Gemini vs ChatGPT

The honest answer to Gemini or ChatGPT is that on the specs people usually compare — context window, multimodal input, tool use, coding ability — they have converged. Both offer roughly a million tokens of context. Both take text and images. Both call functions, search the web and drive a computer. Choosing on those axes is choosing on a tie.

The differences that survive scrutiny are about price structure, and they are stranger than the headline rates suggest. Verified against Google's and OpenAI's own documentation on 23 August 2026.

	Gemini 3.7 Flash	Gemini 3.1 Pro (preview)	GPT-5.6 Sol	GPT-5.6 Terra	GPT-5.6 Luna
Input / MTok	$0.75*	$2.00 (≤200k) / $4.00 (>200k)	$4.00	$2.00	$0.20
Output / MTok	$3.75*	$12.00 (≤200k) / $18.00 (>200k)	$20.00	$12.00	$1.20
Context	—	—	1.05M	1.05M	1.05M
Knowledge cutoff	—	—	16 Feb 2026	16 Feb 2026	16 Feb 2026

* Promotional. See below.

Gemini 3.7 Flash's price doubles on 1 January 2027

This is the single most important thing in the comparison and it is a footnote on Google's pricing page. Gemini 3.7 Flash and 3.6 Flash are $0.75 input / $3.75 output through 31 December 2026, then $1.50 / $7.50 from 1 January 2027.

Not an estimate, not a rumour — a published, dated doubling.

If you are pricing a product on Gemini Flash today, you are modelling on a rate with a known expiry. A margin that works at $0.75 needs to work at $1.50, and anything with a twelve-month payback assumption crosses the boundary before it pays back. Build the sensitivity analysis at the post-January number and treat the current rate as a discount you are enjoying rather than a price you are relying on.

Note also that Gemini 3.5 Flash is more expensive than 3.7 and 3.6 Flash — $1.50 / $9.00. The newer models are cheaper than the older one. Version numbers do not predict price here, so read the table rather than assuming.

GPT-5.6 Luna is the cheapest serious option by a wide margin

At $0.20 input / $1.20 output, Luna is roughly a quarter the input cost of Gemini 3.7 Flash's promotional rate, and an eighth of it after January. Against its own family it is 20× cheaper on input than Sol and 10× cheaper on output.

It still carries the 1.05M context window, the February 2026 knowledge cutoff, and the same capability list — functions, web search, file search, computer use. On paper the concessions are quality-related rather than structural, which makes it the natural default for high-volume, well-specified work: classification, extraction, routing, summarisation.

The obvious move for anything running at scale is to test Luna first and escalate only where it demonstrably fails. Most production pipelines are not doing frontier-difficulty reasoning; they are doing a lot of medium-difficulty transformation, and paying Sol rates for that is the most common way to overspend on inference.

Gemini 3.1 Pro charges double for long prompts

Google's Pro tier prices by prompt length: $2 / $12 for prompts of 200k tokens or fewer, $4 / $18 above that.

This is a structural trap for exactly the use case that makes a large context window attractive. Cross 200,000 tokens — one large codebase dump, one long document set, one conversation that grew — and your input cost doubles while output rises 50%. It is a cliff, not a gradient, and a single request either side of it costs very differently.

If your workload sits near that boundary, the highest-leverage optimisation is not a cheaper model. It is keeping prompts under 200k: retrieve the relevant sections instead of pasting everything, summarise conversation history, and drop transcript boilerplate. OpenAI's flat per-token pricing has no such cliff, which for long-context work can matter more than the headline rate.

The Gemini lineup is Flash-first, and Pro is still preview

Something structural shows up in Google's model list: the stable models are almost all Flash — 3.7, 3.6, 3.5, plus Flash-Lite variants — while Gemini 3.1 Pro is preview. The newest stable Gemini, 3.7 Flash, is described for complex coding and agentic workflows, work that would once have implied the Pro tier.

OpenAI's lineup is shaped differently: three general-purpose GPT-5.6 models — Sol, Terra, Luna — all stable, all at the same context window and knowledge cutoff, differing in price and capability rather than in release status.

For production, that difference is not academic. Preview models carry no stability promise. If you need a Pro-class Gemini under a support commitment today, check its status before designing around it.

How to actually choose

Since the capability specs have converged, decide on the things that don't:

• Running at high volume with well-defined tasks? Start with GPT-5.6 Luna. Nothing else is close on price, and the work most pipelines do doesn't need more.
• Prompts regularly above 200k tokens? Prefer flat pricing, or engineer the prompts down. The Gemini Pro cliff is easy to trip and easy to miss on the invoice.
• Budgeting past this year? Model Gemini Flash at $1.50 / $7.50, not $0.75 / $3.75.
• Already inside a cloud? Vertex AI or Azure adjacency will usually outweigh these differences for procurement and data-residency reasons, and that is a legitimate way to decide.
• Everything else? Run your own task on both. Two models this close on paper cannot be separated by a table; they can be separated in ten minutes by your actual prompts.

One caution about the whole genre: these numbers were accurate on 23 August 2026 and the promotional rate above proves the point — some of them have expiry dates already printed on them. Before you commit, check the source.

---

Verified 23 August 2026 against Google's Gemini API pricing and models documentation (ai.google.dev) and OpenAI's models documentation (developers.openai.com). Google's model list does not publish per-model context windows; those cells are left blank rather than filled from secondary sources.