diff --git a/CLAUDE.md b/CLAUDE.md
index 5282bcd..366e6a6 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -5,8 +5,10 @@ Static site for **ByteFuture**, served from GitHub Pages (`.nojekyll`, `CNAME` 
 Layout:
 
 - `index.html` — ByteFuture home: company intro, the hybrid-inference thesis, the two products, and a "Writings" teaser.
-- `blog/` — the **Writings** section (listing + articles). Rules below.
-- `posts.json` — content manifest for Writings (repo root; the site fetches it at `/posts.json`).
+- `src/content/writings/<lang>/<slug>.md` — the article sources (Astro content collection); the build generates `/blog/<slug>[-lang].html`. See **Authoring model — Astro** below.
+- `blog/` — the **Writings** section listing pages + article assets/covers (the listing `index*.html` are still hand-maintained static files).
+- `posts.json` (+ `posts-{zh,ja,ko}.json`) — the Writings manifest served at `/posts.json`, **generated at build** from the collection by `src/pages/posts*.json.ts` (the root `posts*.json` files are stale legacy, no longer served).
+- `sitemap.xml` — hand-maintained (no Astro sitemap integration); new articles need manual `<url>` entries.
 - `CNAME`, `.nojekyll` — GitHub Pages config.
 
 ## HARD RULE — NEVER touch analytics or ads tracking
@@ -152,7 +154,20 @@ The `blog/` folder is the **ByteFuture Writings** section. It contains:
 - `blog/<slug>.html` — one file per published article.
 - cover images and article assets, also under `blog/`.
 
-The post index is driven by `posts.json` (repo root); `blog/index.html` fetches it at `/posts.json` and renders the cards. The **HARD RULE above applies to every page under `blog/`.**
+The post index is driven by `posts.json`; `blog/index.html` fetches it at `/posts.json` and renders the cards. The **HARD RULE above applies to every page under `blog/`.**
+
+### Authoring model — Astro (current, authoritative)
+
+The site is now built with **Astro** (`astro.config.mjs`, `output: 'static'`, `format: 'file'`). Articles are authored as **Markdown in a content collection**, not as hand-written HTML. This section is authoritative and supersedes the older per-file HTML instructions further down (the `cp blog/post-template.html`, "fill every `▼ EDIT ▼` marker", "copy the redirect script / GA / hreflang / view-counter verbatim into each page" steps): those describe the pre-Astro static pages and are kept only as reference for the legacy files still on disk.
+
+What changed, and where things live now:
+
+- **One Markdown file per language per article:** `src/content/writings/<lang>/<slug>.md` where `<lang>` ∈ `en | zh | ja | ko`. Frontmatter fields (validated by `src/content.config.ts`): `slug`, `lang`, `title`, `summary`, `category`, `date` (`YYYY-MM-DD`), `cta` (URL, defaults to Token Station intro), `cover` (optional, repo-root-relative), `draft` (optional bool). The article body is Markdown/HTML below the frontmatter.
+- **URLs are generated, unchanged:** `src/pages/blog/[...slug].astro` emits `/blog/<slug>.html` for `en` and `/blog/<slug>-<lang>.html` for the others (the `-zh/-ja/-ko` suffix is derived from `lang`, so you never hand-name files).
+- **Chrome is centralized — do NOT copy it per article.** `src/layouts/WritingLayout.astro` injects, for every article automatically: the viewport meta, the i18n redirect script + `bf_lang` cookie logic, `canonical` / `og:locale` / all four `hreflang` alternates, the **GA4 tag** (`G-KQ0EX9QGK6`), the nav + language `<select>`, the footer, the **GoatCounter** view counter, and the `.prose` / mobile CSS. The HARD RULE still protects the GA and GoatCounter snippets — they now live in the layout, so never strip them there.
+- **The manifests are generated, not hand-edited:** `src/pages/posts.json.ts` (+ `posts-zh/ja/ko.json.ts`) build `/posts.json` etc. from the collection via `src/lib/posts-manifest.ts` (drops drafts, filters by `lang`, sorts newest-first). Add/rename/retire an article by adding/renaming/removing its Markdown files — the manifests follow. The stale root `posts*.json` files are legacy and are no longer served (this is why `scripts/sync-legacy-public.mjs` stopped copying them).
+- **Still hand-maintained static files** (copied verbatim into `public/` by `scripts/sync-legacy-public.mjs`): the listing pages `blog/index.html` + `index-{zh,ja,ko}.html`, the home `index*.html`, and **`sitemap.xml`**. Consequences that still bind (see the rules below): the **category closed set** is enforced in those listing files, and **`sitemap.xml` entries are added by hand** — there is no `@astrojs/sitemap` integration.
+- **Build & verify:** `npm run build` (runs `sync-legacy-public.mjs` then `astro build` into `dist/`); `npm run check:legacy-links` and `npm run check:mobile -- <url>` for the checks.
 
 ### Branding (non-negotiable)
 
@@ -258,7 +273,7 @@ Root-level `posts.json` is a JSON array. Each object:
 
 Rules:
 
-- **Categories are a closed set.** They must match the filter pills in `blog/index.html` (`#cat-filter` `data-cat`) **and** the `validCats` map in its posts engine. To add a category, update all three.
+- **Categories are a closed set** (`engineering`, `product`, `research`, `tutorial`) and it's the `category` frontmatter field. This still binds under Astro: the value must match, in each `blog/index*.html` listing, the filter pills (`#cat-filter` `data-cat`), the `validCats` map, **and** the `CAT_LABELS` table (all four languages) in its posts engine. On top of that, `WritingLayout.astro` prints the raw `category` string in the article header and uses it as the Topics-link label + `?cat=` param — so an off-list value leaks as untranslated literal text (e.g. `model-launches`), can't be filtered, and its Topics link dead-ends to **All**. To add a genuinely new category, update the pills + `validCats` + `CAT_LABELS` (×4 langs) in every `index*.html`; otherwise pick one of the four.
