Claude vs ChatGPT vs Gemini: Cost Per Task in 2026

Picking between Claude, ChatGPT and Gemini is no longer a taste question, and by mid 2026 it is no longer even a capability question. The three frontier families are close enough on raw output that the deciding factor is the one your billing dashboard shows you: the real cost per finished task.

The model lineups changed faster than any article can keep up. Anthropic launched Sonnet 5 on June 30 and Opus 5 on July 24. OpenAI put GPT-5.6 Sol, Terra and Luna into public hands on July 10. Google reshaped its Flash lineup on July 21. If your mental model still runs on GPT-5, Gemini 2.5 and Claude Opus 4.7, you are making decisions on retired prices.

This guide is the 2026 refresh of the cost-per-task framework. Direct numbers, no vendor cheerleading, and a decision tree you can actually use on Monday morning.

The current line-ups and prices

As of early August 2026, here is where each family sits. Prices are public US list rates per million tokens, standard (non-batch, non-cached).

Tier	Anthropic	OpenAI	Google
Flagship in/out	Opus 5, $5 / $25	GPT-5.6 Sol, $5 / $30	Gemini 3.5 Pro (premium)
Workhorse in/out	Sonnet 5, $2 / $10	GPT-5.6 Terra, $2.50 / $15	Gemini 3.6 Flash, $1.50 / $7.50
Cheap in/out	Haiku	GPT-5.6 Luna, $1 / $6	Gemini 3.5 Flash-Lite, $0.30 / $2.50

Three things in this table matter more than the raw numbers.

First, the spread between flagship and cheap is now 5x to 25x at a single provider, narrower than the 50x to 200x gap of late 2025. Google, in particular, has been cutting token prices hard. Gemini 3.6 Flash dropped to $1.50 in, $7.50 out, and it also uses fewer output tokens per task than its predecessor.

Second, the old rule “Google is always the cheap option” is dead. Gemini 3.5 Flash Pro sits at premium pricing, and the flashy discount is real only if you target the right tier.

Third, and this is the one people miss: Sonnet 5’s $2 / $10 rate is an intro offer that expires August 31. After that it jumps about 50 percent. If you are pricing a September budget on today’s Sonnet numbers, you will be surprised. Lock in the plan or plan for the higher rate.

How to think about cost per task, not cost per token

A task is one complete unit of work: a pull request review, a blog draft, a research brief, a support reply. The cost is whatever you actually spend to finish it, including retries, tool calls and context bloat.

Three numbers decide cost per task:

Input tokens, dominated by context. A 40-page contract plus instructions is roughly 25,000 to 30,000 input tokens.
Output tokens, dominated by reasoning. Frontier models that think out loud can emit 5,000 to 20,000 output tokens on one hard task.
Number of turns. One-shot is one round. An agentic task can be 20.

Output tokens are still where the bill explodes. They cost 4x to 8x more than input across every provider, and 2026 reasoning models generate far more of them per task than the 2024 chat models did. Everything that follows assumes you read per-token prices with this in mind.

Task 1: Coding

Coding remains the highest-stakes head-to-head, and it is the one where the 2026 lineups genuinely changed the answer.

Claude Opus 5 is the new coding flagship. It launched on July 24 at $5 / $25, same as Opus 4.8, with benchmark wins over Fable 5 on Frontier-Bench, GDPval-AA and CursorBench. It is especially strong at self-verification and multi-file refactors.
Claude Sonnet 5 remains the price-to-performance coding default. It clears the overwhelming majority of real work at $2 / $10 while the intro lasts.
GPT-5.6 Sol and Terra are strong on greenfield problems and math-heavy code. Terra at $2.50 / $15 is a credible Sonnet alternative.
Gemini 3.6 Flash is the value play. It is genuinely good on moderate refactors and extremely cheap, and it uses fewer output tokens than before. It is still weaker than Opus on large agentic edits.

