← /editions
#64

Cheaper Agents, and a Wallet to Match

// Google shipped Gemini 3.7 Flash three weeks after 3.6 at half the price, Mercury gave AI agents their own corporate cards, and Meta open-weighted a 30B local agent model — while attackers impersonate ClaudeBot to hunt for AI tool credentials.

$ today_in_10_seconds

Google shipped Gemini 3.7 Flash three weeks after 3.6 Flash, at half the price — while its Pro flagship still hasn't landed. Mercury gave AI agents their own corporate cards, and Meta open-weighted a 30B agent model that runs on a single consumer GPU. Capability is getting cheaper and autonomy is getting a budget line; agent identity is still the unsolved part.

$ lead_story

Google's Gemini 3.7 Flash targets coding and agents with a 50% introductory price cut

Google released Gemini 3.7 Flash on Thursday, just three weeks after 3.6 Flash, pitching it as its most capable workhorse model for coding and agents. The coding gains are real: 43.6% on FrontierCode 1.1 Main against 34.4% for 3.6 Flash, 65.3% on DeepSWE v1.1, and a WebDev Arena Elo of 1588 versus 1538 — and notably, the model card frames this as algorithmic work on the reasoning core rather than a new pretraining run. For agent builders the more interesting claim is behavioural: Google says the model is more disciplined at completing multi-step tasks, including planning and tool calls, and recovers better when it hits obstacles, which is the part that actually determines whether your loop burns tokens on retries. But read the pricing carefully — $0.75/$3.75 per million tokens holds through the end of 2026, then rises to $1.50/$7.50 on January 1, 2027. Meanwhile Google gave no release date for Gemini 3.5 Pro, despite it having previously been described as in partner testing.

// why_it_matters: Build your agent's unit economics on the January price, not the promo — you have four and a half months to find out whether fewer retries actually offsets a doubling.

read more →

$ also_worth_attention

  • Mercury gives AI agents their own corporate cards

    Mercury Spend lets businesses issue Agent Cards — virtual cards created by a human but usable autonomously by an agent to complete purchases end-to-end, with their own limits and their own audit trail, separate from any human on the team. Transactions outside budget or merchant rules decline at the point of sale, and the agent can't raise its own ceiling. The design choice worth copying: the agent gets a distinct identity and spend envelope rather than borrowing a human's credential, which is the same principle that should govern every other capability you delegate.

    link →
  • Meta open-weights Muse Glimmer, a 30B agent model for one GPU

    Meta Superintelligence Labs released Muse Glimmer on August 10 — a 30B model tuned for always-on local agent workflows, with weights under Apache 2.0. It's compressed to roughly 4-bit with block-level speculative decoding so it fits 24GB VRAM and answers fast enough to sit inside a real agent loop, and Meta says it's trained to resist prompt injection from untrusted content and minimise oversharing. The Apache 2.0 terms — no Llama-style user cutoffs — are what make it deployable without a legal review cycle.

    link →
  • Attackers are impersonating ClaudeBot to hunt for AI tool credentials

    Known Agents reports an active campaign impersonating AI crawlers to scan sites for vulnerabilities, targeting credential and configuration paths written by AI coding tools. Bots are around 35% of the traffic it measures across 5,000+ sites, 29% of that is AI-related, and it counts a visit as spoofed when it claims a known agent identity but fails that agent's own authentication. Robots.txt compliance sits at 98.5% — a number that only describes bots identifying themselves honestly. A user-agent string is a claim, not a credential.

    link →

$ bigger_picture

The trend behind the headlines

Three of today's four stories are about the cost of autonomy falling — cheaper frontier tokens, a free local model good enough for a real agent loop, a bank making it trivial to hand an agent a spending limit. The fourth is about what happens when the number of automated actors on the internet grows faster than our ability to tell which ones are real. Mercury's answer and the spoofing problem are the same problem viewed from opposite ends: agents need identities that are issued, scoped, and verifiable, not inherited from whoever happened to start the process. Google's discount window is a nice forcing function for cost modelling, but the harder deadline is architectural. The teams that survive the next incident will be the ones who treated agent identity as infrastructure a year before they needed to.

$ tool_of_the_day

Web Bot Auth

An authentication method that uses cryptographic signatures in HTTP messages to verify a request actually comes from the bot it claims to be, built on RFC 9421 HTTP Message Signatures. It relies on two IETF drafts: a directory for publishing crawler public keys, and a protocol for attaching that identity to requests. If you allowlist AI crawlers or expose endpoints to agents, this is the replacement for trusting a header string — note the drafts are still moving, so pin what you implement against.

explore Web Bot Auth →

$ subscribe --free

Get the next edition in your inbox

// DAIly Agent covers agentic AI three times a week

// free, no spam, exit anytime