 - **The hero is automatic: the newest post by `date` always leads.** On the **All** tab the most recent post renders in the hero and is pulled out of the grid; the rest follow newest-first. There is no manual `featured` flag (the engine ignores it) — to put an article on top, give it the latest `date`.
 - Dates are real publish dates; newest sorts first.
 
diff --git a/blog/asset-sources/gpt-5-6-token-station-cover.html b/blog/asset-sources/gpt-5-6-token-station-cover.html
new file mode 100644
index 0000000..f657130
--- /dev/null
+++ b/blog/asset-sources/gpt-5-6-token-station-cover.html
@@ -0,0 +1,99 @@
+<!DOCTYPE html>
+<html>
+<head>
+<meta charset="UTF-8">
+<style>
+  @import url('https://fonts.googleapis.com/css2?family=Space+Grotesk:wght@400;500;600;700&family=JetBrains+Mono:wght@400;500;600&family=Source+Serif+4:ital,opsz,wght@1,8..60,400&display=swap');
+  * { margin:0; padding:0; box-sizing:border-box; }
+  body { width:1200px; height:630px; background:#faf9f5; font-family:'Space Grotesk',sans-serif; overflow:hidden; position:relative; }
+
+  .grid {
+    position:absolute; inset:0;
+    background-image: radial-gradient(rgba(24,24,27,0.07) 1.2px, transparent 1.2px);
+    background-size: 28px 28px;
+  }
+  .wash {
+    position:absolute; top:-220px; right:-160px; width:680px; height:680px; border-radius:50%;
+    background: radial-gradient(circle, rgba(37,99,235,0.16) 0%, rgba(13,148,136,0.08) 45%, transparent 70%);
+  }
+
+  /* one endpoint routing to three variants, upper-right */
+  .routes { position:absolute; right:70px; top:118px; width:430px; height:250px; opacity:0.98; }
+
+  .content { position:absolute; inset:0; padding:64px 72px; display:flex; flex-direction:column; }
+
+  .kicker {
+    font-family:'JetBrains Mono',monospace; font-size:17px; font-weight:600;
+    letter-spacing:0.16em; text-transform:uppercase; color:#0D9488; margin-bottom:24px;
+  }
+
+  h1 { font-size:78px; font-weight:700; letter-spacing:-0.035em; line-height:1.02; color:#18181b; max-width:720px; }
+  h1 .model { color:#2563EB; }
+  h1 .sub { display:block; font-family:'Source Serif 4',serif; font-style:italic; font-weight:400; font-size:38px; letter-spacing:-0.01em; color:#52525b; margin-top:22px; }
+
+  .bottom { margin-top:auto; display:flex; align-items:flex-end; justify-content:space-between; }
+  .chips { display:flex; gap:14px; align-items:center; }
+  .chip {
+    background:#ffffff; border:1.5px solid #e4e4e7; border-radius:50px; padding:13px 26px;
+    font-size:22px; font-weight:600; color:#18181b;
+  }
+  .chip.accent { border-color:rgba(37,99,235,0.40); color:#2563EB; background:#eff4ff; }
+
+  .brand { display:flex; align-items:center; gap:13px; }
+  .brand span { font-size:23px; font-weight:600; color:#52525b; }
+
+  /* route node graphic */
+  .node-label { font-family:'JetBrains Mono',monospace; font-size:19px; font-weight:600; fill:#18181b; }
+  .node-hub { font-family:'JetBrains Mono',monospace; font-size:17px; font-weight:600; fill:#2563EB; letter-spacing:0.04em; }
+</style>
+</head>
+<body>
+  <div class="grid"></div>
+  <div class="wash"></div>
+
+  <svg class="routes" viewBox="0 0 430 250" fill="none">
+    <!-- links from single endpoint to three routes -->
+    <path d="M60 125 C 170 125, 190 34,  300 34"  stroke="#2563EB" stroke-width="4" stroke-linecap="round" opacity="0.55"/>
+    <path d="M60 125 C 170 125, 190 125, 300 125" stroke="#2563EB" stroke-width="4" stroke-linecap="round" opacity="0.55"/>
+    <path d="M60 125 C 170 125, 190 216, 300 216" stroke="#2563EB" stroke-width="4" stroke-linecap="round" opacity="0.55"/>
+
+    <!-- endpoint hub -->
+    <circle cx="60" cy="125" r="28" fill="#2563EB" fill-opacity="0.10" stroke="#2563EB" stroke-width="2.4"/>
+    <circle cx="60" cy="125" r="8.5" fill="#2563EB"/>
+
+    <!-- three variant nodes -->
+    <circle cx="300" cy="34"  r="14" fill="none" stroke="#0D9488" stroke-width="3"/>
+    <circle cx="300" cy="125" r="14" fill="none" stroke="#0D9488" stroke-width="3"/>
+    <circle cx="300" cy="216" r="14" fill="none" stroke="#0D9488" stroke-width="3"/>
+
+    <text x="326" y="41"  class="node-label">Sol</text>
+    <text x="326" y="132" class="node-label">Terra</text>
+    <text x="326" y="223" class="node-label">Luna</text>
+  </svg>
+
+  <div class="content">
+    <div class="kicker">OpenAI &middot; GPT-5.6 &middot; Coding agents</div>
+    <h1><span class="model">GPT-5.6</span> is now supported
+      <span class="sub">Sol, Terra, and Luna from one endpoint.</span>
+    </h1>
+
+    <div class="bottom">
+      <div class="chips">
+        <div class="chip">Sol</div>
+        <div class="chip">Terra</div>
+        <div class="chip">Luna</div>
+        <div class="chip accent">on Token Station</div>
+      </div>
+      <div class="brand">
+        <svg width="40" height="40" viewBox="0 0 32 32" fill="none">
+          <rect x="2.5" y="2.5" width="27" height="27" rx="8" fill="#2563EB" fill-opacity="0.10" stroke="#2563EB" stroke-width="1.6"/>
+          <circle cx="12" cy="16" r="3.4" fill="#2563EB"/>