Rough cost per coding task:

Task	Best fit	Typical cost
Quick bugfix, one file	Sonnet 5 or Gemini 3.6 Flash	$0.01 to $0.08
Multi-file refactor	Claude Sonnet 5 or GPT-5.6 Terra	$0.10 to $0.50
Big architectural change, 30+ files	Claude Opus 5 or GPT-5.6 Sol	$0.80 to $3
Repo-wide audit, 500k+ tokens	Gemini Pro or Opus	$0.30 to $2

If you live inside a coding agent all day, Sonnet 5 is the sensible default until the intro price dies. Opus 5 earns its cost only when the diff has to be right the first time, which for architecture and security-critical changes it often does.

Task 2: Long-form writing

Drafting reports, articles and decks. This is where no price table will settle the argument, because output quality differs less than the models’ voices.

Claude writes with the most natural rhythm and the least filler, even after Sonnet 5’s token-efficiency changes.
ChatGPT is the most flexible across formats but still drifts toward listicle structure unless pushed.
Gemini is the driest of the three but the most factual on news-adjacent topics, thanks to search grounding.

Cost per 1,500-word draft, two to three revision turns:

Provider	Tier	Typical cost
Claude	Sonnet 5	$0.08 to $0.18
OpenAI	GPT-5.6 Luna	$0.02 to $0.05
Google	Gemini 3.6 Flash	$0.01 to $0.04

If you write a newsletter and care about voice, Sonnet wins while the intro price holds. If you write at high volume and want to run five candidates, Gemini 3.6 Flash at 1 to 4 cents per draft is the rational pick. The drop in quality is small and the drop in price is large.

Task 3: Research and synthesis

Reading twenty PDFs and producing a three-page brief. Two factors decide this: corpus size and whether you need citations.

For corpora under 200k tokens, every family handles it. For corpora above 500k tokens, Google still owns the context window, and that advantage has not changed. Gemini Pro keeps the whole thing in context where Claude and OpenAI clip or chunk.

One 2026 caveat: the gated frontier models complicate things. GPT-5.6 Sol had a government-only preview before its July public release, and the process surrounding it is still settling. Most teams do not need Sol for research, and the general models are enough.

If you need real source citations, ground the call in a retrieval layer. All three providers will confidently misattribute under load.

Typical cost for a 60-document research brief:

Approach	Cost
Gemini Pro, single big-context call	$0.30 to $1.00
Claude Sonnet 5 with a retrieval shim	$0.15 to $0.70
GPT-5.6 Terra with chunked summarization	$0.25 to $0.90

Task 4: Vision and document parsing

Reading invoices, screenshots, charts and handwritten notes.

Claude is the most accurate on dense structured documents, tables and forms.
Gemini is the most accurate on natural images and charts, and the cheapest at volume.
GPT-5.6 sits in the middle and now includes computer-use capabilities in Sol.

For a batch of 1,000 receipts:

Provider	Tier	Typical cost
Claude	Haiku	$3 to $8
OpenAI	GPT-5.6 Luna	$0.40 to $1.50
Google	Gemini 3.5 Flash-Lite	$0.20 to $1

Gemini Flash-Lite on vision is a steal. If you do OCR at any volume, default to it and only fall back to Claude on the documents that fail validation.

Task 5: Customer support and high-volume chat

Replying to 100,000 tickets a month. At this volume the cheap tiers are the whole game, and a 10x price gap is real money.

For a 500-token-in, 200-token-out reply:

Provider	Tier	Cost per reply	Cost per 100k replies
Anthropic	Haiku	$0.0012	$120
OpenAI	GPT-5.6 Luna	$0.0006	$60
Google	Gemini 3.5 Flash-Lite	$0.0001	$10

Luna beats the old GPT-5-nano estimate at $1 / $6, and Gemini Flash-Lite remains the volume king at roughly 10x cheaper than Haiku. For most support flows the quality is indistinguishable. Reserve Haiku for escalations.

