Issue #1 · 30 Oct 2026 · 5-minute read

Claude gets pooled budgets, Mistral Large 4 halves its price, GLM-5.3 output +157%

Prices re-checked 2026-10-30 ~13:00 CET against the OpenRouter models API (and official pages where noted). Workload math: agentcosthub.com/methodology.

🧮 What will your agent cost next month? Plug in steps, tokens and models and see the bill across 80+ LLMs in seconds. Free, no sign-up. → https://agentcosthub.com/

💸 Spend controls & pricing news

Anthropic adds pooled budgets and smart reports to Claude Enterprise (3 minute read) Claude Enterprise admins can now set a shared monthly budget per team on top of per-user caps, with alerts at 50/75/95/100%. New reports break cost down per session and by output type. For teams, this finally shows which Claude workflows burn money, and stops one heavy user from blocking everyone else.

Mistral Large 4 goes open weight at half the price (2 minute read) Mistral released Large 4’s weights and cut API pricing to $0.70 input / $2.10 output per 1M tokens (cache reads $0.07), down from $1.36 / $4.18. On our coding-agent workload (50 tasks/day, 15 steps) that’s about $386/month, roughly a third of GPT-6.1 Sol. Open weights also make self-hosting an option at scale.

Snowflake Cortex AI Gateway adds per-user quotas and model allow-lists (4 minute read) Snowflake’s gateway can now block a user within minutes of hitting a quota and restrict which models each role may call. An allow-list is the simplest cap on worst-case cost per request: if nobody can call the $25/M model, nobody gets a surprise bill from it.

⚡ Quick links

• Salesforce Agent Fabric gets hard spend limits and a kill switch (2 min) — caps stop requests before overspend; Salesforce claims up to 75% token savings from caching/compression.
• Copilot Studio per-agent spending policies (MC1530368) (1 min) — Copilot Credit policies per environment or agent; GA by 20 Nov.
• BetterCloud tracks AI spend per tool and employee (2 min) — one view to find and reclaim unused AI seats.
• Qwen quietly cuts cache and input prices (1 min) — Qwen3.6 Plus cache reads $0.050 → $0.028/M (−45%); Qwen3.5 Plus and Qwen-Plus input $0.40 → $0.12/M (−71%).
• Kimi K3 reprices on OpenRouter — input $0.239 → $0.192/M (−20%) but output $10.80 → $15/M (+39%); output-heavy jobs get pricier.
• DeepSeek V4.1 Flash: same model, 2× the price depending on provider — DeepSeek direct is $0.15 / $0.60 per 1M tokens; nine OpenRouter providers (Together, Fireworks, Nebius and others) charge $0.30 / $1.20. Pin your provider.
• 📬 Forward this to whoever pays your AI bill — or subscribe free for weekly LLM & agent pricing changes. → https://agentcosthub.com/newsletter/

📈 Price increases (check before your bill arrives)

• GLM-5.3 on OpenRouter: output $1.40 → $3.60/M (+157%), input $0.07 → $0.08/M. Z.ai’s official list price is still $1.40 / $4.40 (checked 30 Oct), so the cheap OpenRouter listing is closing the gap.
• DeepSeek V4 Pro 0424 (high) on BenchLeader’s feed: input $0.44 → $1.67/M (+280%), output $0.87 → $3.34/M (+284%).
• Kimi K3 output +39% (see above).

Corrections: GLM-5.3-flash’s OpenRouter listing now matches the official $0.15 / $0.50 we flagged last time. Mistral Large 4 confirmed at $0.70 / $2.10.

💡 Cost tip of the week: stop paying full price for the same words

Tokens are the chunks of text (about ¾ of a word each) that AI providers bill you for. An agent works in steps, and at every step it re-sends everything: its instructions, the list of tools it can use, and the whole conversation so far. Most of that text is identical each time.

Prompt caching lets the provider remember that repeated beginning and charge you about 10% of the normal price for it.

Example at $2 per million input tokens: each step sends 10,000 tokens, 8,000 of them unchanged. Without caching a step costs $0.02; with caching about $0.006. At 15 steps × 50 tasks a day, that’s $450/month vs. about $126.

How to turn it on:

• Anthropic: add cache_control to your request (docs).
• OpenAI: automatic for prompts of 1,024+ tokens (docs).

Keep the unchanging part first; a date or user ID at the top breaks the cache.

🛠️ Three tool picks

• Langfuse — open-source LLM observability; see cost per trace. Hobby free (50k units/mo), Core $29/mo. → https://agentcosthub.com/tools/langfuse/
• Firecrawl — web data for agents in clean markdown (fewer junk tokens). Free 1k credits; Hobby $19/mo. → https://agentcosthub.com/tools/firecrawl/ (affiliate link on that page)
• Make — no-code automation; Free 1,000 credits/mo, Core from $9/mo (billed annually). → https://agentcosthub.com/tools/make/ (affiliate link on that page)

Full directory of 25 tools, filterable by free tier and price: https://agentcosthub.com/directory/

Forward this to whoever pays your AI bill. Got it from a friend? Subscribe: https://agentcosthub.com/newsletter/

Read on the web: https://agentcosthub.com/newsletter/2026-10-30/ · Methodology · Affiliate disclosure

Some links lead to pages with affiliate links; we may earn a commission at no cost to you. It never affects prices shown. https://agentcosthub.com/disclosure/