This article provides a step by step guide to comparing the new Gemini 4 models with Gemma 4 on equal terms. Seven models, Gemma 4 26B and 31B, Gemini 3.5 Flash-Lite and 3.8 Flash, and Gemini 4 Hydrogen, Carbon and Argon, run through the same Gemini API endpoint with the same questions, settings and scoring, started at the same moment.
https://github.com/xbill9/gemma4-dev/tree/main/gemini-api-google-hydrogen4
The scores: on 1,319 GSM8K math problems Argon scored 0.9841, Hydrogen and Carbon 0.9803, 3.8 Flash 0.9788, Gemma 4 26B 0.9742, Gemma 4 31B 0.9735 and 3.5 Flash-Lite 0.9629. On 400 BFCL function calls Argon and Gemma 4 31B tied at 0.9675. Gemma 4 ran free; the Gemini runs cost $0.92 to $5.02. Hydrogen answered fastest of the Gemini 4 models, 2.84 s per GSM8K problem against 7.63 s for Argon.
The surprise is in the cost, and it takes the whole picture to see it. Hydrogen is the cheapest Gemini 4 model per token, $7.50 per million output tokens against $10 for Carbon and Argon, and it cost the most per run, $5.02 against $3.46 for Argon. Its thinking eats the token budget: Google's docs say reasoning cannot be turned off on Gemini 4 models, it ran on all 1,319 GSM8K problems at the lowest setting, and it was 55% of the GSM8K bill. Price per token, tokens per answer and thinking per answer together set what a model costs.
Gemini 4 Hydrogen, Carbon and Argon are on the Gemini API next to the 3.x Flash models and Gemma 4. Google's pricing page, saved on 2026-11-14 (paid tier, USD per million tokens; output includes thinking tokens):
| Model | Input | Output | Free tier |
|---|---|---|---|
gemma-4-31b-it ,gemma-4-26b-a4b-it |
free tier only | free tier only | yes |
gemini-3.5-flash-lite |
$0.30 | $2.50 | yes |
gemini-3.8-flash |
$0.75 | $3.75 | yes |
gemini-4-hydrogen |
$1.50 | $7.50 | yes |
gemini-4-carbon |
$2.00 | $10.00 | no |
gemini-4-argon |
$2.00 | $10.00 | no |
Carbon and Argon go to $4 / $20 on 2026-12-01, and 3.8 Flash to $1.50 / $7.50 on 2027-01-01.
Every model got:
/v1beta/openai/chat/completions on the Gemini API
Step 2 shows that setting for each model, and the two models that keep thinking at it.
~/gemini.key
gemma4-dev
The Gemini API answers OpenAI chat requests with the key as a Bearer token. Gemma 4 and Gemini take the same request body; only model changes:
curl -sS https://generativelanguage.googleapis.com/v1beta/openai/chat/completions \
-H "Authorization: Bearer $(cat ~/gemini.key)" -H 'Content-Type: application/json' \
-d '{"model":"gemma-4-31b-it","temperature":0,"max_tokens":2048,"reasoning_effort":"minimal",
"messages":[{"role":"user","content":"Janet has 16 eggs, eats 3, bakes with 4, sells the rest at $2. How much does she make? End with Answer: <number>"}]}'
The reply ends Answer: 18.
The OpenAI endpoint sets thinking with reasoning_effort. levels.sh sends the question to every model with each value:
gemma-4-31b-it minimal thinking 0 finish stop Answer: 18
gemma-4-31b-it none rejected
gemma-4-31b-it low rejected
gemma-4-26b-a4b-it minimal thinking 0 finish stop Answer: 18
gemma-4-26b-a4b-it none rejected
gemini-3.5-flash-lite minimal thinking 0 finish stop Answer: 18
gemini-3.5-flash-lite none rejected
gemini-3.8-flash minimal rejected
gemini-3.8-flash none thinking 70 finish stop Answer: 18
gemini-4-hydrogen minimal rejected
gemini-4-hydrogen none thinking 81 finish stop Answer: 18
gemini-4-carbon minimal rejected
gemini-4-carbon none thinking 0 finish stop Answer: 18
gemini-4-argon minimal rejected
gemini-4-argon none thinking 0 finish stop Answer: 18
Gemma 4 and Flash-Lite go down to minimal, and there they think zero tokens. Hydrogen, 3.8 Flash, Carbon and Argon go down to none; Carbon and Argon think zero tokens there, and Hydrogen and 3.8 Flash keep thinking.
