094
Claude Sonnet 5.5	$0.015	$0.132	$0.25
GPT-6.1 Sol	$0.015	$0.132	$0.25
Gemini 4 Argon	$0.015	$0.132	$0.25
GPT-5.6 Terra	$0.016	$0.144	$0.26
Claude Opus 5.5	$0.030	$0.264	$0.50
GPT-6 Astra	$0.075	$0.660	$1.25
Claude Fable 5.1	$0.075	$0.660	$1.25

At the bottom of the table, GPT-6 Luna and Claude Haiku 5.5 share the same $0.10/$0.50 pricing and produce identical per-task costs. GPT-5.6 Luna sits a step above at $0.20/$1.20, making it roughly twice as expensive per task as those two. Gemini 3.5 Flash-Lite's $0.30/$2.50 rate is higher still, and its heavier output price means it costs more than either OpenAI or Anthropic's cheapest option despite the low input price. The budget tier is not uniform across providers: there is a real spread between the cheapest and least-cheap options in that band.

Mid-tier models cluster more tightly. Claude Sonnet 5.5 at $2.00/$10.00, GPT-6.1 Sol at $2.00/$10.00, and Gemini 4 Argon at $2.00/$10.00 (introductory) all produce the same medium-task cost of $0.132.

The real differences in this table are between tiers, not between providers within the same tier at the mid and flagship levels.

The Pricing Changes Coming Soon

Two prices in the table above are temporary. Gemini 4 Argon's $2.00/$10.00 rate runs through November 30, 2026; from December 1 it moves to $4.00/$20.00. That doubles the medium-task cost from $0.132 to $0.264, putting it in line with Claude Opus 5.5 rather than Claude Sonnet 5.5.

Gemini 3.8 Flash carries a similar caveat: the $0.75/$3.75 promotional rate holds through December 31, 2026, then rises to $1.50/$7.50. At the post-promotion price, its medium-task cost climbs from $0.050 to $0.099.

If you are evaluating Gemini for an agent workload right now, factor in where those prices land in Q1 2027 before committing.

What Runs at Which Tier

Not all tasks need the same model. The cost gap between budget and flagship is roughly 100x at the medium task size, so picking the wrong tier for a high-volume workload is expensive.

Budget tier (Haiku 5.5, GPT-6 Luna): straightforward classification, extraction, routing decisions, format conversion, and tasks where the correct answer is narrow and checkable. These models work well when the output doesn't require nuanced judgment and a wrong answer is easy to catch.

Mid tier (Sonnet 5.5, GPT-6.1 Sol, Gemini 4 Argon at intro price): multi-step research, summarization of complex documents, code generation for well-scoped functions, customer-facing reply drafting where tone matters. The cost premium over the cheapest budget models only pays off when those budget models produce outputs you'd reject in production.

Flagship tier (Opus 5.5, GPT-6 Astra, Fable 5.1): hard reasoning tasks with ambiguous or long-horizon requirements, safety-critical decisions, tasks where a wrong answer has a high downstream cost. At $0.66 per medium task, the bar for using a flagship model should be high.

The practical question is not which model is best in a benchmark but which model is good enough for a specific step. The article on hardcoding a cheaper model per step covers the trade-offs of locking in a tier by hand.

Scale Changes the Math

At a hundred medium agent tasks per day, the cheapest budget models (GPT-6 Luna or Claude Haiku 5.5) cost about $0.66 per day, or roughly $20 per month. A flagship model at the same volume costs $66 per day and roughly $2,000 per month. That is a difference of about $1,960 a month for the same task count.

At 10,000 medium tasks per day, the flagship path costs around $200,000 per month, which is a number that demands a routing strategy. The article on AI cost per user and SaaS margins works through how task costs stack up against subscription pricing.

Batch and Caching Can Cut These Numbers Further

Every provider in this comparison offers a 50% batch discount for asynchronous workloads. If your agent tasks don't need results in real time, submitting them to the Batch API halves all the numbers in the table above.

Prompt caching helps differently. Anthropic's cache hit rate brings input costs to 10% of the standard rate for most models (2.5% for Fable 5.1, 5% for Opus 5.5 and Sonnet 5.5). OpenAI's cached input price for GPT-6 Luna is $0.01 per million tokens, a 90% reduction from the $0.10 standard rate. Because agent tasks are input-heavy and many tasks share the same system prompt and tool definitions, caching the repeated prefix can substantially reduce what the input column actually costs you in practice. The full breakdown on when caching pays off is in the article on prompt caching and real savings numbers.

Routing Costs to the Right Model Per Step

One way to close the gap between budget and flagship costs without degrading quality is to route individual steps to the cheapest model that handles them well. A tool like PromptUnit checks each request as it arrives and, when a cheaper model is just as good for that request, uses it instead. You point your app's AI API base URL at PromptUnit and add your PromptUnit key in the x-promptunit-key header; your own provider keys stay in place. The dashboard shows, per step, which model answered and what it cost. This is a different approach from hardcoding a cheaper model per step, which carries its own failure modes.