Yesterday’s dispatch was open-weight coding week — Qwen3.8-27B, GLM-5.3, DeepSeek’s clock tax. That lane still owns local harness defaults. Today’s lead is closed-stack workhorse economics: the models you actually route high-volume agent steps through, priced like infrastructure instead of trophy shelves.

Lead: half-price Flash + Sol-price Grok

Two ships from Aug 12–13 that operators should pin this week, not next month:

Primary bookmarks: Google Gemini 3.7 Flash post, xAI Grok 4.6 announcement, Grok 4.6 docs, xAI pricing, Reuters on Gemini 3.7 Flash.

# route_sketch — not a purchase order
routes = {
  "volume_agent_steps": "gemini-3.7-flash",  # cheap multi-step + tool loops
  "long_horizon_build": "grok-4.6",          # agents + visual/interactive first pass
  "local_dense": "qwen3.8-27b",              # still from yesterday
  "do_not": "pay Sol rates for every retry",
}

Bold insight: the race this week is cost per completed task with tools, not another flagship press kit. Open weights gave you a laptop default yesterday. Closed APIs just re-priced the always-on middle of the funnel.

Gemini 3.7 Flash: workhorse, not wallflower

Google calls 3.7 Flash its most intelligent workhorse yet for coding and agents — three weeks after 3.6 Flash, half the original 3.6 cost per million on the intro tier through December 31, 2026. Access paths called out in the launch stack: Gemini API / AI Studio, Android Studio, Antigravity, Gemini Enterprise Agent Platform, plus Spark for Pro/Ultra in supported regions.

Vendor-reported jumps vs 3.6 Flash (treat as claims until you reproduce):

Pitch language is pure agent ops: better obstacle handling, cleaner instruction following, more deliberate tool calls, stronger doc/workflow automation. Practical eval list is the boring one: debug a real repo issue, generate UI against your design system, multi-step tool loop with a human gate, measure retries + total dollars per done ticket.

# smoke pattern — pin model id from your console, not folklore
# google genai / vertex / antigravity path depends on your tenant
echo "task: fix flaky test + open PR description" \
  | your_agent --model gemini-3.7-flash --tools repo,tests
# log: tool_calls, tokens_in/out, wall_time, human_edits

If you are buying metal to host the other half of the stack, the Amazon boring kit still applies: RTX-class GPU for local Qwen, NVMe for multi-quant experiments, UPS so a brownout does not eat a half-written GGUF.

Grok 4.6: Sol-tier index, not Sol-tier bill

xAI shipped Grok 4.6 on August 12 with a clear focus: long-running agents and more ambitious interactive/visual first passes. Official release notes: 500K context, text+image in / text out, no text output limit, reasoning effort low|medium|high|xhigh (default high). Pricing under 200K prompt tokens: 2 / 0.50 cached / 6 per 1M; 4 / 1 / 12 once the prompt crosses 200K — and the long-context rate applies to the whole request.

On xAI’s charts, Grok 4.6 lands 61 on the Artificial Analysis Intelligence Index — matching GPT-5.6 Sol Max and sitting under Fable 5 Max at 62. Availability: xAI API (grok-4.6), default in Grok Build, Cursor (all plans), plus gateways (OpenRouter / Vercel / Cloudflare). First-week promo called out 2× included usage in Build and Cursor.

# pin + cache key or you pay cold-server full input forever
from xai_sdk import Client
from xai_sdk.chat import user
import os

client = Client(api_key=os.environ["XAI_API_KEY"])
chat = client.chat.create(model="grok-4.6")
# Responses API: set prompt_cache_key for sticky routing
chat.append(user("Turn this product brief into a working scaffold + tests"))
print(chat.sample().content)

Ops gotcha from the docs, not vibes: set prompt_cache_key (Responses) or the Chat Completions conv id header. Without sticky routing you bounce cold servers and lose cache. Long agent loops also want context compaction. For agent builders comparing harnesses, Cursor and Grok Build are the first-party seats; if you are shopping seats more broadly, some teams still A/B against Claude-oriented tooling and classic SWE references on the shelf.

Secondary: IPO-path money + product gardeners

While Flash and Grok re-priced the middle:

• Anthropic (IPO path): Bloomberg-seen materials put Q2 preliminary revenue over 11.5 billion vs ~787 million a year earlier and 4.73 billion in Q1, with positive adjusted operating income called out in the same pack — CNBC, Yahoo/Bloomberg writeup, The Information. Figures can still revise; treat as investor-doc reporting, not audited 10-Q.
• OpenAI: CFO Sarah Friar told investors enterprise revenue now exceeds consumer, with ~40 billion annualized run rate context in the same week of coverage — CNBC.
• Microsoft: merging Copilot surfaces and pruning consumer experiments (group chats, AI podcasts, Labs bits, Deep Research for consumers) with an August 18 cutover in TechCrunch’s writeup — TechCrunch.
• Calendar note for Aug 17: tracker deprecations list Google’s imagen-4.0-fast-generate-001 → newer Gemini image path — check Google deprecations before hardcoding model strings in image pipelines.

Closed labs sell seats, governance, and SLA routes. Open labs sell weights and timezone clocks. Your agent router still needs both.

Takeaway

Do not wait for another Sol/Fable launch trailer to rebalance the middle of the stack.

1. Smoke Gemini 3.7 Flash on a real multi-tool agent path while intro pricing holds through year-end.
2. Pin grok-4.6 with cache keys and measure long-horizon build quality vs Sol dollars.
3. Keep yesterday’s Qwen/GLM/DeepSeek defaults for local and off-peak bulk — do not throw them out because the closed workhorses got cheaper.
4. Read IPO-path revenue as why agent governance and enterprise routing keep getting funded, not as your personal P&L.

The spaghetti stays the same: tangled routes, real evals, fewer hardcoded model strings — and a spreadsheet that tracks dollars per merged PR, not vanity leaderboard screenshots.

Want this in your inbox every morning? Subscribe to the SpaghettiStories newsletter.

Some links may be affiliate links. If you’re buying hardware to run local models, this affiliate link helps keep the lights on.