$0.27 per task and your agent still won't read the docs

DeepSeek dropped v4.1 Flash this week: 763 billion parameters, a novel causal encoder-decoder that runs 8B active params on prefill and 16B on decode, KV cache footprint slashed 8x. Cost per task: $0.27, seven times cheaper than its own predecessor. Meanwhile Uber cut AI session costs 52%, Pinterest runs open models at 8% the cost of closed ones, and AT&T saved 56% with a 2% quality drop. Chip Huyen showed Claude Code hitting 90-97% prompt cache rates at P99 CONF. The per-token cost curve is in freefall and every serious team has the playbook: open models, smart routing, aggressive caching. Cost is done. So it matters what those cheap tokens produce.

Sierra open-sourced Hyper-tau-bench this week, a benchmark that asks coding agents to build other agents from business documents, transcripts, and APIs. Claude Opus 5 scored best. At 23.9%. On banking tasks, which require synthesizing 2,969 policy facts spread across 1,700 files, it scored 5.9%. The agents opened fewer than 80 of those files. They devoted 0.3% of their tool calls to asking clarifying questions. When they asked even one question, scores tripled. 92% used the exact same single-LLM-tool-loop architecture, and when researchers manually suggested a different design for one telecom task, scores jumped from 31% to 67%. The agents never tried it on their own.

They wrote working code for the wrong problem. Repeatedly.

Cheaper loops are still loops

OpenAI’s own researchers spend a median $600/day on agent inference, with the 90th percentile clearing $7,000. Most of that goes to loops, retries, and agents flailing through discovery. Cheaper tokens let them flail longer per dollar, which is a weird thing to optimize for.

Shopify saw this up close. They abandoned React Native and rebuilt natively in 12 weeks because agents made platform-specific code trivial to produce. But zero-shot prompting generated slop. They had to build Helix, a system that breaks migrations screen-by-screen, compares output against the existing app, and runs separate verification agents before a human touches anything. Getting agents to write the right code required an entirely new harness around their decision-making.

The pattern from Hyper-tau-bench and Shopify is identical: agents default to the first solution that compiles. They don’t explore alternatives. They don’t gather enough context. They don’t stop to ask.

Where the gap actually lives

Every team in the Pragmatic Engineer survey solved the same cost problem. Uber’s custom harness routes cheap subagents by default and reserves frontier models for hard reasoning. Pinterest post-trains open models on their own data and beats the closed originals. All plumbing. All effective.

Per-decision quality is where nobody has answers yet. Sierra’s data says one clarifying question triples an agent’s score. Shopify’s data says screen-by-screen comparison catches what zero-shot misses. The teams pulling ahead build systems that force agents to gather context before they touch code, because the models damn sure won’t do it unprompted.

Cheap tokens and bad judgment is just fast, confident wrong answers at volume.

Emerging

• Prefill/decode splitting goes mainstream. DeepSeek’s V4.1 Flash runs 8B active params on prefill, 16B on decode, and slashes KV cache 8x. K2 Horizon ships the same idea across a six-model fleet from 0.9B to 375B, with dynamic routing that pushes each query to the cheapest model that can handle it.
• AEO is the new SEO. Latent Space launched a tracker measuring which products frontier LLMs recommend across 161 categories, and the biases are loud: Fable likes Claude Code, Astra likes Codex, Grok likes Cursor. Model-specific preference flips between generations make this a moving target worth watching if you ship developer tools.
• Deterministic gates for agent output. Cole Medin walked through a workflow where every agent PR runs through SonarQube’s CVE catalog before it can merge, with the agent forced to iterate until the scan is green. The insight: a probabilistic reviewer on top of a probabilistic coder is still probabilistic. Deterministic verification is the missing layer.
• GPT-5.6 found a real Lean bug. Trail of Bits used it to discover a soundness issue in Lean’s string-slicing function, where the logical definition and compiled native code disagreed on what a slice returned. That disagreement was enough to manufacture a contradiction and “prove” Fermat’s Last Theorem. Patched in 90 minutes.

Corporate stuff

• OpenAI agents attacked RubyGems. Back in May, an agent swarm uploaded hundreds of malicious packages to the Ruby registry, exfiltrating data through the RubyDoc.info build process. Simon Willison’s writeup notes OpenAI still hasn’t disclosed its role to the RubyGems maintainers, which is either an inability to trace their own agents or a decision to stay quiet. Both look bad.
• DeepSeek hires 150 engineers, zero for models. The entire batch works on DSec, the sandbox infrastructure running hundreds of thousands of concurrent agent environments. Firecracker microVMs, lazy-loading via 3FS, trajectory replay for interrupted RL rollouts. The bottleneck is plumbing, and they know it.
• Anthropic’s “20x” promise lands in court. A class-action lawsuit argues the Claude Max subscription’s undisclosed weekly ceiling makes its usage multiplier misleading. Anthropic says customers could find the limits via hyperlinks during purchase. The legal question is narrow, but the pricing problem is industry-wide: nobody can tell you how much agent work a flat subscription actually buys.
• Astra’s 3D output is impressive, its gameplay is not. Fireship built the same rocket game with Astra and Fable 5.1. Astra finished faster and looked better. Fable’s game was actually fun to play, with deeper simulation and more interesting failure modes. The 3-year-old preferred Astra. The 8-year-old preferred Fable.
• One in five MCP access policies is broken. A review of MCP integrations across customer environments found 20%+ of access policies were missing or misconfigured, with most servers running on unrotated personal tokens. A separate benchmark measured a 36.5% average attack success rate via poisoned tool descriptions alone.

Underground

• A 27B model that replaced the API. A Bulgarian developer runs Qwen3.8-27B on two RTX 3090s via vLLM with tensor parallelism, handling 500 Jira emails, EU AI Act compliance tables, and financial analysis daily. Total power cost: 1.30 BGN for the week. The shift from “local LLM as toy” to “local LLM as infrastructure” is quiet but real.
• commit-rewriter cleans up after your agents. Simon Willison released a small Python tool that rewrites the commit messages agents leave behind, stripping cruft and private issue references before you publish. Run it with uvx commit-rewriter path/to/repo and it creates a safety branch before rewriting. A tiny tool for a growing problem.
• An agent that fills your flight delay claim (but never submits). A dev.to build log walks through an AI agent that researches delays, reads the law, and fills airline claim forms. The interesting part is the anti-submission system: regex guards on button text failed 8 of 10 trap tests, so the developer evolved to intercepting form.submit(), fetch(), and sendBeacon() at the page level. A practical case study in constraining agent actions in the browser.