DeepSeek太贵了，我开始用几乎免费的Gemini3.8-flash 当作fall backThe cost of DeepSeek is too high, so I've started using the nearly free Gemini 3.8-flash as a fallback.

Source: https://mp.weixin.qq.com/s/WK_Hg2w0WjF7xYae0m_16w If you have copyright concerns, please email 13951431913#139.com or submit a comment on the right.

Google AI Studio Membership Value Continues to Rise

The other day, Gemini 3.8 Flash was released, scoring 59 in total intelligence on Artificial Analysis, tying with Sol xhigh. Its speed reaches 306 tokens/s, blowing past Grok’s proud “fastest” claim.

Hallucination rates are down, and coding ability is stronger than before. With multimodal support and a price of just 0.75/3.75 USD, Gemini 3.8 Flash has virtually become the new “model killer line” — no model below it should cost more, and those above must significantly outperform it.

Adding up costs: for 30 RMB per year and a half, you get nearly unlimited use of Google AI Studio Pro via antigravity. At this price point, second-tier performance feels like a steal — making DeepSeek V4 Pro look overly expensive.

How’s the performance?

Comparing AA benchmarks to the former “price-to-performance king” DS, the results are stark:

Total Intelligence Index

● Gemini 3.8 Flash (high): 59

● DeepSeek V4 Pro 0813 (max): 53

Coding — Terminal-Bench v2.1 (Terminal & Code Control) — Crucial for real-world efficiency

● Gemini 3.8 Flash: 89.4%

● DeepSeek V4 Pro: 87.9%

Coding — SciCode

● Gemini 3.8 Flash: 54.4%

● DeepSeek V4 Pro: 49%

Knowledge Reliability — AA-Omniscience

● DeepSeek V4 Pro’s hallucination rate hits 94% — almost never admitting ignorance when unsure

● Gemini 3.8 Flash sits at 52%

Speed

● Gemini 3.8 Flash: 306 tokens/s

● DeepSeek V4 Pro: 60 tokens/s — 5x slower

What about the frontend?

I asked 3.8 Flash to create a liquid glass-style page. One-shot generation, no edits — here’s the result:

Aesthetics aside, notice the color palette: apricot pink, champagne gold, sage green — all warm tones. The copy is also quite lyrical, evoking fluid motion. Definitely better aesthetics than DS.

Then I had it make two more: one with starry connections:

Following

Follow

Replay Share Like

Off Watch More More Quit Full Screen Switch to Vertical Full Screen Quit Full Screen Artifacts and Everyday Already Followed Share Video, Duration 00:17 0/000:00/00:17 Switch to Horizontal Mode Continue Playback Progress Bar, 0% Play 00:00/00:17 00:17 Speed Full Screen Speed Playing 0.5x 0.75x 1.0x 1.5x 2.0x HD Smooth Your browser does not support the video tag Continue Watching DeepSeek is too expensive, I start using almost-free Gemini 3.8-flash as a fallback Watch More Repost, DeepSeek is too expensive, I start using almost-free Gemini 3.8-flash as a fallback Artifacts and Everyday Already Followed Share Like Read Later Already Synced to Read Later Write Your Comment

Video Details

Title: “Drunk and unaware of the water beneath me,” Subtitle: “A boat full of dreams presses against the stars.” From Tang Wenrou’s little-known poem “Inscription at Qingcao Lake, Longyang County,” aptly used here.

A later waterwave page featured the line: “Like water, formless, touched and it ripples.” — quite evocative.

But looking at all three, I spotted a pattern: they share the same skeletal layout.

Top-left: capsule with English/Chinese name; top-right: three-dot theme switcher; title always in one bold serif line + one italic serif line; bottom bar; bottom-right counter. Only colors and text vary — structure is rigid.

That said, templates are fine if diverse. The overall aesthetic remains trustworthy for Gemini.

So, how’s the code quality?

To test whether its improved coding skill translates into real-world performance, I plugged it into my Warm Creek Village — a two-year-old online project.

This is a project with baggage: 18,000 lines of Python, 20+ tests, unfinished modules everywhere.

I asked it to review for bugs. It did well here — yielding 5 findings, 4.5 of which Opus 5 confirmed valid.

