Claude/readme fast models

README.md

@@ -18,6 +18,8 @@ from MCP servers (Yahoo Finance, Sharesight) and web search. It runs on **xAI (G
 1. Create a bot with @BotFather. Run `/setprivacy` and choose **Disable**, so the bot sees every
    message in a group (remove and re-add it to groups it is already in). Who may use the bot is
    controlled in Telegram/BotFather; the bot itself has no allow-list.
+   - **IMPORTANT:** if the bot has access to Sharesight, enable "Restrict bot usage" in BotFather (this
+     enables user whitelisting).
 2. Python 3.13.5 (production; GitHub CI runs the latest stable Python and Node.js as an early warning): `pip install -r requirements.txt` (plus Node.js for npx-based MCP servers such as Yahoo Finance).
 3. Export `TELEGRAM_BOT_TOKEN` and **exactly one** of `XAI_API_KEY`, `OPENROUTER_API_KEY` or `ZAI_API_KEY`.
 4. Optional: edit `mcp_servers.json` (see [MCP data tools](#mcp-data-tools)).
@@ -239,18 +241,28 @@ token size of the tool definitions. The `Starting ...` line shows the git commit
 Figures from search results and Artificial Analysis at the time of writing (October 2026), so
 treat them as rough; speed varies by provider.
 
-| Model | $ per 1M in / out | Output speed | Quality index |
-|---|---|---|---|
-| `xiaomi/mimo-v2.6-pro` | 0.43 / 0.87 | ~28-46 tok/s | 46 (top open-weight model) |
-| `xiaomi/mimo-v2.6-flash` | 0.14 / 0.28 | ~56 tok/s | 38 |
-| `z-ai/glm-5.3-flash` (default on OpenRouter) | 0.15 / 0.50 | ~50 tok/s (other hosts up to ~270) | 57 |
-| `grok-4.3` / `x-ai/grok-4.3` (default on xAI) | 1.25 / 2.50 | ~105-146 tok/s | 25 (at high reasoning) |
-| `glm-5.3-flash` (default on z.ai) | 0.15 / 0.50 (cached input 0.03) | ~50 tok/s | 57 |
-| `xiaomi/mimo-v2.5-pro` | 0.30 / 0.61 | ~29-46 tok/s | unreliable (retires 21 Oct 2026) |
-| `xiaomi/mimo-v2.5` | 0.12 / 0.24 | ~44-58 tok/s | not found (retires 21 Oct 2026) |
+| Model | $ per 1M in / out | Output speed | Quality index | Fast sibling for `FAST_MODEL` |
+|---|---|---|---|---|
+| `xiaomi/mimo-v2.6-pro` | 0.43 / 0.87 | ~28-46 tok/s | 46 (top open-weight model) | `xiaomi/mimo-v2.6-flash` |
+| `xiaomi/mimo-v2.6-flash` | 0.14 / 0.28 | ~56 tok/s | 38 | itself (a flash model) |
+| `z-ai/glm-5.3-flash` (default on OpenRouter) | 0.15 / 0.50 | ~50 tok/s (other hosts up to ~270) | 57 | itself (a flash model) |
+| `grok-4.3` / `x-ai/grok-4.3` (default on xAI) | 1.25 / 2.50 | ~105-146 tok/s | 25 (at high reasoning) | none (see below) |
+| `glm-5.3-flash` (default on z.ai) | 0.15 / 0.50 (cached input 0.03) | ~50 tok/s | 57 | itself (a flash model) |
+| `glm-5.3-flashx` (z.ai) | 0.37 / 1.25 (cached input 0.075) | not found | not found | itself (a flash model) |
+| `glm-5.3` (z.ai) | 1.40 / 4.40 (cached input 0.26) | not found | not found | `glm-5.3-flash` |
+| `glm-5.2` (z.ai) | 1.40 / 4.40 (cached input 0.26) | not found | not found | `glm-5.3-flash` |
+| `xiaomi/mimo-v2.5-pro` | 0.30 / 0.61 | ~29-46 tok/s | unreliable (retires 21 Oct 2026) | `xiaomi/mimo-v2.5` (also retiring) |
+| `xiaomi/mimo-v2.5` | 0.12 / 0.24 | ~44-58 tok/s | not found (retires 21 Oct 2026) | itself |
 
 Most of the delay is reasoning time before the first answer token, not typing speed: use
-`REASONING=low`, and `FAST_MODEL=xiaomi/mimo-v2.6-flash` so simple messages skip the big model.
+`REASONING=low`, and set `FAST_MODEL` so simple messages skip the big model. `FAST_MODEL` runs on the
+same provider as your API key:
+- **OpenRouter:** any model, for example `FAST_MODEL=xiaomi/mimo-v2.6-flash` or `z-ai/glm-5.3-flash`
+  alongside a bigger `MODEL`.
+- **z.ai:** a GLM model: for example `MODEL=glm-5.3` with `FAST_MODEL=glm-5.3-flash` (or `glm-5.3-flashx`).
+  A z.ai model missing from the price table in `lib/llm/zai.py` shows $0 cost.
+- **xAI:** a Grok model, and there is no fast one to pick: Grok 4 Fast and Grok 4.1 Fast were reportedly
+  retired on 15 May 2026 and now redirect to `grok-4.3`.
 
 ## Development
 ```

lib/llm/zai.py

@@ -16,7 +16,12 @@
 
 BASE_URL = "https://api.z.ai/api/paas/v4/"
 # USD per million tokens (input, cached input, output); a model missing here shows $0.
-PRICES = {"glm-5.3-flash": (0.15, 0.03, 0.50)}
+PRICES = {
+    "glm-5.3-flash": (0.15, 0.03, 0.50),
+    "glm-5.3-flashx": (0.37, 0.075, 1.25),
+    "glm-5.3": (1.40, 0.26, 4.40),
+    "glm-5.2": (1.40, 0.26, 4.40),
+}
 # z.ai models always think and accept only low, high or max. Low costs almost no reasoning
 # tokens, so it is the default; the bot's REASONING values map onto z.ai's.
 EFFORT = {"": "low", "low": "low", "medium": "high", "high": "high", "max": "max"}

tests/test_llm_backends.py

@@ -136,6 +136,18 @@ async def test_zai_reasoning_maps_to_its_efforts_and_none_leaves_tools_out():
     assert client.kwargs[1]["extra_body"] == {"reasoning_effort": "max"} and "tools" not in client.kwargs[1]
 
 
+def test_zai_prices_cover_the_published_glm_models():
+    from types import SimpleNamespace as N
+
+    from lib.llm.zai import PRICES
+    assert set(PRICES) == {"glm-5.3-flash", "glm-5.3-flashx", "glm-5.3", "glm-5.2"}
+    b = ZaiBackend(config.load(ZAI_ENV), FakeChatClient())
+    usage = N(prompt_tokens=1_000_000, completion_tokens=1_000_000, prompt_tokens_details=N(cached_tokens=500_000))
+    # 500k uncached in + 500k cached in + 1M out, per million tokens
+    assert b._cost(usage, "glm-5.3") == pytest.approx(0.5 * 1.40 + 0.5 * 0.26 + 4.40)
+    assert b._cost(usage, "glm-5.3-flashx") == pytest.approx(0.5 * 0.37 + 0.5 * 0.075 + 1.25)
+
+
 async def test_zai_warns_once_about_models_without_a_price(caplog):
     ZaiBackend(config.load({**ZAI_ENV, "MODEL": "glm-x", "FAST_MODEL": "glm-5.3-flash"}), FakeChatClient())
     assert [r.getMessage() for r in caplog.records if "No z.ai price" in r.getMessage()] == [