
Gemini 4 vs Gemini 3.1 Pro: Price, Benchmarks & Upgrade Path
By Emma Reed
MyClaw Editorial
Compare in MyClaw
Compare Gemini 4 Argon and Gemini 3.1 Pro on your own work
Give both models the same task and compare quality, speed, and cost.
AI Takeaway
- Does Google show Gemini 4 beating Gemini 3.1 Pro? No. Google hasn't run that test in public. Its Gemini 4 (Argon) launch table has no 3.1 Pro column. The two Google tables share only one test name, Terminal-Bench, and in different versions. Google does state one clear gain: Gemini 4 lifts the output limit from 64K to 1M tokens.
- Can you still use Gemini 3.1 Pro? Yes. It is still a preview model,
gemini-3.1-pro-preview, released on February 19, 2026. Google lists no shutdown date and hasn't shipped a stable version. - Is Gemini 4 cheaper than Gemini 3.1 Pro? It depends on prompt size and timing. For now, Google prices Gemini 4 input at $2 and output at $10 per million tokens. Later, that rises to $4 and $20. Gemini 3.1 Pro charges $2 and $12 for prompts of 200K tokens or less, and $4 and $18 for longer ones.
- When can most teams call Gemini 4? Google hasn't said. Access is opening in waves, and the Fairwind Program came first. There are no dates for anyone else. Gemini 3.1 Pro is open to paid Gemini API users now.
- Does MyClaw support Gemini 4 and Gemini 3.1 Pro together? Yes. Both models run on MyClaw. One agent can process an identical recurring job on each, and you score each answer.
Gemini 4 vs Gemini 3.1 Pro: Core Specs
Gemini 4 vs Gemini 3.1 Pro sets Google's September 2026 flagship against the February Pro preview it follows.
| Item | Gemini 4 (Argon) | Gemini 3.1 Pro |
|---|---|---|
| API model ID | None published | gemini-3.1-pro-preview, plus a custom-tools endpoint |
| Status | Staged rollout | Preview |
| Input context | Not published | 1,048,576 tokens |
| Output limit | 1M tokens | 65,536 tokens |
| Inputs | Google stresses long video, charts and documents | Text, images, video, audio and PDF |
| Thinking | Tested at maximum thinking; other levels undocumented | Low, medium or high; defaults to high |
| Training cutoff | Not published | January 2025 |
Output is the spec Google states most plainly. Its launch post says Gemini 4 can write 1M tokens, "up from the previous 64K." That 64K matches the 3.1 Pro cap. Most other Gemini 4 rows are still blank. For the details Google has released, see our first look at Gemini 4 Argon.

Two Pro Models, Two Release Paths
Any Gemini 4 Argon vs Gemini 3.1 Pro decision starts with access. Google released 3.1 Pro as a preview on February 19, 2026. It replaced Gemini 3 Pro Preview, which shut down on March 9. Seven months later, 3.1 Pro is still a preview. Google's deprecations page lists no shutdown date, and no stable 3.1 Pro has appeared. Our Gemini 3 Pro and 3.1 Pro comparison covers that handover.
The model meant to follow it never shipped. Google announced Gemini 3.5 Pro and then dropped it, which we cover in the story of the missing 3.5 Pro. That makes Gemini 4 the next model above 3.1 Pro. But it arrives gated: staged access, no model ID and no published context window.
So your Gemini 4 vs Gemini 3.1 Pro choice is plain. You can call 3.1 Pro today as a long-running preview. Gemini 4 is a successor most teams can't call yet. Is 3.1 Pro being retired? Not on any announced schedule.
Benchmarks: Google Never Put Them in One Table
Google published two tables seven months apart, and they don't overlap. Its 3.1 Pro table (February, Thinking High) and its Gemini 4 table (September) share no row. So a clean Gemini 4 vs 3.1 Pro scorecard doesn't exist.
| Area (Google-reported) | Gemini 4 | Gemini 3.1 Pro | Comparable? |
|---|---|---|---|
| Terminal-Bench | v4.0: 57.4 | v2.0: 68.5 | No, different versions |
| Long context | GraphWalks: 99.7 up to 128K, 84.2 at 256K–1M | MRCR v2: 84.9 at 128K, 26.3 at 1M | No, different tests |
The Terminal-Bench row is the easiest to misread. Version 4.0 is newer and uses different tasks, so Gemini 4's lower number is not a step backwards.
Each table also has scores the other lacks. The 3.1 Pro rows include ARC-AGI-2 at 77.1, GPQA Diamond at 94.3, SWE-Bench Verified at 80.6, MCP Atlas at 69.2 and BrowseComp at 85.9. Gemini 4's own rows include DeepSWE v1.1 at 77.9, AutomationBench at 51.3, Vals Finance Agent v2 at 65.4 and LVBench at 91.7. Independent leaderboards have started testing both. Their scores aren't Google's, so we leave them out. Every Gemini 4 row Google published is in our breakdown against Fable 5.1.
Gemini 4 vs Gemini 3.1 Pro Pricing: The 200K Line
| Rate (Google) | Input, $/1M | Output, $/1M | Details |
|---|---|---|---|
| Gemini 4 at launch | $2 | $10 | Cache reads $0.10; Google hasn't said when this ends |
| Gemini 4 after launch | $4 | $20 | No long-prompt, batch or storage rules published yet |
| Gemini 3.1 Pro, prompts up to 200K | $2 | $12 | Batch $1/$6; cache $0.20 plus storage |
| Gemini 3.1 Pro, prompts over 200K | $4 | $18 | Batch $2/$9; cache $0.40 plus storage |
The 200K line hits long-document agents hardest. Here are two example calls (our math, at Google's list prices):
- 150K input + 20K output: Gemini 4 $0.50 at launch, $1.00 later; Gemini 3.1 Pro $0.54
- 500K input + 20K output: Gemini 4 $1.20 at launch, $2.40 later; Gemini 3.1 Pro $2.36
At launch, Gemini 4 wins on price at both sizes. Once the launch rate ends, 3.1 Pro is cheaper on shorter prompts and about even on long ones. Two caveats apply. 3.1 Pro bills thinking tokens as output. And the two models may spend different token counts on one task. If a job doesn't need a Pro model at all, read our case for starting with 3.8 Flash.

