“Which is cheaper, GPT, Claude, or Gemini?” has no single answer, because each lab ships a ladder of models, and the three ladders don’t line up rung for rung. The honest comparison is tier by tier. Here’s how the three families stack up as of our 2026-09-06 snapshot, using each model’s separate input and output price plus a blended figure (a 3:1 input-to-output mix, (3×in + out)/4) for quick ranking.
This guide was substantially rewritten on 2026-09-06. We first published it in July, and OpenAI repriced its entire GPT-5.6 line twice in the seven weeks that followed. Every tier verdict below changed. The “What changed since July” section at the end has the details — it’s the most useful part of this update.
Flagship tier
The top-end, high-capability models:
| Model | Input / 1M | Output / 1M | Blended | Context |
|---|---|---|---|---|
| Claude Opus 5 | $5.00 | $25.00 | $10.00 | 1M |
| GPT-5.6 Terra | $2.00 | $12.00 | $4.50 | 1.05M |
| Gemini 3.1 Pro Preview | $2.00 | $12.00 | $4.50 | 1.05M |
GPT-5.6 Terra and Gemini 3.1 Pro Preview are now priced identically — $2 in, $12 out, to the cent, with near-identical context windows. When two labs land on exactly the same numbers, price has stopped being a differentiator at this tier; pick on capability and latency instead.
Claude Opus 5 sits at more than twice the blended cost of either. Anthropic positions Opus as a premium reasoning flagship and prices it like one, and it has held that price without a single change since launch. If your flagship workload is output-heavy the gap widens further — Opus’s $25 output is more than double the other two.
That doesn’t make Opus the wrong choice; it makes it a choice you should only pay for when the task needs it. Many teams route the hard 10% of requests to a flagship and everything else down a tier.
Anthropic also now ships a rung above Opus: Claude Fable 5.1 at $10 / $50 (blended $20, 1M context) — double Opus 5 across the board.
Mid tier — the workhorses
Where most production traffic actually runs:
| Model | Input / 1M | Output / 1M | Blended | Context |
|---|---|---|---|---|
| Claude Sonnet 5 | $2.00 | $10.00 | $4.00 | 1M |
| GPT-5.6 Sol | $2.00 | $10.00 | $4.00 | 1.05M |
| Gemini 3.5 Flash | $1.50 | $9.00 | $3.38 | 1.05M |
Another exact tie: Claude Sonnet 5 and GPT-5.6 Sol are the same price, $2 / $10. Gemini 3.5 Flash undercuts both by about 16% blended.
The spread at this tier is now narrow — $3.38 to $4.00, about 18% — where in July it was $2.25 to $4.00, or 78%. This is the tier to study hardest, because it’s where you’ll spend the most money over a year and where the models are closest in capability. With price converging, capability-per-dollar is doing more of the work than the sticker price.
Budget tier
For classification, routing, extraction, and high-volume simple tasks:
| Model | Input / 1M | Output / 1M | Blended | Context |
|---|---|---|---|---|
| GPT-5.6 Luna | $0.20 | $1.20 | $0.45 | 1.05M |
| Gemini 3.1 Flash Lite | $0.25 | $1.50 | $0.56 | 1.05M |
| Claude Haiku 4.5 | $1.00 | $5.00 | $2.00 | 200K |
This tier inverted. GPT-5.6 Luna is now the cheapest of the three, narrowly under Gemini 3.1 Flash Lite, with both carrying million-token context windows. Claude Haiku 4.5 costs 4.4× Luna on a blended basis and offers a 200K window rather than 1M.
Luna’s position is recent and worth understanding rather than trusting: it launched in July at $1 / $6, fell to $0.60 output by the end of that month, and settled at $1.20 in August after doubling back. It is the cheapest option in this comparison and also the most repriced model in it.
Two things the headline price hides
Context window. Claude’s flagship and mid models carry 1M tokens; Google and OpenAI push to ~1.05M across most tiers; Anthropic’s budget Haiku drops to 200K. If your workload depends on stuffing large documents into a single call, the cheaper model with the bigger window can win twice.
Cache rates. All three families discount cached input heavily — around 90% off on the models here (Claude Sonnet 5 caches at $0.20 vs $2.00 fresh; GPT-5.6 Luna at $0.02 vs $0.20; Gemini 3.1 Pro at $0.20 vs $2.00). If your prompts reuse a big fixed prefix, the cached input price matters more than the sticker input price. One exception worth knowing: Claude Fable 5.1 reads cache at 97.5% off ($0.25 vs $10), a deeper discount than the flat 90% across the rest of Anthropic’s line.
What changed since July
This is the part a static pricing page can’t give you. Between our first publication (2026-07-14 data) and now:
| Model | July | September | |
|---|---|---|---|
| GPT-5.6 Sol | $5 / $30 | $2 / $10 | output −67% |
| GPT-5.6 Terra | $2.50 / $15 | $2 / $12 | output −20% |
| GPT-5.6 Luna | $1 / $6 | $0.20 / $1.20 | output −80% |
| Claude Opus 4.8 / Opus 5 | $5 / $25 | $5 / $25 | unchanged |
| Claude Sonnet 5 | $2 / $10 | $2 / $10 | unchanged |
| Gemini 3.1 Pro Preview | $2 / $12 | $2 / $12 | unchanged |
| Gemini 3.5 Flash | $1.50 / $9 | $1.50 / $9 | unchanged |
Every change in this comparison came from one lab. Anthropic and Google did not move a single price in the three tiers above across seven weeks; OpenAI cut all three of its rungs, one of them by 80%.
Two consequences worth carrying away:
OpenAI’s naming ladder no longer matches its price ladder. In July the order was Sol ($30) > Terra ($15) > Luna ($6). Today Sol ($10 output) is cheaper than Terra ($12) — the model that was the most expensive of the three is now the middle one. If you selected a model by tier name in July, verify what you’re actually paying for.
Our July verdicts are all obsolete. We wrote that Gemini undercut GPT at the flagship tier, that GPT led the mid tier by 44%, and that Gemini Flash Lite was “in a class of its own” on price. Today: flagship is a tie, GPT does not lead the mid tier, and Flash Lite has been undercut. None of those calls survived seven weeks.
So which is cheapest?
- Flagship: GPT = Gemini, tied to the cent; Claude Opus is the premium option at 2.2× either.
- Mid tier: Gemini < GPT = Claude, and the whole tier now sits within 18%.
- Budget: GPT < Gemini, with Claude Haiku 4.4× above both and a much smaller context window.
OpenAI moved from “wins one tier” to “cheapest or joint-cheapest in all three” over seven weeks — not by out-designing anyone, but by cutting prices while the others held. Google remains consistently aggressive and never repriced. Anthropic prices for the top end and asks you to pay for it, unchanged.
But “cheapest blended” is a starting filter, not a verdict — your real input:output ratio, your context size, your cache hit rate, and (above all) which model actually solves your task will move the answer.
And note how fast this page aged: of the eight models in our July tables, two had repriced within seven weeks — and both by enough to change the tier rankings. Each model on computetrail carries its own daily price history, so before you build on any standing in this guide, check whether it still holds.
Figures from computetrail’s 2026-09-06 price snapshot (updated 2026-09-06; first published 2026-07-19 using the 2026-07-14 snapshot). Blended figures assume a 3:1 input-to-output mix and are for ranking, not budgeting. Independent and not affiliated with any provider. See our methodology.