+          <circle cx="21.5" cy="16" r="3.4" fill="none" stroke="#0D9488" stroke-width="1.8"/>
+          <path d="M15.4 16h2.7" stroke="#0D9488" stroke-width="1.8" stroke-linecap="round"/>
+        </svg>
+        <span>ByteFuture</span>
+      </div>
+    </div>
+  </div>
+</body>
+</html>
diff --git a/blog/gpt-5-6-token-station-cover.png b/blog/gpt-5-6-token-station-cover.png
new file mode 100644
index 0000000..430a4dc
Binary files /dev/null and b/blog/gpt-5-6-token-station-cover.png differ
diff --git a/scripts/sync-legacy-public.mjs b/scripts/sync-legacy-public.mjs
index 626051d..448486b 100644
--- a/scripts/sync-legacy-public.mjs
+++ b/scripts/sync-legacy-public.mjs
@@ -9,10 +9,6 @@ const copyNames = [
   'index-zh.html',
   'index-ja.html',
   'index-ko.html',
-  'posts.json',
-  'posts-zh.json',
-  'posts-ja.json',
-  'posts-ko.json',
   'CNAME',
   '.nojekyll',
   'robots.txt',
diff --git a/sitemap.xml b/sitemap.xml
index 6a44a31..471828f 100644
--- a/sitemap.xml
+++ b/sitemap.xml
@@ -85,6 +85,26 @@
     <lastmod>2026-06-18</lastmod>
     <priority>0.6</priority>
   </url>
+  <url>
+    <loc>https://bytefuture.ai/blog/gpt-5-6-token-station.html</loc>
+    <lastmod>2026-07-13</lastmod>
+    <priority>0.7</priority>
+  </url>
+  <url>
+    <loc>https://bytefuture.ai/blog/gpt-5-6-token-station-zh.html</loc>
+    <lastmod>2026-07-13</lastmod>
+    <priority>0.6</priority>
+  </url>
+  <url>
+    <loc>https://bytefuture.ai/blog/gpt-5-6-token-station-ja.html</loc>
+    <lastmod>2026-07-13</lastmod>
+    <priority>0.6</priority>
+  </url>
+  <url>
+    <loc>https://bytefuture.ai/blog/gpt-5-6-token-station-ko.html</loc>
+    <lastmod>2026-07-13</lastmod>
+    <priority>0.6</priority>
+  </url>
   <url>
     <loc>https://bytefuture.ai/blog/glm-5-2-zai-stock-jump.html</loc>
     <lastmod>2026-06-15</lastmod>
diff --git a/src/content.config.ts b/src/content.config.ts
index 2dfe4de..80b0ab7 100644
--- a/src/content.config.ts
+++ b/src/content.config.ts
@@ -15,6 +15,7 @@ const writings = defineCollection({
     category: z.string().default('tutorial'),
     date: z.coerce.date(),
     cta: z.string().url().default('https://models.bytefuture.ai/intro.html'),
+    cover: z.string().optional(),
     draft: z.boolean().default(false),
   }),
 });
diff --git a/src/content/writings/en/gpt-5-6-token-station.md b/src/content/writings/en/gpt-5-6-token-station.md
new file mode 100644
index 0000000..ed7deef
--- /dev/null
+++ b/src/content/writings/en/gpt-5-6-token-station.md
@@ -0,0 +1,126 @@
+---
+slug: gpt-5-6-token-station
+lang: en
+title: "GPT-5.6 is now on Token Station: try it in your coding agent routes"
+summary: "Token Station now supports GPT-5.6 across OpenAI-compatible API and coding-agent routes, including Codex and Copilot surfaces. Use one endpoint to compare Sol, Terra, Luna, and Claude Fable 5 in real workflows."
+category: product
+date: 2026-07-13
+cta: https://models.bytefuture.ai/intro.html
+cover: blog/gpt-5-6-token-station-cover.png
+---
+
+GPT-5.6 is now available on Token Station.
+
+For coding-agent teams, what matters is that GPT-5.6 can be tested through the routes developers already use: a direct OpenAI-compatible API, OpenAI Codex-style workflows, and GitHub Copilot routes where the supported catalog exposes the model.
+
+That turns GPT-5.6 from a launch announcement into something you can actually route, compare, and adopt without rebuilding your agent stack.
+
+## What Token Station supports
+
+Token Station exposes the GPT-5.6 family through direct OpenAI-compatible routes:
+
+- `openai/gpt-5.6`
+- `openai/gpt-5.6-sol`
+- `openai/gpt-5.6-terra`
+- `openai/gpt-5.6-luna`
+
+And GitHub Copilot routes where the supported catalog publishes availability:
+
+- `github-copilot/gpt-5.6-sol`
+- `github-copilot/gpt-5.6-terra`
+- `github-copilot/gpt-5.6-luna`
+
+OpenAI Codex-style workflows use these same `openai/` routes, so there is no separate Codex route to set up.
+
+The result is simple: you can test the same model family from the surface that matches your workflow instead of treating each provider interface as a separate integration project.
+
+## Try GPT-5.6 from one endpoint
+
+The endpoint is the standard Token Station OpenAI-compatible API:
+
+```bash
+curl https://models.bytefuture.ai/v1/chat/completions \
+  -H "Authorization: Bearer TOKEN_STATION_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "openai/gpt-5.6-sol",
+    "messages": [
+      {"role": "user", "content": "Plan a safe refactor for a pricing module and list the tests to run."}
+    ]
+  }'
+```
+
+To try a Copilot route, use:
+
+```json
+{
+  "model": "github-copilot/gpt-5.6-luna"
+}
+```
+
+Your application still talks to the same endpoint. The route changes; the integration stays stable.
+
+## GPT-5.6 Sol, Terra, and Luna
+
+A coding agent is not one API call. A real session may include planning, repository search, patch generation, test repair, code review, and delegated subtasks. Those steps do not all need the same model tier.