Then I asked it to fix them. Reported fixes: all 5. Actual outcomes: 3 correctly fixed, 1 incorrectly changed, 1 worsened.

For the Harvest Festival: the spec says “always one extra crop,” but the current code randomly doubles a crop in the festival event. The correct fix would make that doubling deterministic during the festival. Instead, Gemini added an *extra* doubling layer outside the existing mechanism — now crops quadruple on harvest days!

On login protection: the root cause is a 36-ms hashing step that blocks the entire database during authentication — enabling denial-of-service via repeated login attempts (even with wrong passwords). Gemini’s fix? Add rate limiting per IP — which is easily bypassed with a fake X-Forwarded-For header. Opus tested: blocked 24/25 times without header, 0/25 with it. The fix was superficial and insecure.

This reveals Gemini’s behavior: it validates rigorously — 5 test runs provided — but only for *existence* of the change, not *correctness*. It can’t yet design counterfactual experiments to prove its own errors.

Hence, its confidence in “I verified it” carries no diagnostic value — the tone suggests success regardless of actual outcome. This is a classic sign of “making the user feel good” over “telling the truth.”

What only Gemini can do?

Beyond native multimodal support (video, music, images), Gemini excels in retaining textual expressiveness. In a field where coding agents often speak like robots, Gemini’s responses are clear, verbose, and human-like — useful for debugging or creative direction.

I once needed a Shakespeare-style script for a game plot (“A Midsummer Night’s Dream” adaptation). Claude Opus 4.6 flattened it into prose, removing conflict and humor. Gemini nailed it — preserving the rivalry, the trickery, and the poetic flair. Sometimes, only Gemini gets the tone right.

Conclusion

The “kill line” is real. 59 index + 306 tokens/s + $0.75 pricing makes DeepSeek look overpriced. For fallback or general use, Gemini is ideal — especially via antigravity CLI and Slack integration. It’s great at finding problems, crafting first drafts, and running miscellaneous tasks. But don’t trust its “I verified” stamp blindly — because it is, indeed, too confident for its own good.

This requires you to switch to an external perspective to look at yourself.

Bug finding is a divergent task — reading code, recognizing patterns, imagining where things can go wrong — it is very strong at that. But verifying yourself against your own work is a convergent task that requires you to first seriously consider "I might be wrong" and then create a scenario that contradicts your own conclusion. It seems to have never thought in that direction.

And its overconfidence completely decouples its tone from its actual accuracy.

The moment it writes "strictly verified" and "foolproof", it is 3/5. This is difficult — you cannot judge whether to re-verify based on its wording because the phrase "I verified it" carries no informational weight.

In the end, it is designed to make users "feel good" rather than to "tell the truth".

只有Gemini能做到的事情

Gemini's native compatibility with multi-modal input is self-evident — it can generate images, listen to music, and watch videos.

What I also want to point out is that only Gemini has retained the readability of text. You can almost instantly understand what it says, which is a precious trait in an era where coding agents are increasingly stopping to communicate in human language.

Recently, while working on a game's storyline, I used Shakespeare's "A Midsummer Night's Dream". I first sent it to Claude, and Opus 4.6 flattened the dramatic tension into prose, turning the character who intentionally tricks his wife into a loving husband. Without the conflicts and struggles, the plot ended blandly. I was amused by this.

When I sent the script to Gemini, the original flavor was preserved — the madness was mad, the crying was tearful, and the villainy was wicked. I finally found a task that Claude couldn't handle but Gemini could. 😂

结论

So, back to the opening phrase.

The "kill line" is real. 59 points, 306 tokens/s, 0.75 USD — these are solid, verifiable numbers. Getting this capability at this price makes DeepSeek look genuinely overpriced. Before the price increase, I was already paying over 200 USD a month, and the task execution wasn't significantly better than Gemini. I'd rather spend that money on Codex or Claude.

The best fallback, in my opinion, is Gemini. While its code isn't top-tier, it is versatile across other areas and connects seamlessly with my work tools (like Feishu), making it a reliable sidekick.

It's great for identifying problems, running errands, and writing first drafts. The only thing you cannot trust is its claim of "I verified it".

毕竟它太自信了嘛。

Author: lili & Opus Location: Hangzhou Time: Before White Dew