Google documents this. Its OpenAI compatibility page says "Reasoning cannot be turned off for Gemini 2.5 Pro, Gemini 3, or Gemini 4 models", and its thinking page lists Hydrogen's levels as low, medium, high and xhigh, with minimal "Not supported on Gemini 4 models or gemini-3.8-flash". The same page documents none for 2.5 models only. On Hydrogen, 3.8 Flash, Carbon and Argon the endpoint accepts none and treats it as low. On the first 100 GSM8K problems ( levels-sample/), thinking tokens per problem:
| Model | none |
low |
medium |
not set |
|---|---|---|---|---|
| Gemini 4 Hydrogen | 239.0 | 238.5 | 350.6 | 311.8 |
| Gemini 3.8 Flash | 145.3 | 139.9 | 229.8 | 244.6 |
| Gemini 4 Carbon | 28.6 | 24.6 | 136.6 | 125.3 |
| Gemini 4 Argon | 2.3 | 4.9 | 67.7 | 74.7 |
So this run's none is each model's lowest documented level, low. The native API's other controls give the same result on the test question. thinking-controls.sh tries each on Hydrogen and reads the API's own counter, thoughtsTokenCount:
gemini-4-hydrogen {"thinkingBudget":0} answer 88 thoughts 134 total 292 thought-parts 0
gemini-4-hydrogen {"thinkingLevel":"NONE"} rejected: Invalid value at 'generation_config.thinking_config.thinking_level' (type.googleapis.com/g
gemini-4-hydrogen {"thinkingLevel":"MINIMAL"} rejected: Thinking level MINIMAL is not supported for this model. Please retry with other thinking l
gemini-4-hydrogen {"thinkingLevel":"LOW"} answer 70 thoughts 106 total 246 thought-parts 0
gemini-4-hydrogen {"thinkingBudget":0,"includeThoughts":true} answer 73 thoughts 79 total 222 thought-parts 1
With thinkingBudget: 0 and includeThoughts: true, the reply carries the thought itself, beginning "Okay, here's my breakdown of this little egg scenario". 3.8 Flash reports 74 thinking tokens on that question under every control.
Across the full run, Hydrogen thought on 1,319 of 1,319 GSM8K problems and all 400 BFCL calls. 3.8 Flash thought on 1,293 GSM8K problems and 113 BFCL calls, Carbon on 268 GSM8K problems, Argon on 66, and Gemma 4 and Flash-Lite on none. Every one of those tokens is billed as output. Hydrogen also thought on "Say hi." (more-checks.sh).
completion_tokens
On this endpoint completion_tokens counts only the visible answer. Thinking is total_tokens - prompt_tokens - completion_tokens, and Google bills it as output. The scoring script records it as reasoning_tokens on every row.
gemma4-dev names a rig <platform>-<runtime>-<hardware>-<model>. On the Gemini API the platform is gemini, the runtime api and the hardware google, since Google publishes neither; the model slot is the Gemma size or the Gemini family:
gemini-api-google-31b
gemini-api-google-26b
gemini-api-google-flashlite35
gemini-api-google-flash38
gemini-api-google-hydrogen4
gemini-api-google-carbon4
gemini-api-google-argon4
Each tpu.env names the model and the lowest thinking value it accepts, and the seven run.sh files are identical:
MODEL_NAME=gemini-4-hydrogen
LOWEST_REASONING_EFFORT=none
jev-tpu-v5e1/gen_eval.py scores every Gemma 4 rig in the repo. GSM8K is right when the number after Answer: matches; BFCL simple is right when the one function call passes BFCL's own checks. Four settings point it at a hosted API:
| Setting | What it does |
|---|---|
JEV_CHAT_URL |
full chat URL in place of <server>/v1/chat/completions |
JEV_API_KEY_FILE |
sends the file's contents as a Bearer token |
JEV_REASONING_EFFORT |
sends reasoning_effort |
JEV_RETRIES ,JEV_TIMEOUT_S |
retries on HTTP 429/5xx and network errors, and a per-request time limit |
for r in 31b 26b flashlite35 flash38 hydrogen4 carbon4 argon4; do
gemini-api-google-$r/benchmarks/runs/2026-11-14-accuracy-api/run.sh &
done
2026-11-14 09:11:16 BEGIN gemini-4-argon reasoning_effort=none
2026-11-14 09:18:51 END bfcl_simple rc=0 {"task": "bfcl_simple", "max_tokens": 512, "n": 400, "right": 387, "accuracy": 0.9675, "errors": 0, "truncated": 0, "wall_s": 454.8}
2026-11-14 09:48:10 END gsm8k rc=0 {"task": "gsm8k", "max_tokens": 2048, "n": 1319, "right": 1298, "accuracy": 0.9841, "errors": 0, "truncated": 0, "wall_s": 1759.2}
2026-11-14 09:48:10 DONE
Three Gemma 4 requests hit the free tier's rate limit (HTTP 429) after every retry. resend_errors.py sent those three again with the same settings and replaced only those rows; all three came back right. Every answer in the table below is a model answer.
