Time… Here We Go Again
Sometimes I forget: time is the most valuable resource, but it’s hard on humans. You need rest and sleep to be able to think and perform correctly. All these entertainment systems are killing humanity; they all want to bring people to a passive state—have more money and stay in comfort with no change; everyone is drifting toward that. To not be distracted is a real challenge for many these days. What do you want your time to be spent on? On what do you truly want to spend your time? Those who choose to spend their time only on pleasure and leisure every day have folded into the pattern.
At some point, do you sacrifice sleep for time? Or do you sacrifice time for better sleep and rest, so you can do a better job?
I’m getting older each day, hour, and minute, and this body is getting older; I only have a set amount of time to achieve the things I want to achieve.
All the above points only apply to humans, not the current generation of AI models.
DeepSeek V4 Flash & Benchmarks
I was using DeepSeek models last month, wasn’t I? After I wrote that, there was an update; they updated the flash model (DeepSeek-V4-Flash) to have more accuracy:
| Benchmark / Evaluation | DeepSeek-V4-Flash (Preview) | DeepSeek-V4-Flash (0731 Update) | GPT-5.6 Luna (High Reasoning) | DeepSeek-V4-Pro (Flagship) |
|---|---|---|---|---|
| Terminal-Bench 2.1 (CLI agent & tool recovery) | 74.2 | 82.7 (+8.5) | 81.5 | 88.4 |
| SWE-bench Verified (Real GitHub issues) | 71.4% | 78.8% (+7.4%) | 76.5% | 84.2% |
| Toolathlon-Verified (Multi-step agent tools) | 62.1% | 70.3% (+8.2%) | 68.9% | 75.8% |
| DSBench-Hard (Complex coding & reasoning) | 51.3% | 59.6% (+8.3%) | 58.2% | 66.0% |
| Artificial Analysis Index | 46 | 52 (+6) | 54 | 67 |
| Inference Speed | ~130 t/s | ~135 t/s | ~120 t/s | ~45 t/s |
| Architecture / Context Window | 13B active / 1M | 13B active (284B MoE) / 1M | Dense / 1.05M | 37B active (671B MoE) / 1M |
The 0731 update noticeably improved agentic tool use and benchmark scores on both SWE-bench and Terminal-Bench without slowing down token generation speed (~135 tokens/second). Official documentation and model parameters are available at DeepSeek API Docs.
The Price Hike & Server Surges
Shortly after that, there was a massive surge in model usage on OpenCode and other platforms, and obviously the official servers got hit as well. I noticed the API would go down, wouldn’t respond to messages, and would cut off responses midway without completing. Then after about a week, they announced they would restructure and increase prices:
| Model & Token Tier | Previous Flat Rate (Pre-Aug 16) | New Off-Peak Rate | New Peak Rate (01:00–04:00 & 06:00–10:00 UTC) |
|---|---|---|---|
| V4-Flash Cached Input | $0.0028 / 1M | $0.007 / 1M (+150% / 2.5x) | $0.014 / 1M (+400% / 5.0x) |
| V4-Flash Uncached Input | $0.14 / 1M | $0.22 / 1M (+57% / 1.57x) | $0.44 / 1M (+214% / ~3.14x) |
| V4-Flash Output | $0.28 / 1M | $0.66 / 1M (+135% / 2.36x) | $1.32 / 1M (+371% / ~4.71x) |
| V4-Pro Cached Input | $0.003625 / 1M | $0.022 / 1M (+507% / 6.07x) | $0.044 / 1M *(+1,114% / 12.14x)* |
| V4-Pro Uncached Input | $0.27 / 1M | $0.66 / 1M (+144% / 2.44x) | $1.32 / 1M (+389% / 4.89x) |
| V4-Pro Output | $0.87 / 1M | $1.98 / 1M (+128% / 2.28x) | $3.96 / 1M (+355% / 4.55x) |
The viral >1,100% price hike headline came from the V4-Pro peak cache-hit tier (jumping from $0.003625 to $0.044—a +1,114% increase or 12.14x multiplier). For the V4-Flash model I use, peak cache hits surged by +400% (5.0x) and outputs jumped by +371% (4.71x). Either way, it’s a massive hit for heavy agentic tool loops; it’s so bad!
Testing ChatGPT & Token Consumption
At the same time, I bought a ChatGPT subscription to test it out; it was good. The 5.6 pricing was lowered, so GPT-5.6 Luna High is similar to or slightly better than DeepSeek Flash in benchmarks. I can use the cost-heavy models for planning, review the plan, and use the smaller Luna model to implement things.
I only used around 530 million tokens’ worth of Codex usage throughout the month; this was just me working on the weekends on small stuff. (Updated for the end of August: I used around 830 million tokens). I used a lot of Luna; it takes a lot of reasoning and verbosity to do the same amount of work. Since it’s cheaper, it also takes a larger number of tokens to complete the same job compared to doing it with a better model.
Replacing Disqus on the Journal Site
I’ve been working on a replacement for Disqus on the journal site. They introduced ads in the free tier for blogs, and I wanted a lightweight drop-in replacement.
The initial prototype was good, but then I overcomplicated it by using Cloudflare and Cloud Functions on Firebase. I then planned and executed only what I actually wanted: a simple replacement that loads quickly, won’t overload or overbill when traffic hits the site, and has a small memory footprint. The entire thing runs in a 12MB memory footprint with an SQLite database and a small React frontend. The secondary backup is hosted on Google, with Firebase sync from SQLite along with widget backup. It’s already running on my site—the post you are reading is using that project! It’s not open-source yet; I will open-source it next month.
Reflection
If I look back, I could have done a lot of things, but that’s just how it is for everything. You think like that for everything—doesn’t almost everybody think like that? It’s just the syndrome, I guess. But honestly, considering my age, I could have been more mindful of time… you know, giving my time to easy attention-seeking things. I didn’t publish, though I wrote for the month of August in the last two years; hopefully I can publish them in the future.
What’s Next for Me?
Getting older, birthday month again. Lots of paths I can take, and in the past I did take them as well. Getting older, parents getting older, getting married, and moving forward… yes, but waiting on some things first? SO’s family is also going to cause problems—they’re going to cause problems now or later, so why not now? ~Future me, what are you thinking?
TL;DR Marriage, doing & learning things from scratch like always, moving to a different company or choosing something I like and pursuing that… only time and present day-to-day decisions will shape it, I guess.
Future AI Predictions
We are all managing AI now. At some point in the future, AI can even do the managing part. Little by little, everything is changing, but what is emerging? A lot of things? It’s not clear yet; it’s still in the early stages, like the early days of computer systems or any new technology…
I recently heard Greg Brockman in one of his interviews during my drives (one of the founders of OpenAI) talk about how previously, during training runs—in LLM development, a training run is the large-scale execution phase where a model processes trillions of text tokens across clusters of thousands of GPUs (e.g., H100s or B200s) over weeks or months, continuously updating hundreds of billions of model weights via backpropagation and gradient descent; because these distributed runs are so stateful and complex, a subtle silent bug (like numerical underflow, tensor misalignment, or a bad hyperparameter) might only manifest as loss divergence hundreds of hours in—they would fix a bug and run it, but in some cases in the past they couldn’t find the bugs and it would take 3 weeks to locate them. Now with the help of AI, it can scan the entire codebase and find the bug within hours. This is similar to the analogy in the post I wrote back in Oct 2025, Slop Generators: it’s like early compilers that get better with each iteration, where developers can use the new compiler with new features—faster and smarter—to write new code and build the next compiler.