Both models are callable on the Gemini API as of October 15, 2026, so the question is no longer whether to wait for Argon. It’s which requests go where. Gemini 3.8 Flash (gemini-3.8-flash) is stable, has a free tier, and costs $0.75 input and $3.75 output per 1M tokens through 2026-12-31, then $1.50 and $7.50. Gemini 4 Argon (gemini-4-argon) went generally available on October 13 on the paid tier only, at $2 and $10 through November 30, then $4 and $20. Route volume to Flash, and send the hard tail, including anything that needs more than 65,536 output tokens, to Argon.
This post compares the two on specs, caching, the benchmark rows both launch posts share, cyber scores and the cost of an identical call, then gives you routing rules. For background, see what Gemini 4 Argon is and what Gemini 3.8 Flash is. The last section shows the Apidog setup that runs one saved request against both.
Side by side
| Gemini 3.8 Flash | Gemini 4 Argon | |
|---|---|---|
| Model ID | gemini-3.8-flash (stable) |
gemini-4-argon (stable, GA October 13) |
| API tiers | Free and paid | Paid only |
| Price per 1M (in / out) | $0.75 / $3.75 through 2026-12-31, then $1.50 / $7.50 | $2 / $10 through 2026-11-30, then $4 / $20 |
| Cached input per 1M | $0.075, then $0.15 | $0.10, then $0.20 (implicit caching only) |
| Batch per 1M (in / out) | $0.375 / $1.875 through 2026-12-31 | $1 / $5, then $2 / $10 (Flex too) |
| Context | 1,048,576 input | 1,048,576, shared by input, thinking and output |
| Output limit | 65,536 | 65,536 default; up to 1,048,576 minus input with long decode |
| Thinking levels | low, medium (default), high; minimal returns HTTP 400 |
low, medium (default), high; no minimal |
| Consumer access | Gemini app on AI Pro and Ultra, AI Mode in Search | Gemini app and Search on AI Pro and Ultra; Antigravity on Ultra |
A few cells need context. Argon’s window is unified, so a 200,000-token prompt leaves 848,576 tokens for thinking and output. Going past 65,536 output tokens takes long decode: the Interactions API with background: true and a max_output_tokens above 65,536. Our 1M output tokens guide walks through it. Argon’s thinking levels match Flash’s names and default, so a medium request carries over unchanged. Flash’s levels are on Google’s thinking docs and in our 3.8 Flash thinking levels guide.
Argon also drops some knobs, per Google’s developer guide. Explicit caching (cachedContents, caches.create, cached_content) isn’t supported, and temperature, top_p and top_k are ignored if you send them. Argon doesn’t support the Live API, audio generation or image generation, so keep those workloads on other models.
Flash’s free tier is rate-limited, and Google says free-tier data is “used to improve our products.” Argon has no free tier; our Argon free access explainer covers the cheapest routes.
The benchmark rows both launch posts share in text
Google’s launch tables for the two models share two rows in text. These are Google-reported numbers, and Argon’s methodology attributes both rows to Vals AI.
| Benchmark | Gemini 3.8 Flash (Sep 2 table) | Gemini 4 Argon (Sep 30 table) |
|---|---|---|
| Vals Finance Agent v2 | 61.4% | 65.4% |
| Harvey Legal Agent Benchmark | 10.0% | 19.6% |
Google published these tables a month apart, against different rival sets, so read the gap as directional rather than a controlled test. Even so, the legal row stands out: Argon’s 19.6% is nearly double Flash’s 10.0%. The finance gap is 4 points.
Everything else in Argon’s table has no 3.8 Flash counterpart in text. On October 1, Arena’s Text leaderboard, a third-party ranking, had Argon (High) at #1 on preliminary votes and Gemini 3.8 Flash (High) at #11. Since GA, Cursor reports 52.8% for Argon on CursorBench at default effort, “on par with Opus 5.5,” with no Flash figure to compare. The full 19-row Argon table, with who measured each row, is in our Argon benchmarks breakdown.
Cyber: Argon vs 3.8 Flash Cyber
The DeepMind cyber page compares Argon with Gemini 3.8 Flash Cyber, the allowlisted cyber sibling of Flash:
| Cyber eval | Gemini 4 Argon | Gemini 3.8 Flash Cyber |
|---|---|---|
| Real-world vulnerability discovery | 85.8% | 71.0% |
| Wiz Penetration Test Benchmark | 70.9% | 58.2% |
Flash Cyber isn’t the general Flash model, so this table doesn’t describe your gemini-3.8-flash traffic. Argon is now on the paid API for any paid-tier project, but Google gave Fairwind partners its full cyber capabilities first while it iterated on guardrails before wider release. Test your own security prompts on the public gemini-4-argon endpoint rather than assuming these scores carry over.
Cost of the same call on both
Take one call with 100,000 input tokens and 8,000 output tokens. Both models bill thinking tokens as output, so the 8,000 includes them.
| Model and period | Input cost | Output cost | Total per call |
|---|---|---|---|
| 3.8 Flash, through 2026-12-31 | 0.1M x $0.75 = $0.075 | 0.008M x $3.75 = $0.030 | $0.105 |
| 3.8 Flash, from 2027-01-01 | 0.1M x $1.50 = $0.150 | 0.008M x $7.50 = $0.060 | $0.210 |
| Argon, through 2026-11-30 | 0.1M x $2 = $0.200 | 0.008M x $10 = $0.080 | $0.280 |
| Argon, from 2026-12-01 | 0.1M x $4 = $0.400 | 0.008M x $20 = $0.160 | $0.560 |
The two price changes land a month apart, so the ratio moves with the calendar:
- Through November 30: Argon costs 8/3 (about 2.67x) Flash per token.