Task 6: Agentic workflows

This is the category that breaks naive cost estimates, and it is the fastest-growing part of the market. An agent that runs 30 tool calls burns context, retries on errors and racks up output tokens on every reasoning pass.

Two rules still hold:

Flagship for the planner, workhorse for the workers. Opus 5 or Sol picks the plan. Sonnet 5, Terra or Gemini Flash executes the steps. This usually cuts the bill by 60 to 80 percent.
Cache aggressively. Anthropic and OpenAI both discount cached input heavily. If your system prompt is 8,000 tokens and you run 50 agent steps, that is 400,000 tokens you should not pay full price for.

A medium-complexity agent run (10 steps, 50k tokens of context):

Setup	Cost
All Opus 5	$2 to $6
Opus 5 planner + Sonnet 5 workers	$0.60 to $1.50
All Sonnet 5	$0.30 to $0.90
All Gemini 3.6 Flash	$0.15 to $0.50

If the agent is reliable on Sonnet or Gemini Flash, do not pay for Opus. If it is not reliable, fix the prompts before throwing money at the flagship.

A decision tree you can actually use

When a task lands on your desk:

Is it volume-bound (1,000+ runs a day)? Gemini Flash-Lite or GPT-5.6 Luna. Test on 50 samples first.
Is it quality-bound and one-shot (a launch post, a board memo, a contract)? The flagship of whichever provider you trust most. The cost is rounding error versus the consequence.
Is it coding? Sonnet 5 by default while the intro holds. Opus 5 when correctness is non-negotiable. Gemini Pro when you need 500k+ tokens of context.
Is it research with a giant corpus? Gemini Pro.
Is it vision at volume? Gemini Flash-Lite.
Is it agentic? Mixed model: flagship planner, workhorse workers, cache the system prompt.

Where this gets expensive without you noticing

Three patterns quietly drain a budget, and they are unchanged:

Reasoning overhead. Frontier models output more thinking tokens than they used to. A simple question can cost fifty cents on Opus if reasoning is set to high. Default to medium and turn it up only when needed.
Context bloat. Re-sending an entire conversation on every turn is the default in most SDKs. By turn 20, half your bill is recycled context. Trim, summarize or cache.
Wrong-tier overkill. Most people pick the flagship because it is safe. For 70 percent of real tasks the workhorse is indistinguishable. Run a one-week experiment swapping flagship for workhorse and measure the complaint rate.

There is also a 2026-specific trap: pricing churn. Anthropic cut and refined the Opus line, OpenAI muddied the picture with Terra and Luna price adjustments, and Google has moved token prices in opposite directions twice this year. Budgets built on last quarter’s API rates are already wrong.

Where tokenkarma fits

The model-picking framework above gets sharper when you can see the bill it produces. Every provider’s dashboard tells you the month’s total, but none of them tells you that 60 percent of your spend went to one agentic workflow that should have run on Gemini Flash instead of Opus.

That is the gap tokenkarma closes. It pulls usage from Claude, ChatGPT and Gemini into one view, breaks spend down by workflow and by model, and surfaces the exact tasks where you are paying flagship prices for workhorse-grade work. When Sonnet’s intro price expires or Google shifts a tier, the readout shows you the impact on your real workload, not on the abstract per-token rate.

Bottom line

Treat the three providers as a portfolio, not a religion.

Claude Sonnet 5: default for coding and writing, but re-check after the August 31 intro expires.
Claude Opus 5: when correctness is critical and you want the strongest 2026 coder.
GPT-5.6 Sol: frontier reasoning and computer use when you need them.
GPT-5.6 Terra / Luna: cost-sensible mid and high-volume tiers.
Gemini 3.6 Flash / Flash-Lite: value at volume, vision, and agent workers.
Gemini Pro: giant-context research and codebase audits.

The right answer to “Claude, ChatGPT or Gemini” in 2026 is still “yes, depending on the task”. The discipline is knowing which one belongs on which job, and checking the actual bill instead of the sticker price.