<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: ViceRay</title>
    <description>The latest articles on DEV Community by ViceRay (@viceray82).</description>
    <link>https://dev.to/viceray82</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4253451%2Fc4f580fe-d1d8-431b-a165-7f708f7d9b67.png</url>
      <title>DEV Community: ViceRay</title>
      <link>https://dev.to/viceray82</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/viceray82"/>
    <language>en</language>
    <item>
      <title>Cutting LLM API costs by up to 95% without touching your code</title>
      <dc:creator>ViceRay</dc:creator>
      <pubDate>Sat, 21 Nov 2026 14:46:27 +0000</pubDate>
      <link>https://dev.to/viceray82/cutting-llm-api-costs-by-up-to-95-without-touching-your-code-45pb</link>
      <guid>https://dev.to/viceray82/cutting-llm-api-costs-by-up-to-95-without-touching-your-code-45pb</guid>
      <description>&lt;p&gt;Two months ago my side projects were bleeding money on LLM calls. Not because the projects were expensive to run — because official API pricing punishes exactly the usage pattern hobby projects have: bursty, unpredictable, and never large enough to qualify for volume tiers.&lt;/p&gt;

&lt;p&gt;So I built a relay instead of paying it. This post is the numbers, the code, and the parts that break.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Disclosure: it's mine.&lt;/strong&gt; Everything below is measured from my own account, and yes there's a link at the end.&lt;/p&gt;

&lt;h2&gt;
  
  
  The numbers that made the decision obvious
&lt;/h2&gt;

&lt;p&gt;Per 1M tokens, input / output, USD — official price first, then mine:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Model&lt;/th&gt;
&lt;th&gt;Official&lt;/th&gt;
&lt;th&gt;Here&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Grok 4.7&lt;/td&gt;
&lt;td&gt;2 / 6&lt;/td&gt;
&lt;td&gt;0.10 / 0.30&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;GPT-6 Astra&lt;/td&gt;
&lt;td&gt;10 / 50&lt;/td&gt;
&lt;td&gt;0.60 / 3.00&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;GPT-5.6 Sol&lt;/td&gt;
&lt;td&gt;4 / 20&lt;/td&gt;
&lt;td&gt;0.24 / 1.20&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;GLM-5.3&lt;/td&gt;
&lt;td&gt;1.4 / 4.4&lt;/td&gt;
&lt;td&gt;0.105 / 0.33&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Gemini 4 Argon&lt;/td&gt;
&lt;td&gt;4 / 20&lt;/td&gt;
&lt;td&gt;0.48 / 2.40&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;DeepSeek V4 Pro&lt;/td&gt;
&lt;td&gt;1.32 / 3.96&lt;/td&gt;
&lt;td&gt;0.198 / 0.594&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Claude Sonnet 5&lt;/td&gt;
&lt;td&gt;2 / 10&lt;/td&gt;
&lt;td&gt;0.945 / 4.72&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;All of them answer to the same key and drain the same balance.&lt;/p&gt;

&lt;h2&gt;
  
  
  The design choice that actually matters: 1:1
&lt;/h2&gt;

&lt;p&gt;Most relays in this corner of the market price in points or apply a multiplier on top of a token cost. You end up doing mental arithmetic to find out what a dollar bought you.&lt;/p&gt;

&lt;p&gt;This one is 1:1. Pay $1, hold $1 of balance. Usage is deducted at the model's rate. No top-up bonus, no multiplier, no minimum tier.&lt;/p&gt;

&lt;p&gt;That property is worth more than any single discount, because it makes cost predictable: the price table is the bill.&lt;/p&gt;

&lt;h2&gt;
  
  
  Calling it is a one-line change
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="k"&gt;from&lt;/span&gt; &lt;span class="n"&gt;openai&lt;/span&gt; &lt;span class="k"&gt;import&lt;/span&gt; &lt;span class="nc"&gt;OpenAI&lt;/span&gt;

&lt;span class="n"&gt;client&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nc"&gt;OpenAI&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;
    &lt;span class="n"&gt;api_key&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="s2"&gt;"sk-..."&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="n"&gt;base_url&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="s2"&gt;"https://api.daluobo.cc.cd/v1"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
&lt;span class="p"&gt;)&lt;/span&gt;

&lt;span class="n"&gt;resp&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;client&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;chat&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;completions&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;create&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;
    &lt;span class="n"&gt;model&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="s2"&gt;"gpt-6.2-astra"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="n"&gt;messages&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="s2"&gt;"role"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"user"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"content"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"explain etag caching briefly"&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
&lt;span class="p"&gt;)&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Curl works the same way. So do Cherry Studio, Lobe Chat, Cline and Continue — they all just take a &lt;code&gt;base_url&lt;/code&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where it breaks, honestly
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;The &lt;code&gt;/v1&lt;/code&gt; suffix is not optional.&lt;/strong&gt; Drop it and you get an HTML page back instead of JSON, which surfaces in most clients as a bare &lt;code&gt;request failed&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Model IDs are relay-side names, not vendor names.&lt;/strong&gt; &lt;code&gt;gpt-6.2-astra&lt;/code&gt; may not resolve on the official API, and vice versa. Keep a per-project constant instead of string literals scattered through the codebase.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Streaming is fine, exotic vendor extensions are not.&lt;/strong&gt; Chat Completions with SSE works. Provider-specific params, proprietary tool formats and some multimodal shapes depend on whichever upstream route you land on that minute.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;It is not the official API and does not pretend to be.&lt;/strong&gt; No SLA. For anything revenue-bearing, keep a second provider configured and a way to switch.&lt;/p&gt;

