--- title: Agent tokens break OpenRouter's pricing game, leaked AWS... description: Today's big theme is that AI agents behave differently enough from humans to break assumptions baked into existing infrastructure. Agent token usage is up... canonical: https://daily.dev/posts/agent-tokens-break-openrouter-s-pricing-game-leaked-aws-keys-keep-working-for-years-lzrcw4aei twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Agent tokens break OpenRouter's pricing game, leaked AWS keys keep working for years | daily.dev og:description: Today's big theme is that AI agents behave differently enough from humans to break assumptions baked into existing infrastructure. Agent token usage is up... og:url: https://daily.dev/posts/agent-tokens-break-openrouter-s-pricing-game-leaked-aws-keys-keep-working-for-years-lzrcw4aei og:image: https://api.daily.dev/og/posts/lZrCW4Aei.png og:image:alt: Agent tokens break OpenRouter's pricing game, leaked AWS keys keep working for years og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Agent tokens break OpenRouter's pricing game, leaked AWS keys keep working for years **[Backend Digest](https://daily.dev/sources/backend_digest)** · 4 min read · 7 upvotes · 1 comments ## Summary Today's big theme is that AI agents behave differently enough from humans to break assumptions baked into existing infrastructure. Agent token usage is up roughly 14x since February and locks into a single provider mid-task, undermining OpenRouter's price-shopping model. Separately, a Harvard/Stanford paper shows LLM-based retrieval barely beats embeddings at over 1,400x the cost. And the AWS leaked-keys story from yesterday got worse: Corey Quinn points out AWS's quarantine policy for compromised keys still leaves real room for damage. ## Content **TLDR:** Today's big theme is that AI agents behave differently enough from humans to break assumptions baked into existing infrastructure. Agent token usage is up roughly 14x since February and locks into a single provider mid-task, undermining OpenRouter's price-shopping model. Separately, a Harvard/Stanford paper shows LLM-based retrieval barely beats embeddings at over 1,400x the cost. And the AWS leaked-keys story from yesterday got worse: Corey Quinn points out AWS's quarantine policy for compromised keys still leaves real room for damage. --- ## Agent traffic locks into a single LLM provider mid-task, weakening OpenRouter's price arbitrage Agents now consume tokens at roughly 5x the human rate, and agent usage on OpenRouter is up about 14x since February. Because agents repeatedly resend the same long context block, over 85% of their tokens are cached prompt reuses rather than fresh ones, and that cache only lives on whichever provider is currently holding the warm prefix. Switch providers mid-task and you eat the full prefill cost again, which effectively locks an agent to its starting provider until the task finishes. That's a structural problem for any router whose whole pitch is playing providers off each other on price. [Read more](https://daily.dev/posts/nNTZrxT49) ## Harvard/Stanford paper says embeddings still beat LLM retrieval on cost, not much on accuracy The Embedder's Dilemma paper ran 10 LLMs against 26 embedding models across 37 tasks and found Gemini 3.1 Pro (77.6) and the embedding model Octen-8B (77.2) were statistically tied. The catch: running the LLM benchmark cost $154.14 versus $0.11 for Octen-8B, a 1,431x difference for basically the same score. LLMs did pull ahead specifically on retrieval tasks requiring cross-document reasoning, so the real takeaway is task-specific routing, not a blanket verdict either way. [Read more](https://daily.dev/posts/OtuAmI6q9) ## AWS quarantine on leaked keys doesn't stop lateral movement or log deletion Following up on the 9,300+ exposed AWS keys Truffle Security found still active, Corey Quinn flags that AWS's automatic quarantine policy for detected leaked keys is leaker than it sounds. A quarantined key can still assume other IAM roles, run commands on EC2, stop CloudTrail logging, delete audit trails, and write to S3 buckets, including ones under compliance-mode object lock, which can't be shortened or removed by anyone, including AWS support, without deleting the whole account. Quarantine is containment theater if you're relying on it as your actual safety net. [Read more](https://daily.dev/posts/HDSgnXiNv) ## PagedAttention explainer: how vLLM cut KV-cache waste from 60-80% to under 5% A deep-dive on the technique behind vLLM's memory management borrows the OS virtual-memory idea directly: fixed-size KV-cache blocks in a shared pool, with a per-request block table mapping logical to physical blocks. The original vLLM paper measured naive serving at 20-40% GPU memory utilization; PagedAttention pushes that to roughly 96% and enables 2-4x more concurrent requests per GPU. Worth reading if you're tuning inference infrastructure and want to understand why block size and copy-on-write for shared prefixes matter. [Read more](https://daily.dev/posts/lRfvUzIgR) --- ## Also notable - **LinkedIn's multi-agent code review: 63.9% overall acceptance, 100% on concurrency bugs:** Across 5,230 sampled comments in 1,727 PRs, LinkedIn's internal multi-model reviewer platform saw 100% acceptance on concurrency bug flags versus just 40.6% on security fixes. [Read more](https://daily.dev/posts/7LpKuHKzI) - **Anthropic