derived.py builds the table from the saved replies. The cost is the token counts the API returned times the list price.
| Model | GSM8K | BFCL simple | GSM8K problems with thinking, at the lowest setting | Cost, both tasks |
|---|---|---|---|---|
| Gemma 4 31B | 0.9735 | 🥇 0.9675 | 0 | 🥇 free tier |
| Gemma 4 26B | 0.9742 | 0.9575 | 0 | 🥇 free tier |
| Gemini 3.5 Flash-Lite | 0.9629 | 0.9350 | 0 | $0.92 |
| Gemini 3.8 Flash | 0.9788 | 0.9400 | 1,293 | $1.81 |
| Gemini 4 Hydrogen | 🥈 0.9803 | 0.9550 | 1,319 | $5.02 |
| Gemini 4 Carbon | 🥈 0.9803 | 0.9600 | 268 | $4.02 |
| Gemini 4 Argon | 🥇 0.9841 | 🥇 0.9675 | 66 | $3.46 |
Against Gemma 4 31B on the same questions, counting the questions each got right that the other missed:
| Model | GSM8K | BFCL simple |
|---|---|---|
| Gemma 4 26B | level (+10 / −9) | level (+5 / −9) |
| Gemini 3.5 Flash-Lite | behind (+6 / −20) | behind (+5 / −18) |
| Gemini 3.8 Flash | level (+10 / −3) | behind (+2 / −13) |
| Gemini 4 Hydrogen | ahead (+11 / −2) | level (+5 / −10) |
| Gemini 4 Carbon | ahead (+14 / −5) | level (+5 / −8) |
| Gemini 4 Argon | ahead (+17 / −3) | level (+7 / −7) |
Ahead and behind mean the 95% range of the difference, from 2,000 resamples of the questions, sits entirely above or below zero.
By list price Hydrogen is the cheapest Gemini 4 model, $7.50 per million output tokens against $10 for Carbon and Argon. It cost the most because it thought on all 1,319 GSM8K problems at its lowest setting, 248.7 thinking tokens each on average, which made thinking 55% of its GSM8K bill, against 11% for Carbon and 2% for Argon. Argon thought on 66 problems and scored the highest.
The Gemini 4 models also count more input tokens for the same text: the GSM8K prompts are 160,432 tokens to Hydrogen, Carbon and Argon and 120,902 to Gemma 4 31B, Flash-Lite and 3.8 Flash, a mean of 121.6 against 91.7 per problem.
Thinking also shares the output limit with the answer. Google's Hydrogen page says max_output_tokens "caps combined thinking and output tokens". At this run's limits of 2,048 and 512 tokens, thinking used the limit before the answer arrived on 4 GSM8K problems and 2 BFCL calls for 3.8 Flash and on 5 GSM8K problems for Carbon; all 11 scored wrong. Sent again with a 16,384-token limit (cut-rows.py), 2 of the 4 on 3.8 Flash and 3 of the 5 on Carbon came back right. Hydrogen had no cut answers.