- December 2026: Argon’s $4/$20 against Flash’s $0.75/$3.75 is about 5.33x.
- From January 1, 2027: back to about 2.67x.
At 1,000 of these calls a day, that’s $105 on Flash against $280 on Argon now, and $560 on Argon in December.
Caching narrows the gap on repeated context. Argon’s caching is implicit: prompts of 4,096 tokens or more get 95% off on cache hits, with no cache to create. Cached input costs $0.10 on Argon against $0.075 on Flash through November 30, only 1.33x.
Three things break the tidy ratio:
- Token counts won’t match. The same prompt produces different output and thinking token counts on different models. Argon’s published evals ran at the highest thinking settings, and on 3.8 Flash Google warns that higher effort levels can use more tokens.
- Long outputs change the math. A 200,000-token answer doesn’t fit in Flash’s 65,536-token cap, so you’d split it across several calls. With Argon’s long decode it’s one response: 0.2M x $10 = $2.00 through November 30, or $4.00 after.
- Output caps don’t port. Google’s Argon guide says apps with caps under 32K should raise them to at least 32,768, because thinking eats the budget. A Flash config capped at 8,192 can leave Argon too little room to answer.
Both models discount Batch by half: Flash at $0.375 and $1.875 through December 31, Argon at $1 and $5 through November 30, per the Gemini API pricing page. Our Argon pricing guide and 3.8 Flash pricing guide go deeper.
When to route to which
Route to 3.8 Flash when
- You’re cost-bound or high volume. Flash costs 3/8 of Argon per token now, and 3/16 in December.
- You’re prototyping. Flash has a free tier; Argon doesn’t.
- The answer has to come back fast. Google says hard Argon prompts can take minutes.
- Your responses fit in 65,536 tokens. Most API workloads do.
Route to Argon when
- The work is long-horizon. Google says Argon is “built to sustain deep reasoning across complex, long-horizon workflows,” per its launch post.
- You need outputs beyond 64K. Full-module rewrites, long reports and large migrations are the case for long decode.
- You run legal or finance agents. Those are the two rows where Google published both models, and Argon leads both.
- The task is hard coding work. Cursor reports 52.8% on CursorBench at default effort, on par with Opus 5.5.
You don’t have to pick once. Keep the model name per request, send the bulk of traffic to Flash, and send the hard tail to Argon.
One saved request, two models
Hold the model name in a variable and run the same request against both:
MODEL="${GEMINI_MODEL:-gemini-4-argon}" # or gemini-3.8-flash
curl -s -X POST \
"https://generativelanguage.googleapis.com/v1beta/models/${MODEL}:generateContent" \
-H "x-goog-api-key: $GEMINI_API_KEY" \
-H "Content-Type: application/json" \
-d '{"contents":[{"parts":[{"text":"Summarize this contract clause in 3 bullets."}]}],
"generationConfig":{"thinkingConfig":{"thinkingLevel":"medium"},"maxOutputTokens":32768}}'
There’s no temperature in the body, since Argon ignores it, and the cap sits at Google’s recommended minimum for Argon, 32,768. Every generateContent response ends with usageMetadata. In Apidog, create two environments that share GEMINI_API_KEY and set GEMINI_MODEL to each model, save this request, and add assertions: status 200, the fields your app reads, and a ceiling on usageMetadata.thoughtsTokenCount sized to your cost cap. Add it to a test scenario, run it in both environments, and compare the reports side by side.
Two caveats. Long decode past 65,536 output tokens runs on the Interactions API in background mode, not this call; the Argon API guide has that shape. And hard Argon prompts can take minutes, so raise your client timeout (Google suggests at least 15 minutes for synchronous calls) before you compare latency.
FAQ
Is Gemini 4 Argon better than Gemini 3.8 Flash? On the two rows both launch tables share in text, yes: 65.4% vs 61.4% on Vals Finance Agent v2 and 19.6% vs 10.0% on Harvey’s Legal Agent Benchmark. It also costs about 2.67 times as much per token through November, and about 5.33 times as much in December.
Can I call both models now? Yes. Flash runs on the free and paid tiers. Argon, as gemini-4-argon, runs on the paid tier only.
Gemini 3.8 Flash or Argon for a new project? Start on Flash with the model in a variable. Move the requests that need long outputs, hard reasoning, or legal and finance depth to Argon once you’ve tested it on your own prompts.
Can I reuse my Flash request on Argon? Mostly. The thinking level names and default match. Remove explicit caching, drop sampling parameters, raise output caps to at least 32,768, and lengthen timeouts.
Does Argon replace Gemini 3.8 Flash? Google hasn’t said. Google calls Argon its new frontier model, and Reuters reports it’s larger than the previous Pro line, so it’s a different tier from Flash.
Next step
Save one request, run it on both models, and keep both reports. A day of real traffic will show you which requests earn Argon’s price. Do it before December 1, when Argon’s price doubles and the gap to Flash widens to about 5.33x. Download Apidog to build the comparison.