Moving a 3.1 Pro Workflow to Gemini 4
If you already run 3.1 Pro, a Gemini 4 vs Gemini 3.1 Pro switch needs a few checks first.
- Keep the preview ID live. Gemini 4 has no public model ID, so
gemini-3.1-pro-previewstays in production until Google publishes one. - Re-test thinking settings. 3.1 Pro offers low, medium and high. Google hasn't documented Gemini 4's levels, so measure latency and cost again.
- Check custom tools. If your agent depends on the 3.1 Pro custom-tools endpoint, confirm Gemini 4 has an equivalent before you switch.
- Cap output length. With a 1M-token limit, one runaway reply can cost far more than a 64K one. Set hard limits.
- Keep a fallback. Leave 3.1 Pro wired in as a backup while Gemini 4 access widens.
Gemini 4 vs Gemini 3.1 Pro Verdict
Pick by access and workload. Test Gemini 4 for long outputs, long multi-document work and agentic coding once you can call it. Keep 3.1 Pro for production work today and for custom-tool agents. It also suits short-prompt jobs after Gemini 4's launch rate ends. Across Gemini 4 Argon vs Gemini 3.1 Pro, the newer model's edge is still mostly on paper.
Gemini 4
Pros:
- Google's new top model
- Output limit of 1M tokens
- Launch price at or below 3.1 Pro
- Strong Google-reported agent and long-context scores
Cons:
- Gated access and no public timeline
- Missing model ID and context size
- Price doubles after the launch period
- No benchmark shared with 3.1 Pro
Gemini 3.1 Pro
Pros:
- Callable today through the Gemini API
- 1M-token input with audio, video and PDF support
- Documented thinking levels
- Batch processing at half price
Cons:
- Still a preview after seven months
- 64K output cap
- Input price doubles past 200K tokens
- Training data stops at January 2025
Put Gemini 4 and Gemini 3.1 Pro Side by Side on One Recurring Task

Google's tables can't show which of the two suits your work, so settle Gemini 4 vs 3.1 Pro on a real task. MyClaw supports both models. One hosted agent, OpenClaw or Hermes Agent, can take a recurring task, such as a contract summary or a codebase audit, and run it on each. Track answer quality, retry counts and token costs over roughly two weeks. Keep 3.1 Pro as the fallback while Gemini 4 access widens. For other picks, read our OpenClaw model guide.
Three Steps to Test Gemini 4 Against 3.1 Pro on MyClaw
- Launch an agent. Open MyClaw and spin up Hermes Agent or OpenClaw.
- Add both models. Turn Gemini 4 on alongside Gemini 3.1 Pro. After that, connect your chat app.
- Compare the output. Run the same recurring task on both, then score accuracy, latency and retries.
Compare in MyClaw
Compare Gemini 4 Argon and Gemini 3.1 Pro on your own work
Give both models the same task and compare quality, speed, and cost.