GPT-6 Luna vs Gemini 3.8 Flash: The Cheap Tier Just Got Replaced

G​oogle released G​emini 4 Hydrogen on November 12, 2026, and its launch post describes it as "a major standalone upgrade for teams migrating from Gemini 3.8 Flash." That sentence does something no benchmark table can: it changes what a comparison between GPT-6 Luna and Gemini 3.8 Flash is for. The G​emini side of this matchup is no longer a model you adopt - it is a model you are being migrated off. The Luna side is still O​penAI's cheapest G​PT-6 rung, eight weeks past launch and unchanged. So the honest question is not which of these two is better. It is whether the thing that made 3.8 Flash worth its $0.75 per million input tokens still holds for Hydrogen at $1.50 - and whether Luna's $0.10 was ever in the same market.

Here is the short version before the detail. Gemini 3.8 Flash is the more capable model on almost every independent measure and roughly seven and a half times the price per token. GPT-6 Luna writes twice as much output per request, costs about a tenth as much per task on Artificial Analysis's suite, and is markedly slower to start answering. Hydrogen does not resolve that - it moves the G​emini tier up in both price and capability, leaving the cheap end of the market to Luna by default.

The spec sheets, side by side

• Rate card, both halves - GPT-6 Luna at $0.10 in and $0.50 out per 1M against Gemini 3.8 Flash at $0.75 and $3.75; the same 7.5x appears on input and output, which is unusual and means no workload shape can arbitrage the ratio

• Cached input per 1M - GPT-6 Luna $0.01 vs Gemini 3.8 Flash $0.075, and note that G​oogle quotes this as a cache-hit rate rather than a cache-write rate, so it is not directly comparable to O​penAI's two-number scheme

• Long-context pricing - GPT-6 Luna doubles input and cache and multiplies output by 1.5 above 272,000 prompt tokens; Gemini 3.8 Flash quotes a single flat rate across its whole window

• Context window - 1,050,000 tokens for GPT-6 Luna vs 1,048,576 for Gemini 3.8 Flash, a difference of 1,424 tokens that no application will ever notice

• Maximum output - GPT-6 Luna 128,000 tokens vs Gemini 3.8 Flash 65,536, which is a real difference the first time you ask for a long document

• What each one will actually ingest - Gemini 3.8 Flash takes video and audio as well as text, images and files; GPT-6 Luna takes the last three and refuses the first two

• The effort dial - Luna's ladder runs from an outright none (medium if you say nothing) to max; G​emini's starts higher, defaults to high, and G​oogle warns it burns more tokens the further up you push it

• Release - Gemini 3.8 Flash on September 2, 2026; GPT-6 Luna on September 22, 2026; Hydrogen, its replacement, on November 12, 2026

• Artificial Analysis Intelligence Index v5.0.1 - 31 for Gemini 3.8 Flash at high, 28 for GPT-6 Luna at max, so the cheaper model gives up three points while running its most expensive setting

• Waiting for the first word - 16.98 s on Gemini 3.8 Flash against 79.33 s on GPT-6 Luna, measured by Artificial Analysis; the G​emini model is slower to finish a whole task, 15 minutes against Luna's 11.8

• What the index run cost to complete - Gemini 3.8 Flash $7.77 a task, or $5,764 for the full suite, against GPT-6 Luna's $0.15 and $136

Seven and a half times the price, and where it shows

The intelligence gap is three points: 31 against 28. That is smaller than the price ratio would suggest and smaller than most people guess, but it is not the interesting part of the comparison, because the two scores were taken at different effort settings - G​emini at high, Luna at max. Luna's maximum-effort score is the best it can do; G​emini's high-effort score is its default. G​oogle's model has headroom Luna does not.

Where the three points come from is more useful than the number. Split the index into its components and a pattern appears: the two models are close on coding-style evaluations and far apart on anything requiring recall or refusal.