Each reply records how long its request took. All seven models ran at once, eight requests at a time each, over the same connection, so the times are comparable with each other on that morning:
| Model | GSM8K median | GSM8K 90th percentile | Whole GSM8K run | BFCL median |
|---|---|---|---|---|
| Gemma 4 31B | 8.96 s | 23.93 s | 2,044.1 s | 1.69 s |
| Gemma 4 26B | 8.16 s | 25.80 s | 2,022.8 s | 1.23 s |
| Gemini 3.5 Flash-Lite | 🥇 1.10 s | 🥇 1.54 s | 🥇 195.8 s | 🥇 0.61 s |
| Gemini 3.8 Flash | 3.79 s | 5.43 s | 684.9 s | 1.93 s |
| Gemini 4 Hydrogen | 🥈 2.84 s | 🥈 5.39 s | 🥈 620.8 s | 1.95 s |
| Gemini 4 Carbon | 7.34 s | 20.04 s | 1,803.7 s | 5.29 s |
| Gemini 4 Argon | 7.63 s | 17.57 s | 1,759.2 s | 6.32 s |
Hydrogen writes about 147.8 tokens per second per request, thinking included, against 26.1 for Argon and 29.7 for Carbon. It finished GSM8K in about a third of Argon's time while thinking on every problem. Flash-Lite was the fastest model overall.
| Gemma 4 31B | 3.5 Flash-Lite | 3.8 Flash | Hydrogen | Carbon | Argon | |
|---|---|---|---|---|---|---|
| GSM8K against Gemma 4 31B | — | behind | level | ahead | ahead | 🥇 ahead |
| BFCL against Gemma 4 31B | 🥇 | behind | behind | level | level | 🥇 level |
| Thinks at its lowest setting | no | no | most problems | every problem | some problems | rare |
| Cost, both tasks | 🥇 free tier | $0.92 | $1.81 | $5.02 | $4.02 | $3.46 |
| GSM8K median per request | 8.96 s | 🥇 1.10 s | 3.79 s | 🥈 2.84 s | 7.34 s | 7.63 s |
| Weights you can download | 🥇 yes | no | no | no | no | no |
Google documents both facts behind Hydrogen's bill: reasoning cannot be turned off on Gemini 4 models, and output prices include thinking tokens. Four things in the API still make the bill easy to misread:
reasoning_effort: "none" on the OpenAI endpoint and thinkingBudget: 0 on the native API return no error on Hydrogen. none runs as low, and nothing in the reply says thinking is still on. high: 311.8 thinking tokens per problem on the 100-problem sample against 239.0 at Compare models by cost per answer, measured on your own prompts.
For the best scores on short math problems and single function calls with thinking at its lowest setting, Gemini 4 Argon: highest on GSM8K, tied for highest on BFCL, and the cheapest of the three Gemini 4 models per run because it rarely thinks.
For hard, multi-step work, Google points to Carbon. Its Hydrogen page calls Carbon "our strongest frontier model", one that "spends more thinking tokens per turn to test hypotheses and verify edge cases on hard architecture, debugging, science, and security audits", and suggests pairing the two: Carbon plans and resolves hard roadblocks, Hydrogen does the code edits, subagent tasks and test runs. This run held Carbon to its lowest thinking setting, where it thought on 268 of 1,319 GSM8K problems, so that case is untested here. Five of its GSM8K answers were cut at the 2,048-token limit; with a 16,384-token limit 3 of them came back right, which still leaves it below Argon.
Gemini 4 Hydrogen fits four cases:
For high-volume simple work on a paid tier, Gemini 3.5 Flash-Lite: $0.92 for both tests, no thinking, 1.10 s per problem, at the cost of being behind Gemma 4 31B on both tests.
For tool calling at no cost, Gemma 4 31B: tied with Argon on BFCL, 0.0106 behind it on GSM8K, free on the Gemini API, and the same weights run on your own hardware.
The goal of this article was to compare the new Gemini 4 models with Gemma 4 on equal terms. The key to the solution was running all seven models through one Gemini API endpoint, started together, with the same questions, scoring script and settings, with thinking set as low as each model accepts. The results were:
Scope: one laptop in the US calling the Gemini API on 2026-11-14, one run per model, all seven started together. Thinking was set as low as each model accepts: minimal for Gemma 4 and Flash-Lite, none for the rest, which matched low on a 100-problem sample. The native thinking controls were tried on one question. The scores for 3.8 Flash and Carbon include the 11 answers cut by thinking at the 2,048 and 512 token limits. Three Gemma 4 requests that hit the free tier's rate limit were sent again with the same settings. Costs are token counts times list prices. Run times are from one run with all seven sharing one connection, and will move with load on the service. Hard multi-step tasks and higher thinking levels were not measured, which is the work Google positions Carbon for. Google does not say what hardware or engine serves any of these models.
The strategy for comparing Gemini and Gemma 4 on equal terms was validated with an incremental step by step approach.
Hope you enjoyed it!