incident: one broad token let an agent delete 200 workloads in 90 seconds:** A filter bug plus an overly broad access token let an autonomous agent wipe about 200 workloads across 20 engineers in 90 seconds, prompting Anthropic to propose replacing binary tokens with rate-limited, undo-testable 'budgets.' [Read more](https://daily.dev/posts/SRsOi1sAL) - **Microsoft's Token Ops framework claims 78% spend reduction on agent workflows:** Benchmarked against open-source agents like browser-use and MetaGPT, Microsoft's run-level token governance approach reportedly cut average spend 78% while lifting completion rate from 67% to 96% versus simple throttling. [Read more](https://daily.dev/posts/eC2Hma6fh) - **Australia's banking API mandate: 50 banks, one identical read-only contract:** All fifty CDR-accredited Australian banks re-host the same Consumer Data Standards Banking API v1.36.0 with GET-only access, and only two banks (Up and partially Commonwealth Bank) ship anything beyond the mandated baseline. [Read more](https://daily.dev/posts/bpGwLWrBx) - **UC Berkeley finds an agent scoring near-perfect on SWE-Bench without solving any tasks:** The finding surfaced alongside AWS's new open-source aws-bench for evaluating agents on live cloud tasks, underscoring why static-fixture benchmarks are losing credibility. [Read more](https://daily.dev/posts/0sHW5bgj3) ## Community discussion Top comments from developers on daily.dev. **@gabibeyo** · 0 upvotes > The through-line is ugly and useful: agents do not consume infrastructure the way humans do. > > They stick to a provider mid-task. They hold keys longer than your rotation theater expects. They turn temporary credentials into long-lived automation. > > That is not a pricing footnote. It is a behavior problem. > > If your controls assume a person shopping models and pasting keys, agents will look like abuse in your graphs and normal in your policy. Build for sticky sessions, tool fan-out, and secret reuse. Not for the human workflow you wish they had. ## Similar posts on daily.dev - [Local AI Weekly \#3: The Token Rug Pull](https://daily.dev/posts/local-ai-weekly-3-the-token-rug-pull-h4ipcuev9) · It's Foss · 1 upvotes · 0 comments - [State of AI](https://daily.dev/posts/state-of-ai-dlul2m2pu) · Hacker News · 3 upvotes · 0 comments --- Tags: [#security](https://daily.dev/tags/security), [#aws](https://daily.dev/tags/aws), [#ai-agents](https://daily.dev/tags/ai-agents), [#ai-inference](https://daily.dev/tags/ai-inference), [#openrouter](https://daily.dev/tags/openrouter) [View this post on daily.dev](https://daily.dev/posts/agent-tokens-break-openrouter-s-pricing-game-leaked-aws-keys-keep-working-for-years-lzrcw4aei) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"DiscussionForumPosting","mainEntityOfPage":"https://daily.dev/posts/agent-tokens-break-openrouter-s-pricing-game-leaked-aws-keys-keep-working-for-years-lzrcw4aei","headline":"Agent tokens break OpenRouter's pricing game, leaked AWS keys keep working for years","text":"Today's big theme is that AI agents behave differently enough from humans to break assumptions baked into existing infrastructure. Agent token usage is up roughly 14x since February and locks into a single provider mid-task, undermining OpenRouter's price-shopping model. Separately, a Harvard/Stanford paper shows LLM-based retrieval barely beats embeddings at over 1,400x the cost. And the AWS leaked-keys story from yesterday got worse: Corey Quinn points out AWS's quarantine policy for compromised keys still leaves real room for damage.","url":"https://daily.dev/posts/agent-tokens-break-openrouter-s-pricing-game-leaked-aws-keys-keep-working-for-years-lzrcw4aei","datePublished":"2026-08-23T04:18:06.197Z","dateModified":"2026-09-13T21:14:35.883Z","author":{"@type":"Organization","name":"Backend Digest","logo":"https://media.daily.dev/image/upload/s--9uCSrsAb--/f_auto,q_auto/v1773839398/logos/backend_digest?_a=BAMAMiiu0","url":"https://daily.dev/sources/backend_digest"},"interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":7},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":1}],"comment":[{"@type":"Comment","text":"The through-line is ugly and useful: agents do not consume infrastructure the way humans do.\nThey stick to a provider mid-task. They hold keys longer than your rotation theater expects. They turn temporary credentials into long-lived automation.\nThat is not a pricing footnote. It is a behavior problem.\nIf your controls assume a person shopping models and pasting keys, agents will look like abuse in your graphs and normal in your policy. Build for sticky sessions, tool fan-out, and secret reuse. Not for the human workflow you wish they had.","datePublished":"2026-08-24T10:51:24.490Z","url":"https://daily.dev/posts/lZrCW4Aei#c-UWUaEj1Up","author":{"@type":"Person","name":"Gabi Beyo","url":"https://daily.dev/gabibeyo","image":"https://lh3.googleusercontent.com/a/ACg8ocLemxROLDZ6RYrC2zzv0zJ9ezzj62VOCNwxNCdVNa_M9tqvU4JI=s96-c"}}],"isPartOf":{"@type":"WebPage","url":"https://daily.dev/sources/backend_digest","name":"Backend Digest"}} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Backend Digest","item":"https://daily.dev/sources/backend_digest"},{"@type":"ListItem","position":3,"name":"Agent tokens break OpenRouter's pricing game, leaked AWS keys keep working for years"}]} ```