• Terminal-Bench 4.0 - Gemini 3.8 Flash 19.7% vs GPT-6 Luna 12.6%; both are poor, and neither is a terminal agent

• Terminal-Bench-Science 0.1 - Gemini 3.8 Flash 10.0% vs GPT-6 Luna 8.6%, which is a tie in practical terms

• HLE-Diamond - Gemini 3.8 Flash 33.1% vs GPT-6 Luna 22.2%, an eleven-point spread on hard reasoning that neither model handles well

• FrontierCode-AA 2.0 - Gemini 3.8 Flash 18.7% vs GPT-6 Luna 25.2%, the one coding evaluation where Luna leads, though both sit far below the models that clear 46%

• AA-Omniscience v1.1 net - Gemini 3.8 Flash +23 vs GPT-6 Luna −8; this is the widest and most consequential gap in the table

• AA-Briefcase v1.1 Elo - GPT-6 Luna 1337 vs Gemini 3.8 Flash 1203, a knowledge-work measure where the cheaper model wins

That last pair is the one to sit with. On questions of what a model knows and whether it will admit not knowing, Gemini 3.8 Flash nets +23 and Luna nets −8 - meaning Luna's fabricated answers outweigh its correct ones often enough to pull the total below zero. On briefcase-style knowledge work, Luna scores 134 Elo higher. Two evaluations, opposite verdicts, same two models. The resolution is that Briefcase rewards following instructions over a supplied corpus, while Omniscience rewards knowing things without one. If your application ships the documents, Luna's weakness is contained. If it answers from memory, it is not.

Why cost per task is 50x when the rate card is 7.5x

The list prices differ by a factor of 7.5. The measured cost to complete Artificial Analysis's Intelligence Index differs by a factor of 50-odd: $7.77 for Gemini 3.8 Flash against $0.15 for GPT-6 Luna. The gap between those two ratios is entirely token consumption, and it is the most under-reported number in every model comparison published this year.

Gemini 3.8 Flash is a heavy reasoner. G​oogle says so directly: on complex tasks it "exhibits greater diligence - executing extra reasoning steps, and calling tools iteratively," and "might use more tokens to maximize performance, especially at higher effort levels." The board's own accounting agrees - 98,900 output tokens per task for G​emini against 75,300 for Luna, with 60,100 and 57,600 of those being reasoning tokens respectively. Multiply that token count by a rate card that is already 7.5x higher and the per-task ratio compounds to 50x. This is the single strongest argument for Luna that exists, and it is not about intelligence at all.

It also cuts the other way on a different workload. If your prompts are long and your answers are short - retrieval over a fixed corpus, classification with a one-token output, log triage - then output tokens are a rounding error on both cards and the 7.5x input ratio is the whole story. G​emini's verbosity never gets a chance to compound, and Luna's advantage shrinks from 50x to about 7x. The number to model is your own ratio of input to output tokens, not the one on the benchmark.

The feature gap nobody puts in a comparison table

Gemini 3.8 Flash ingests video and audio natively. GPT-6 Luna does not - text, image and file, nothing else. For most text workloads that is irrelevant, and it is precisely why the two models get compared at all: both are positioned as the cheap tier, both have roughly a million tokens of context, both take images.

For a workload involving recorded calls, screen recordings, or meeting audio, the comparison ends there. Luna cannot do the job at any price, and the 7.5x is not a premium but a category difference. The same logic applies in reverse to output length: Luna returns up to 128,000 tokens in one response against G​emini's 65,536, which matters when one call is supposed to produce a whole document rather than a paragraph of it.

What Hydrogen changes, and what it does not

G​oogle's own framing of the migration is unusually blunt about the delta: Hydrogen finishes tasks 2.5 times faster at its high default than 3.8 Flash, more than triples Terminal-Bench 4.1 (57% against 18.2%), cuts fabricated claims by more than 5x, and runs 2x to 2.8x faster with 12% to 23% less token use in partner testing. On Artificial Analysis's index it posts 51 at its default high thinking level against Argon's 50, and 54 at xhigh.

It is also flat-priced across its full context window at $1.50 per million input tokens and $7.50 per million output - double 3.8 Flash on both sides. So the model that replaces 3.8 Flash is twice the price of the model it replaces, and by G​oogle's own figures roughly 20 points higher on the intelligence index. That is a straightforward trade up the curve, and it removes the middle of the market: the buyer who wanted G​emini-class competence at Flash-class prices now has a model above them and a model below them, with the old option in migration.

None of that changes Luna's position, because Luna was never competing with a 31-index model. It competes with the floor. At $0.10 input and $0.50 output, Luna is still the cheapest way to process a large volume of text with a competent model behind it, and Hydrogen's arrival at $1.50 makes that more true, not less - the G​emini option that used to be 7.5x Luna's price is now 15x.