+
+The three named variants give teams a useful testing ladder:
+
+- **GPT-5.6 Sol**: the flagship route for the hardest coding-agent steps.
+- **GPT-5.6 Terra**: a middle route for repeated implementation and debugging loops.
+- **GPT-5.6 Luna**: a lower-cost route for exploration, triage, and subtask fan-out.
+
+A practical routing pattern looks like this:
+
+- Use a cheaper route for exploration, triage, and repeated iteration.
+- Move to a stronger route for hard reasoning, risky patches, and final review.
+- Keep the endpoint stable so your harness, agent, and evaluation scripts do not need to change every time you compare a model.
+
+Agent workloads are uneven. A repository-wide plan, a subtle failing test, and a boilerplate file edit should not automatically use the same cost tier. Model choice becomes a route name, not a new integration project.
+
+## Pricing and cache accounting
+
+Token Station exposes GPT-5.6 pricing in the categories agent operators actually need to track: input, output, cached input, cache writes, and long-context tiers above 272K input tokens.
+
+For coding agents, cache accounting matters because repository context repeats. Agents send file summaries, diffs, test output, task state, and previous plans again and again. Prompt caching can lower repeated-context cost, but only if cache writes and cached reads are handled separately.
+
+Token Station normalizes GPT-5.6 `cache_write_tokens` usage so cache writes are charged in the cache-creation bucket without double-counting them as ordinary input tokens.
+
+A practical price frame inside Token Station:
+
+- GPT-5.6 Sol / `openai/gpt-5.6`: $5/M input, $30/M output, $0.50/M cached input, and $6.25/M cache writes up to 272K input tokens.
+- GPT-5.6 Terra: $2.50/M input and $15/M output up to 272K input tokens.
+- GPT-5.6 Luna: $1/M input and $6/M output up to 272K input tokens.
+- Claude Fable 5: $10/M input, $50/M output, $1/M cache reads, $12.50/M prompt-cache writes, and $20/M one-hour cache writes.
+
+Above 272K input tokens, GPT-5.6 uses a long-context tier: Sol doubles input and cached-input prices and moves output to $45/M; Terra moves to $5/M input and $22.50/M output; Luna moves to $2/M input and $9/M output.
+
+## How to compare GPT-5.6 with Claude Fable 5
+
+Claude Fable 5 is still a natural comparison point for long-running coding agents. It is configured in Token Station with a 1M context window and a higher $10/$50 price profile.
+
+GPT-5.6 brings a different operating shape: OpenAI-native routes, Codex and Copilot surfaces, and multiple price tiers in the same family.
+
+A simple starting point:
+
+- Reach for the GPT-5.6 family when you want OpenAI-native routes, Codex or Copilot surfaces, or a cheaper tier for high-volume steps. Pick Sol, Terra, or Luna by task, as above.
+- Reach for Claude Fable 5 when you specifically want Anthropic's long-running-agent behavior and its 1M context, and are willing to pay the higher rate.
+- Compare them inside Token Station when workflow fit matters more than model-brand loyalty.
+
+## Azure and unavailable provider catalogs
+
+The integration also includes commented Azure OpenAI GPT-5.6 preview templates and support for Azure's `/openai/v1` surface through `azure_api_version = "v1"`.
+
+Those templates are intentionally not enabled with guessed prices. Azure GPT-5.6 meters were not published at the time of integration, so operators should fill in pricing when their Azure deployment exposes the model.
+
+GMI Cloud and AWS Bedrock OpenAI catalogs were not added because their public catalogs did not list GPT-5.6 support at the time. Token Station should make model routing easier, not pretend every surface supports every model on day one.
+
+## Try GPT-5.6 through Token Station
+
+If you already use Token Station, GPT-5.6 is now another route family you can test in your coding-agent workflow.
+
+Start with the direct OpenAI route for a simple API check. Use the same `openai/` route in your Codex setup for terminal coding tasks. Try the Copilot route if your workflow depends on GitHub Copilot's supported model catalog.
+
+Token Station gives you one place to compare those routes without rewriting your agent stack for every model launch.
+
+[Try Token Station](https://models.bytefuture.ai/intro.html)
diff --git a/src/content/writings/ja/gpt-5-6-token-station.md b/src/content/writings/ja/gpt-5-6-token-station.md
new file mode 100644
index 0000000..819ce77
--- /dev/null
+++ b/src/content/writings/ja/gpt-5-6-token-station.md
@@ -0,0 +1,122 @@
+---
+slug: gpt-5-6-token-station
+lang: ja
+title: "GPT-5.6 が Token Station に登場：coding agent のルートでそのまま試す"
+summary: "Token Station は GPT-5.6 を OpenAI-compatible API、Codex、Copilot などの coding-agent ルートでサポートしました。1 つの endpoint から Sol、Terra、Luna、Claude Fable 5 を実際の workflow で比較できます。"
+category: product