&lt;h2&gt;
  
  
  The uncensored models are the sleeper feature
&lt;/h2&gt;

&lt;p&gt;Two models — &lt;code&gt;deepseek-v4.1-flash-uncensored&lt;/code&gt; and &lt;code&gt;-fast&lt;/code&gt; — run with the safety layer removed, at roughly half of official DeepSeek pricing.&lt;/p&gt;

&lt;p&gt;For fiction, roleplay and long-form writing this is a different product, not a discounted one. No preamble, no hedging, no refusal when a scene involves anything commercially licensed.&lt;/p&gt;

&lt;h2&gt;
  
  
  When you should not bother
&lt;/h2&gt;

&lt;p&gt;If your monthly spend is under a few dollars, the official free tiers beat any relay on convenience, and you skip the trust question entirely.&lt;/p&gt;

&lt;p&gt;If your workload is heavily multimodal or depends on provider-specific APIs, a relay adds a translation layer you will eventually fight.&lt;/p&gt;

&lt;p&gt;Everyone else: the arithmetic is not close.&lt;/p&gt;

&lt;p&gt;Base URL &lt;code&gt;https://api.daluobo.cc.cd/v1&lt;/code&gt;. The price list is public at &lt;a href="https://daluobo.cc.cd/pricing" rel="noopener noreferrer"&gt;daluobo.cc.cd/pricing&lt;/a&gt;, and sign-up drops $0.50 of credit on the account so you can measure it before paying anything.&lt;/p&gt;

</description>
      <category>api</category>
      <category>ai</category>
      <category>webdev</category>
      <category>productivity</category>
    </item>
    <item>
      <title>What is actually inside a "GTA VI PC port" — and what breaks in it</title>
      <dc:creator>ViceRay</dc:creator>
      <pubDate>Thu, 19 Nov 2026 09:19:36 +0000</pubDate>
      <link>https://dev.to/viceray82/what-is-actually-inside-a-gta-vi-pc-port-and-what-breaks-in-it-1dc7</link>
      <guid>https://dev.to/viceray82/what-is-actually-inside-a-gta-vi-pc-port-and-what-breaks-in-it-1dc7</guid>
      <description>&lt;p&gt;Rockstar has not announced a PC version of Grand Theft Auto VI. The console release is out, no PC date exists, and by the company's own history the gap runs from 7 months (GTA IV) to 19 months (GTA V). So a group of reverse engineers did the thing Rockstar has not: they decompiled the console package and recompiled it for Win32.&lt;/p&gt;

&lt;p&gt;That is worth looking at from an engineering angle even if you do not care about the game, because the approach is not emulation and the difference is measurable.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why not an emulator
&lt;/h2&gt;

&lt;p&gt;An emulator translates guest CPU instructions and rebuilds shaders at runtime. The cost is well known: stutter on traversal while new pipeline state compiles, and a permanent instruction-translation tax on the CPU.&lt;/p&gt;

&lt;p&gt;This build skips both. The renderer is re-backed to D3D12 with a Vulkan alternative, the platform layer is rewritten against Win32, and the shader pipeline ships pre-compiled. First launch is roughly 40 seconds; frametimes are flat because nothing is being compiled during play.&lt;/p&gt;

&lt;h2&gt;
  
  
  Measured behaviour
&lt;/h2&gt;

&lt;p&gt;Build 1.0.4 on an RTX 4070, 1440p high: 78–92 fps in the city, 110+ in the Everglades. Raw input mouse with no acceleration, correct FOV to 32:9, HDR calibration, DLSS 4 and FSR 3.1 both wired up.&lt;/p&gt;

&lt;h2&gt;
  
  
  What breaks
&lt;/h2&gt;

&lt;p&gt;The honest list, because this is where most write-ups get vague:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;No GTA Online.&lt;/strong&gt; Story mode only. The console online layer talks to Rockstar services the recompile does not speak.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;No Rockstar Launcher.&lt;/strong&gt; No cloud saves, no Social Club achievements.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;AMD reflections.&lt;/strong&gt; On 1.0.4 the canal water reflects from the wrong angle on RDNA cards. The Vulkan back-end is unaffected; the fix is queued.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;8 GB VRAM&lt;/strong&gt; means medium textures. 12 GB is the realistic floor for 1440p.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Supply chain
&lt;/h2&gt;

&lt;p&gt;134.7 GB across 14 parts, sha256 published per part, signed launcher, public manifest. You can verify before unpacking. Public mirrors are open but rate-limited to 40 MB/s; a licence key lifts the cap and opens a priority mirror.&lt;/p&gt;

&lt;p&gt;It is paid — one-time USDT, $17 for the base build, no subscription and no account. Delivery is by e-mail in about six minutes, which is a strange fit for a project maintained by a group of hobbyists rather than a studio.&lt;/p&gt;

&lt;p&gt;Manifest, requirements and the changelog are public: &lt;a href="https://gta6pc.dpdns.org" rel="noopener noreferrer"&gt;gta6pc.dpdns.org&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Not affiliated with Rockstar or Take-Two. If you have run it on RDNA3, I would like the water-reflection repro — the maintainers are collecting them.&lt;/p&gt;

</description>
      <category>gamedev</category>
      <category>performance</category>
      <category>gta</category>
      <category>showdev</category>
    </item>
  </channel>
</rss>