The third-party results that disagree

Not every independent measurement puts these models three points apart. StealthMark, a comparison service that runs its own evaluation set, has Gemini 3.8 Flash ahead of GPT-6 Luna by 8.8 points, 47.9 to 39.1, and leading on eight of nine individual benchmarks. And Deepnote's Agent Initiative Benchmark - which scores models on fifty workplace situations inside a simulated company rather than on questions - did not include Gemini 3.8 Flash at all, so the one evaluation that measures proactive agent behaviour has nothing to say about this pairing.

The spread between "three points" and "eight and a half points" is a reminder that benchmark choice is a modelling decision, not a neutral one. The evaluations that favour G​emini weigh knowledge and reasoning depth; the ones that favour Luna weigh instruction-following over supplied context and cost. Neither set is wrong. A team choosing between these two should run a dozen of its own tasks through both before trusting any of the tables, including the four in this article.

Where each one earns its place now

• Video, audio, or any multimodal input - Gemini 3.8 Flash, or Hydrogen if you are starting fresh. Luna cannot take the input.

• Answers drawn from model knowledge with no documents attached - Gemini 3.8 Flash. The +23 versus −8 Omniscience split is the deciding evidence.

• Interactive latency where a 79-second first token is a product defect - Gemini 3.8 Flash at 17 seconds, though its 15-minute mean time per task is worse than Luna's 11.8, so the advantage is in the first token rather than the last.

• Long generations, one call producing a whole document - GPT-6 Luna, on the strength of a 128,000-token output ceiling.

• High-volume text processing with documents supplied - GPT-6 Luna, where the 50x per-task advantage is real and the factual floor is contained by the corpus.

• New deployments that want a G​emini tier at all - G​emini 4 Hydrogen, not 3.8 Flash. Adopting a model two days after its replacement shipped needs a reason.

The migration question deserves one more sentence, because it is the practical decision hiding inside this comparison. If you run Gemini 3.8 Flash in production today, Hydrogen's numbers argue for moving - but its price argues for modelling the bill first, since the token savings G​oogle reports (12% to 23%) do not offset a doubling of both rates. If you run neither, the choice is between a model being retired and a model that has sat unchanged for eight weeks, and the answer depends entirely on whether your workload needs video input or a factual floor.

Running either without a migration project

The awkward part of a matchup like this is that the two candidates speak different protocols - G​emini's own generate endpoint and O​penAI's chat-completions format - and the model you pick today may be superseded within a quarter, as 3.8 Flash just was. OrcaRouter exposes GPT-6 Luna, Gemini 3.8 Flash, G​emini 4 Hydrogen and more than 200 other models behind one OpenAI-compatible key at 0% markup, provider list price passed through, so G​oogle's introductory pricing and any later change reach the invoice on the same day. That matters more than usual here: 3.8 Flash launched on what G​oogle called an introductory rate, and a model on introductory pricing that is being migrated off is a line item worth watching. Switching a route from one model to another is a routing-DSL change rather than a new SDK, automatic failover keeps a workload up when one provider degrades, and model fusion lets a panel answer where a single cheap classification is too consequential to trust to one model. We do not host models we do not route, and both of these we route.

What to watch next

Two dates decide how this comparison reads in a month. The first is G​oogle's retirement schedule for Gemini 3.8 Flash - Hydrogen shipped November 12 and the migration language is already public, so the useful question is not whether 3.8 Flash goes away but how long the introductory rate survives it. The second is whether O​penAI puts a GPT-6 Luna successor or a price change on the table; Luna has been static since September 22, and a cheaper or more capable model at the same rung would reset the arithmetic the way Hydrogen just reset G​emini's.

Until one of those lands, the summary is deliberately unglamorous. Gemini 3.8 Flash is the better model and the worse buy, and it is on its way out. GPT-6 Luna is the weaker model with a per-task cost that is an order of magnitude lower and a factual floor that limits where it can be trusted. Neither is a default. The default, for anyone starting now, is Hydrogen on the G​emini side and Luna on the floor - and the gap between those two is no longer 7.5x but 15x, which is a different decision than the one this matchup was written to answer.