+date: 2026-07-13
+cta: https://models.bytefuture.ai/intro.html
+cover: blog/gpt-5-6-token-station-cover.png
+---
+
+GPT-5.6 が Token Station で使えるようになりました。
+
+coding-agent チームにとって重要なのは、GPT-5.6 を開発者がすでに使っているルートで試せることです。直接の OpenAI-compatible API、OpenAI Codex スタイルの workflow、そして対応 catalog が公開している GitHub Copilot ルートです。
+
+つまり GPT-5.6 は単なる発表ではなく、agent stack を作り直さずに route して、比較して、採用を判断できる選択肢になります。
+
+## Token Station がサポートするルート
+
+直接の OpenAI-compatible ルート：
+
+- `openai/gpt-5.6`
+- `openai/gpt-5.6-sol`
+- `openai/gpt-5.6-terra`
+- `openai/gpt-5.6-luna`
+
+GitHub Copilot catalog で公開されているルート：
+
+- `github-copilot/gpt-5.6-sol`
+- `github-copilot/gpt-5.6-terra`
+- `github-copilot/gpt-5.6-luna`
+
+OpenAI Codex スタイルの workflow は上記と同じ `openai/` ルートを使うため、別途 Codex 専用ルートを設定する必要はありません。
+
+同じモデルファミリーを、自分の workflow に合う surface から試せます。
+
+## 1 つの endpoint から GPT-5.6 を試す
+
+```bash
+curl https://models.bytefuture.ai/v1/chat/completions \
+  -H "Authorization: Bearer TOKEN_STATION_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "openai/gpt-5.6-sol",
+    "messages": [
+      {"role": "user", "content": "Plan a safe refactor for a pricing module and list the tests to run."}
+    ]
+  }'
+```
+
+Copilot ルートなら：
+
+```json
+{
+  "model": "github-copilot/gpt-5.6-luna"
+}
+```
+
+endpoint は同じです。変えるのは route だけです。
+
+## GPT-5.6 Sol、Terra、Luna
+
+coding agent は 1 回の API call ではありません。計画、repository search、patch generation、test repair、code review、delegated subtasks などが続きます。すべての step が同じ model tier を必要とするわけではありません。
+
+3 つの named variant はチームに実用的なテストの階段を与えます。
+
+- **GPT-5.6 Sol**：最も難しい coding-agent step に使う flagship route。
+- **GPT-5.6 Terra**：implementation や debugging loop に使う middle route。
+- **GPT-5.6 Luna**：exploration、triage、subtask fan-out に使う lower-cost route。
+
+実用的な routing はこうです。
+
+- 探索、triage、反復には安い route を使う。
+- 難しい reasoning、高リスク patch、final review では強い route に上げる。
+- endpoint を安定させ、harness、agent、evaluation scripts を毎回書き換えない。
+
+agent workload は均一ではありません。全 repository の計画、微妙な失敗テスト、boilerplate の編集を同じ cost tier で動かす必要はありません。モデル選択は route name になり、新しい integration project ではなくなります。
+
+## Pricing と cache accounting
+
+Token Station は GPT-5.6 の input、output、cached input、cache writes、そして 272K input tokens を超える long-context tier を扱います。
+
+coding agent は repository context を何度も送ります。file summaries、diff、test output、task state、previous plans などです。Prompt caching はこの繰り返しコストを下げられますが、cache writes と cached reads を分けて扱う必要があります。
+
+Token Station は GPT-5.6 の `cache_write_tokens` を cache-creation bucket に正規化し、ordinary input tokens と二重計上しないようにします。
+
+実用的な価格フレーム：
+
+- GPT-5.6 Sol / `openai/gpt-5.6`: $5/M input、$30/M output、$0.50/M cached input、$6.25/M cache writes（272K input tokens まで）。
+- GPT-5.6 Terra: $2.50/M input、$15/M output（272K まで）。
+- GPT-5.6 Luna: $1/M input、$6/M output（272K まで）。
+- Claude Fable 5: $10/M input、$50/M output、$1/M cache reads、$12.50/M prompt-cache writes、$20/M one-hour cache writes。
+
+272K input tokens を超えると GPT-5.6 は long-context tier になります。Sol は input と cached-input が倍になり、output は $45/M。Terra は $5/M input と $22.50/M output、Luna は $2/M input と $9/M output です。
+
+## Claude Fable 5 との比較
+
+Claude Fable 5 は long-running coding agent の自然な比較対象です。Token Station では 1M context window と $10/$50 の高い price profile で設定されています。
+
+GPT-5.6 は別の運用形を持ちます。OpenAI-native route、Codex/Copilot surface、そして同じ family 内の複数 price tier です。
+
+まずはこう選べます。
+
+- OpenAI-native route、Codex や Copilot surface、あるいは高頻度な step に安い tier が欲しいなら GPT-5.6 ファミリー。Sol、Terra、Luna は上記のとおりタスクごとに選ぶ。
+- Anthropic の long-running-agent behavior と 1M context を重視し、高い価格を許容するなら Claude Fable 5。
+- workflow fit が model brand より重要なら Token Station 内で比較する。
+
+## Azure と未対応 catalog
+
+Azure OpenAI GPT-5.6 preview template と `azure_api_version = "v1"` による `/openai/v1` surface の support も含まれます。
+
+ただし Azure pricing は推測で有効化していません。統合時点では Azure GPT-5.6 meter が公開されていなかったため、operator が deployment 側でモデルを公開した時点で価格を埋めるべきです。
+
+GMI Cloud と AWS Bedrock OpenAI catalog は、公開 catalog に GPT-5.6 がなかったため追加していません。
+
+## Token Station で GPT-5.6 を試す
+
+Token Station を使っているなら、GPT-5.6 は coding-agent workflow に追加できる新しい route family です。
+
+direct OpenAI route で API check を行い、Codex では同じ `openai/` ルートを設定して terminal coding task を試し、GitHub Copilot catalog に依存する workflow なら Copilot route を試せます。
+
+[Token Station を試す](https://models.bytefuture.ai/intro.html)
diff --git a/src/content/writings/ko/gpt-5-6-token-station.md b/src/content/writings/ko/gpt-5-6-token-station.md
new file mode 100644
index 0000000..c35958c
--- /dev/null
+++ b/src/content/writings/ko/gpt-5-6-token-station.md
@@ -0,0 +1,122 @@
+---
+slug: gpt-5-6-token-station
+lang: ko
+title: "GPT-5.6이 Token Station에 추가되었습니다: coding agent route에서 바로 테스트하기"
+summary: "Token Station은 GPT-5.6을 OpenAI-compatible API, Codex, Copilot 같은 coding-agent route에서 지원합니다. 하나의 endpoint로 Sol, Terra, Luna, Claude Fable 5를 실제 workflow에서 비교할 수 있습니다."
