Claude Code agent summaries cost you tokens every 30 seconds, Anthropic confirms

Anthropic's Claude Code asks each running background agent's own model roughly every 30 seconds to summarize its last action in 3-5 words for the agents panel status line, sending the full transcript each time, and Anthropic engineer edwinarbus confirmed the practice on a 530-point Hacker News thread, saying "we agree that's a lot for a status line." Anthropic's telemetry puts the cost at roughly 2% of spend across users and 6% in the terminal, while independent developer Flamerix AI clocked it at 16.4% of one week's Claude Code spend; a fix for summaries that miss the prompt cache shipped in version 2.1.311+, with a cheaper mode still coming. No setting turns off just the summaries, and Claude Code's docs state the terminal status line runs locally without consuming API tokens, while the agents panel summaries are a separate billed call.

Claude Code agent summaries cost you tokens every 30 seconds, Anthropic confirms Anthropic's own engineer agrees it's "a lot for a status line" and says a fix for the costliest case shipped in 2.1.311. Four more stories inside. Today's lead is Claude Code's background-agent summaries, which spend your tokens whether you look at the agents panel or not. It leads because an Anthropic engineer confirmed it on Hacker News and called it intentional, and because every background agent you run is paying for it. Also today: OpenAI says ChatGPT Pro and Business subscribers get monthly API credits. On Replit's new benchmark, the best coding agents finish at most 65% of multi-feature builds with every test passing. Mistral puts Large 4's weights on Hugging Face. And Gemini 4 Carbon takes first place on Artificial Analysis's Intelligence Index, a model ranking Opus 5.5 led until today, by one point. Of the five, the first is the one already on your bill. What changed this week Claude Code agent summaries cost you tokens every 30 seconds, Anthropic confirms Claude Code asks each running background agent's own model, roughly every 30 seconds, to summarize its last action in 3-5 words for the agents panel's status line, sending the agent's full transcript each time. A gist by independent developer Flamerix AI clocked the practice at 16.4% of one week's Claude Code spend. Anthropic's Claude Code engineer edwinarbus confirmed it on the 530-point Hacker News thread, said "we agree that's a lot for a status line," and put Anthropic's own telemetry at roughly 2% of spend across users 6% in the terminal . A fix for summaries that miss the prompt cache a reuse discount , the costliest case, shipped in 2.1.311+, with a cheaper mode still coming. Claude Code's docs say its terminal status line "runs locally and does not consume API tokens"; the agents panel's summaries are a separate, billed call. No setting turns off just the summaries. OpenAI says ChatGPT Pro, Business subscribers get up to $500 in monthly API credits Eligible ChatGPT Pro subscribers can claim $100 to $500 a month in API credits, by tier; Business workspaces get $20 per paid Standard seat or $100 per Premium seat, pooled up to $500, per a community reply citing OpenAI's Help Center, unreadable to this draft's fetch. The credits build apps and agents through the API, prorated on the first grant, non-rolling after. It echoes Anthropic's own October credits for its Max and Team plans, which excluded Claude Code itself. Replit's ViBench 2.0 benchmark finds top coding agents clear only 65% of long builds ViBench 2.0 starts an agent with an empty project, has it build a product from a short spec, then add eighteen features one at a time, grading each with real-browser tests across seventeen products and 544 test plans. Best scores: OpenAI's GPT-6.1 Sol passes every test in 65% of builds at about $12 each; its GPT-6 Astra, 60% at $90; Anthropic's Opus 5.5, which also graded every run, 49%. Most failures came from features interacting, not one feature alone. Mistral releases Mistral Large 4's weights under a modified MIT license Mistral released the weights of Mistral Large 4, its one-trillion-parameter flagship, under a modified MIT license with one condition: a business selling it as a hosted service needs a commercial license; self-hosting does not. Mistral's own numbers: 19.7% on Cognition's FrontierCode 2.0 coding benchmark and 41% on Terminal-Bench 4.0, both open-weight leads, and prices cut to $0.70 in, $2.10 out per million tokens. Artificial Analysis, independently, scores it 48% on its Coding Agent Index, level with Z.ai's GLM-5.3. Gemini 4 Carbon tops Artificial Analysis's Index by a point, trails Claude Code on coding Artificial Analysis now ranks Google's Gemini 4 Carbon first on its Intelligence Index v5 at 59, one point above Claude Opus 5.5's 58, which led yesterday's issue, and second on its Coding Agent Index: 59 in Google's Antigravity CLI, three behind Claude Code with Opus 5.5 at 62. Per task, Carbon costs $7.76 to Opus 5.5's $10.04, and AA lists no public API yet. For coding agents, that's a cheaper runner-up, not a new leader. Also worth your time - GitHub Copilot Fusion reaches general availability, routing a task across multiple models instead of one https://github.blog/changelog/2026-10-28-route-and-orchestrate-models-with-github-copilot-fusion?ref=thenewway.ai - GitHub Agentic Workflows reaches general availability https://github.blog/changelog/2026-10-28-github-agentic-workflows-generally-available?ref=thenewway.ai Know someone who'd want this in their inbox? Forward it — that's how this grows. And if we got something wrong, or you think we buried the real story today, hit reply. A person reads every one. The New Way is written with AI. It gathers the day's stories, checks them against their sources and drafts every summary. A person decides what runs and reviews every issue before we hit send.