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

For high-volume, low-complexity work, the cheapest efficiency-tier models from every major vendor now cost well under a dollar per million tokens in either direction. For frontier reasoning, the top tier of each vendor's lineup still runs into double-digit dollars per million output tokens. Which family is actually cheapest for you depends entirely on your workload — the ratio between input and output tokens, and how deep into a model's context window you run, can swing the real bill more than the sticker price alone.

Last verified 2026-07-25, via a live AbacusAI model-catalog fetch cross-referenced against each vendor's own published pricing page. Any figure that could not be confirmed against a primary source in that pass is named without a price below, rather than printed as clean data.

Price Comparison

Only figures confirmed against a primary or catalog source are shown here — models whose current price could not be verified this pass are deliberately left out of this table rather than estimated.

How to Read These Numbers

Two things distort a naive "cheapest model" comparison, and both are visible directly in the table above.

First, output tokens cost more than input tokens for every model listed — usually somewhere between four and eight times more per million tokens. That means a workload's real cost depends heavily on how much text the model generates, not just how much you feed it. A short prompt that produces a long, detailed answer spends most of its budget on the output column, so comparing only input prices when your workload is generation-heavy will mislead you about the actual bill.

Second, context-window tiering is real and not always obvious from a single headline price. Gemini 2.5 Pro is the clearest example above: both its input and output price roughly double once a request crosses 200,000 tokens of context. A workload that stays comfortably under that threshold pays the lower tier; the same workload with a much longer document attached pays close to double, on the same model, with no other change.

Put the two together and a cheap-looking model on a long, generation-heavy job can end up costing more in absolute terms than an expensive-looking model on a short, input-heavy one — because total cost is price per token multiplied by tokens actually used, in both directions, not a single per-call sticker price.

Models That Are Already Gone (or Going)

Pricing pages tend to hide a fact that matters just as much as the price itself: several models still referenced in older integration code are no longer callable, or are about to stop being callable, on their vendor's first-party API.

• Claude Sonnet 4 (claude-sonnet-4-20250514) — retired on Anthropic's first-party API since June 15, 2026 (deprecated April 14, 2026). Any integration still targeting this exact model id is already failing.
• Claude Sonnet 3.7 (claude-3-7-sonnet-20250219) — retired since February 19, 2026.
• Claude Sonnet 3.5 (claude-3-5-sonnet-20241022) — retired since October 28, 2025.
• Claude Haiku 3.5 (claude-3-5-haiku-20241022) — retired since February 19, 2026.
• Claude Opus 4.1 (claude-opus-4-1-20250805) — deprecated and scheduled to retire August 5, 2026, only days out from this post's verification date. Anthropic's own guidance points integrations toward a newer Opus-tier model, such as Claude Opus 4.7 above.
• Gemini 2.5 Pro and Gemini 2.5 Flash — both carry a scheduled retirement flag around October 16, 2026, despite being current, actively-priced models as of this post.

No GPT-line model in this verification pass carried a retirement or deprecation flag as of 2026-07-25. That doesn't mean it never will — it means the GPT line happened to be clean on this particular date.

Picking a Model by Workload

Rather than naming one winner, it's more useful to map the table above to what your workload actually asks a model to do:

• Classification and extraction — short, structured outputs from a well-defined input. The nano and mini efficiency tier across every vendor (GPT-5 nano, GPT-4.1 nano, GPT-4o mini, GPT-5 mini, GPT-4.1 mini) is built for exactly this, and stays well under a dollar per million tokens in either direction.
• Long-form generation — drafting, summarizing, or rewriting at length. Output price dominates here, so weigh the output column specifically rather than the input price alone; a model with a low input price and a comparatively high output price can still turn out expensive on a generation-heavy job.
• Multi-step reasoning — planning, complex analysis, or agentic tool-calling where an intermediate step's quality compounds. The reasoning and frontier tiers (o3, o4-mini, GPT-5.5, Claude Opus 4.7, Gemini 2.5 Pro) trade a materially higher per-token price for stronger intermediate reasoning, which is the right trade when a wrong intermediate step costs more than the extra tokens would.

The right tier is a function of what the workload asks the model to do, not which vendor's homepage you happened to read last.

What This Costs Inside PromptHub

PromptHub doesn't bill you the vendor's list price directly per call. Every AI-consuming operation spends PromptHub tokens from your account balance, converted through an internal cost formula that combines a base token count, the amount of text involved, and a per-model multiplier specific to whichever model you've configured. Prompt chains meter cost per node that calls a model, with a pre-flight estimate checked against your balance before a run starts; Agent runs meter cost per turn, settled against the real tokens actually used across every round once the turn completes. No PromptHub-specific price is quoted here, since the real figure depends on which model you've selected and how much text each run actually processes — check your account balance for the number that applies to you.

Vendor pricing changes fast enough that a static comparison is a snapshot, not a promise — the figures above were cross-referenced against primary sources on 2026-07-25 and are refreshed as the catalog changes. Create a free account to try prompts, chains, and Agents against any of PromptHub's supported models yourself.