+category: product
+date: 2026-07-13
+cta: https://models.bytefuture.ai/intro.html
+cover: blog/gpt-5-6-token-station-cover.png
+---
+
+GPT-5.6을 이제 Token Station에서 사용할 수 있습니다.
+
+coding-agent 팀에게 중요한 점은 GPT-5.6을 개발자가 이미 쓰는 route에서 테스트할 수 있다는 것입니다. 직접 OpenAI-compatible API, OpenAI Codex 스타일 workflow, 그리고 지원 catalog가 공개한 GitHub Copilot route입니다.
+
+즉 GPT-5.6은 단순한 출시 소식이 아니라, agent stack을 다시 만들지 않고 route하고 비교하고 도입을 판단할 수 있는 선택지가 됩니다.
+
+## Token Station이 지원하는 route
+
+직접 OpenAI-compatible route:
+
+- `openai/gpt-5.6`
+- `openai/gpt-5.6-sol`
+- `openai/gpt-5.6-terra`
+- `openai/gpt-5.6-luna`
+
+GitHub Copilot catalog가 공개한 route:
+
+- `github-copilot/gpt-5.6-sol`
+- `github-copilot/gpt-5.6-terra`
+- `github-copilot/gpt-5.6-luna`
+
+OpenAI Codex 스타일 workflow는 위의 `openai/` route를 그대로 사용하므로, 별도의 Codex route를 설정할 필요가 없습니다.
+
+같은 모델 패밀리를 workflow에 맞는 surface에서 테스트할 수 있습니다.
+
+## 하나의 endpoint로 GPT-5.6 테스트
+
+```bash
+curl https://models.bytefuture.ai/v1/chat/completions \
+  -H "Authorization: Bearer TOKEN_STATION_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "openai/gpt-5.6-sol",
+    "messages": [
+      {"role": "user", "content": "Plan a safe refactor for a pricing module and list the tests to run."}
+    ]
+  }'
+```
+
+Copilot route는 다음처럼 사용합니다.
+
+```json
+{
+  "model": "github-copilot/gpt-5.6-luna"
+}
+```
+
+endpoint는 그대로입니다. 바뀌는 것은 route입니다.
+
+## GPT-5.6 Sol, Terra, Luna
+
+coding agent는 API call 한 번이 아닙니다. planning, repository search, patch generation, test repair, code review, delegated subtasks가 이어집니다. 모든 step이 같은 model tier를 필요로 하지는 않습니다.
+
+세 가지 named variant는 팀에 실용적인 테스트 사다리를 제공합니다.
+
+- **GPT-5.6 Sol**: 가장 어려운 coding-agent step에 쓰는 flagship route.
+- **GPT-5.6 Terra**: implementation과 debugging loop에 쓰는 middle route.
+- **GPT-5.6 Luna**: exploration, triage, subtask fan-out에 쓰는 lower-cost route.
+
+실용적인 routing pattern은 이렇습니다.
+
+- exploration, triage, 반복 작업에는 더 저렴한 route를 사용합니다.
+- 어려운 reasoning, 위험한 patch, final review에는 더 강한 route로 올립니다.
+- endpoint를 안정적으로 유지해 harness, agent, evaluation script를 매번 바꾸지 않습니다.
+
+agent workload는 균일하지 않습니다. repository 전체 계획, 미묘하게 실패하는 테스트, boilerplate 파일 수정을 같은 cost tier로 처리할 필요는 없습니다. 모델 선택이 route name이 되고, 새 integration project가 아니게 됩니다.
+
+## Pricing과 cache accounting
+
+Token Station은 GPT-5.6의 input, output, cached input, cache writes, 그리고 272K input tokens를 넘는 long-context tier를 다룹니다.
+
+coding agent는 repository context를 반복해서 보냅니다. file summaries, diff, test output, task state, previous plans가 계속 등장합니다. Prompt caching은 반복 context 비용을 낮출 수 있지만, cache writes와 cached reads를 분리해서 처리해야 합니다.
+
+Token Station은 GPT-5.6 `cache_write_tokens`를 cache-creation bucket으로 정규화해 ordinary input tokens와 이중 계산되지 않게 합니다.
+
+실용적인 가격 프레임:
+
+- GPT-5.6 Sol / `openai/gpt-5.6`: $5/M input, $30/M output, $0.50/M cached input, $6.25/M cache writes up to 272K input tokens.
+- GPT-5.6 Terra: $2.50/M input, $15/M output up to 272K.
+- GPT-5.6 Luna: $1/M input, $6/M output up to 272K.
+- Claude Fable 5: $10/M input, $50/M output, $1/M cache reads, $12.50/M prompt-cache writes, $20/M one-hour cache writes.
+
+272K input tokens를 넘으면 GPT-5.6은 long-context tier를 사용합니다. Sol은 input과 cached-input이 두 배가 되고 output은 $45/M, Terra는 $5/M input과 $22.50/M output, Luna는 $2/M input과 $9/M output입니다.
+
+## Claude Fable 5와 비교하기
+
+Claude Fable 5는 long-running coding agent의 자연스러운 비교 대상입니다. Token Station에서는 1M context window와 더 높은 $10/$50 price profile로 설정되어 있습니다.
+
+GPT-5.6은 다른 운영 형태를 제공합니다. OpenAI-native route, Codex/Copilot surface, 그리고 같은 family 안의 여러 price tier입니다.
+
+시작점은 간단합니다.
+
+- OpenAI-native route, Codex나 Copilot surface, 또는 high-volume step에 더 저렴한 tier가 필요하면 GPT-5.6 패밀리를 선택합니다. Sol, Terra, Luna는 위에서 설명한 대로 작업별로 고릅니다.
+- Anthropic의 long-running-agent behavior와 1M context가 필요하고 더 높은 가격을 감수한다면 Claude Fable 5를 사용합니다.
+- workflow fit이 model brand보다 중요하면 Token Station 안에서 비교합니다.
+
+## Azure와 아직 지원하지 않는 catalog
+
+Azure OpenAI GPT-5.6 preview template과 `azure_api_version = "v1"` 기반 `/openai/v1` surface support도 포함됩니다.
+
+하지만 Azure pricing은 추측으로 활성화하지 않았습니다. 통합 시점에는 Azure GPT-5.6 meter가 공개되지 않았기 때문에, operator가 자신의 Azure deployment에서 모델을 노출할 때 pricing을 채워야 합니다.
+
+GMI Cloud와 AWS Bedrock OpenAI catalog는 당시 공개 catalog에 GPT-5.6이 없었기 때문에 추가하지 않았습니다.
+
+## Token Station에서 GPT-5.6 테스트하기
+
+이미 Token Station을 사용하고 있다면 GPT-5.6은 coding-agent workflow에 추가할 수 있는 새로운 route family입니다.
+
+direct OpenAI route로 API check를 하고, Codex에서는 같은 `openai/` route를 설정해 terminal coding task를 테스트하고, GitHub Copilot catalog에 의존하는 workflow라면 Copilot route를 시험할 수 있습니다.
+
+[Token Station 사용해 보기](https://models.bytefuture.ai/intro.html)
diff --git a/src/content/writings/zh/gpt-5-6-token-station.md b/src/content/writings/zh/gpt-5-6-token-station.md
new file mode 100644
index 0000000..4cc9abe
--- /dev/null
+++ b/src/content/writings/zh/gpt-5-6-token-station.md
@@ -0,0 +1,126 @@
+---
+slug: gpt-5-6-token-station
+lang: zh
+title: "GPT-5.6 已接入 Token Station：在你的 coding agent 路由里直接试用"
+summary: "Token Station 现在支持 GPT-5.6，覆盖 OpenAI-compatible API、Codex 和 Copilot 等 coding-agent 路由。用一个 endpoint 比较 Sol、Terra、Luna 和 Claude Fable 5 在真实工作流里的表现。"
+category: product
+date: 2026-07-13
+cta: https://models.bytefuture.ai/intro.html
+cover: blog/gpt-5-6-token-station-cover.png
+---
+
+GPT-5.6 已经接入 Token Station。
+
+对 coding-agent 团队来说，真正的价值在于 GPT-5.6 可以通过开发者已经在用的 route 来测试：直接的 OpenAI-compatible API、OpenAI Codex 风格工作流，以及 GitHub Copilot 已发布支持的 route。
+
+这让 GPT-5.6 不只是一个发布新闻，而是一个可以路由、比较、逐步采用的模型选择，并且不需要重建你的 agent stack。
+
+## Token Station 支持什么
+
+Token Station 暴露直接 OpenAI-compatible GPT-5.6 route：
+
+- `openai/gpt-5.6`
+- `openai/gpt-5.6-sol`
+- `openai/gpt-5.6-terra`
+- `openai/gpt-5.6-luna`
+
+以及 GitHub Copilot 已发布支持的 route：
+
+- `github-copilot/gpt-5.6-sol`
+- `github-copilot/gpt-5.6-terra`
+- `github-copilot/gpt-5.6-luna`
+
+OpenAI Codex 风格的工作流使用的就是上面这些 `openai/` route，所以不需要单独配置一个 Codex route。
+
+结果很简单：你可以从匹配工作流的 surface 测试同一个模型家族，而不是把每个 provider interface 都当成一个单独集成项目。
+
+## 用一个 endpoint 试 GPT-5.6
+
+Endpoint 是标准 Token Station OpenAI-compatible API：
+
+```bash
+curl https://models.bytefuture.ai/v1/chat/completions \
+  -H "Authorization: Bearer TOKEN_STATION_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "openai/gpt-5.6-sol",
+    "messages": [
+      {"role": "user", "content": "Plan a safe refactor for a pricing module and list the tests to run."}
+    ]
+  }'
+```
+
+想试 Copilot route：
+
+```json
+{
+  "model": "github-copilot/gpt-5.6-luna"
+}
+```
+
+应用仍然访问同一个 endpoint。变化的是 route，稳定的是集成层。
+
+## GPT-5.6 Sol、Terra 和 Luna
+
+Coding agent 不是一次 API 调用。真实 session 可能包括规划、搜索代码库、生成 patch、修测试、代码审查和子任务分发。这些步骤不一定需要同一个模型 tier。
+
+三个命名变体给团队一个实用测试阶梯：
+
+- **GPT-5.6 Sol**：用于最难 coding-agent 步骤的旗舰 route。
+- **GPT-5.6 Terra**：用于反复实现和 debugging loop 的中间 route。
+- **GPT-5.6 Luna**：用于探索、triage 和 subtask fan-out 的低成本 route。
+
+一个实用 routing pattern 是：
+
+- 用更便宜的 route 做探索、triage 和反复迭代。
+- 在困难推理、高风险 patch 和最终审查时切到更强 route。
+- 保持 endpoint 稳定，让 harness、agent 和评测脚本不用每次比较模型都重写。
+
+agent 工作负载并不均匀。全仓库计划、微妙的失败测试和样板文件修改，不应该默认使用同一个成本 tier。模型选择变成 route name，而不是新的集成项目。
+
+## 价格和 cache 计费
+
+Token Station 暴露 GPT-5.6 的关键价格类别：input、output、cached input、cache writes，以及 272K input tokens 以上的 long-context tier。
+
+对 coding agent 来说，cache accounting 很重要，因为 repository context 会重复出现。Agent 会反复发送文件摘要、diff、测试输出、任务状态和之前的计划。Prompt caching 可以降低重复上下文成本，但前提是 cache write 和 cached read 被分开处理。
+
+Token Station 会规范 GPT-5.6 的 `cache_write_tokens` 用法，让 cache write 进入 cache-creation bucket，而不是被重复算成普通 input tokens。
+
+Token Station 里的实用价格框架：
+
+- GPT-5.6 Sol / `openai/gpt-5.6`：$5/M input、$30/M output、$0.50/M cached input、$6.25/M cache writes，适用于 272K input tokens 以内。
+- GPT-5.6 Terra：$2.50/M input、$15/M output，适用于 272K input tokens 以内。
+- GPT-5.6 Luna：$1/M input、$6/M output，适用于 272K input tokens 以内。
+- Claude Fable 5：$10/M input、$50/M output、$1/M cache reads、$12.50/M prompt-cache writes、$20/M one-hour cache writes。
+
+超过 272K input tokens 后，GPT-5.6 使用 long-context tier：Sol 的 input 和 cached-input 价格翻倍，output 变为 $45/M；Terra 变为 $5/M input 和 $22.50/M output；Luna 变为 $2/M input 和 $9/M output。
+
+## 如何比较 GPT-5.6 和 Claude Fable 5
+
+Claude Fable 5 仍然是 long-running coding agent 的自然对比对象。它在 Token Station 中配置为 1M context window，并且价格是更高的 $10/$50 profile。
+
+GPT-5.6 的操作形态不同：OpenAI-native route、Codex 和 Copilot surface，以及同一模型家族里的多个价格 tier。
+
+一个简单起点：
+
+- 想要 OpenAI-native route、Codex 或 Copilot surface，或者给高频步骤用更便宜的 tier，就选 GPT-5.6 家族；Sol、Terra、Luna 按上面说的按任务挑。
+- 明确想要 Anthropic 的 long-running-agent 行为和 1M context，并愿意支付更高价格，就用 Claude Fable 5。
+- 当 workflow fit 比模型品牌更重要时，把它们放在 Token Station 里比较。
+
+## Azure 和暂不可用的 provider catalog
+
+这次集成也包含 Azure OpenAI GPT-5.6 preview 模板注释，并通过 `azure_api_version = "v1"` 支持 Azure 的 `/openai/v1` surface。
+
+这些模板没有用猜测价格直接启用。集成时 Azure GPT-5.6 meter 还没有发布，所以 operator 应该在自己的 Azure deployment 暴露模型后再填入价格。
+
+GMI Cloud 和 AWS Bedrock OpenAI catalog 没有添加，因为它们当时的公开 catalog 没有列出 GPT-5.6。Token Station 应该让 model routing 更容易，而不是假装每个 surface 第一天都支持每个模型。
+
+## 通过 Token Station 试 GPT-5.6
+
+如果你已经使用 Token Station，GPT-5.6 现在就是另一个可以放进 coding-agent workflow 测试的 route family。
+
+先用直接 OpenAI route 做简单 API 检查；在 Codex 里配置同一个 `openai/` route 跑终端 coding task；如果 workflow 依赖 GitHub Copilot 的 supported model catalog，就试 Copilot route。
+
+Token Station 让你在一个地方比较这些 route，而不用为每次模型发布重写 agent stack。
+
+[试用 Token Station](https://models.bytefuture.ai/intro.html)
diff --git a/src/lib/posts-manifest.ts b/src/lib/posts-manifest.ts
new file mode 100644
index 0000000..37e95bf
--- /dev/null
+++ b/src/lib/posts-manifest.ts
@@ -0,0 +1,36 @@
+import { getCollection } from 'astro:content';
+
+export type ManifestLang = 'en' | 'zh' | 'ja' | 'ko';
+
+export async function getPostsManifest(lang: ManifestLang) {
+  const entries = await getCollection('writings');
+  return entries
+    .filter((entry) => !entry.data.draft && entry.data.lang === lang)
+    .sort((a, b) => b.data.date.getTime() - a.data.date.getTime())
+    .map((entry) => {
+      const post: {
+        slug: string;
+        title: string;
+        summary: string;
+        category: string;
+        date: string;
+        cover?: string;
+      } = {
+        slug: entry.data.slug,
+        title: entry.data.title,
+        summary: entry.data.summary,
+        category: entry.data.category,
+        date: entry.data.date.toISOString().slice(0, 10),
+      };
+      if (entry.data.cover) post.cover = entry.data.cover;
+      return post;
+    });
+}
+
+export function jsonResponse(data: unknown) {
+  return new Response(JSON.stringify(data, null, 2) + '\n', {
+    headers: {
+      'Content-Type': 'application/json; charset=utf-8',
+    },
+  });
+}
diff --git a/src/pages/posts-ja.json.ts b/src/pages/posts-ja.json.ts
new file mode 100644
index 0000000..ba836b7
--- /dev/null
+++ b/src/pages/posts-ja.json.ts
@@ -0,0 +1,5 @@
+import { getPostsManifest, jsonResponse } from '../lib/posts-manifest';
+
+export async function GET() {
+  return jsonResponse(await getPostsManifest('ja'));
+}
diff --git a/src/pages/posts-ko.json.ts b/src/pages/posts-ko.json.ts
new file mode 100644
index 0000000..05ea61f
--- /dev/null
+++ b/src/pages/posts-ko.json.ts
@@ -0,0 +1,5 @@
+import { getPostsManifest, jsonResponse } from '../lib/posts-manifest';
+
+export async function GET() {
+  return jsonResponse(await getPostsManifest('ko'));
+}
diff --git a/src/pages/posts-zh.json.ts b/src/pages/posts-zh.json.ts
new file mode 100644
index 0000000..c5dbee6
--- /dev/null
+++ b/src/pages/posts-zh.json.ts
@@ -0,0 +1,5 @@
+import { getPostsManifest, jsonResponse } from '../lib/posts-manifest';
+
+export async function GET() {
+  return jsonResponse(await getPostsManifest('zh'));
+}
diff --git a/src/pages/posts.json.ts b/src/pages/posts.json.ts
new file mode 100644
index 0000000..f1473b0
--- /dev/null
+++ b/src/pages/posts.json.ts
@@ -0,0 +1,5 @@
+import { getPostsManifest, jsonResponse } from '../lib/posts-manifest';
+
+export async function GET() {
+  return jsonResponse(await getPostsManifest('en'));
+}