Developers thought AI made them 20% faster.
Measured completion time was 19% slower.
That difference runs through the current wave of AI tools. Models are getting smaller, local inference can be cheaper, and agents can coordinate across team systems. But more activity is not the same as completed work.
Feeling Faster, Finishing Slower
A July 2025 MER control trial followed 16 developers completing 246 real tasks from their normal work.
Almost everyone believed AI made them 20% faster. Measured completion time was 19% slower.
That is a roughly 40 point gap between perception and measurement.
The trial found the gap, but not its cause. It did not identify which checking, correction, or interaction added the delay. The careful conclusion is narrower: in this test, AI assistance felt productive while full tasks took longer.
@NateBJones claimed that agents generate work for humans instead of taking work off their plates. That claim came without measurements, so it should not be treated like the MER result.
The practical question is simple: are teams measuring activity or completed work?
A 27B Model Drew Fast Attention
Alibaba released 51 Quen models and variations in 2026, ranging from under 1 billion parameters to over 70 billion.
Much of the attention went to Quen 3.8 27B. One speaker placed it near GBT 5.6 Luna and Opus 4.6 in the benchmark shown.
In that Code Arena ranking:
- Quen 3.8 27B: ninth
- Gemma 4 31B: 80th
This was one speaker’s comparison. It does not prove that Quen 3.8 27B replaces every cloud model.
The community produced more than 800 quantization variations within days. Versions from Unsloth, LM Studio Community, and GGML ranged from 32 GB to 8 GB.
Quen models also surpassed two billion downloads in 2026 so far.
Quantization stores the model’s numbers with less precision. That makes it smaller and easier to fit into GPU memory, but its answers can move further from the full-precision version.
One Q3 KM version was about 12.5 GB, with KO divergence around 0.07. The brief also reported “Obliterated” and uncensored versions intended to remove safeguards.
The Local Speed Result Was Not a Clean Comparison
A dedicated GPU ran a Q3 version of Quen 3.8 27B at around 41 tokens per second. A DGX Spark ran the full-precision model at about four tokens per second.
That was not a clean hardware comparison.
Memory bandwidth:
- Dedicated GPU: 736 GB per second
- DGX Spark: 237 GB per second
The models also differed in size:
- Full precision: about 55 GB
- Q3: about 13 GB
Making the model roughly 4.2 times smaller and keeping it directly in GPU memory produced about 10 times the throughput in this setup.
Home Electricity Looked Cheaper, but Hardware Was Excluded
At 41 tokens per second, one million tokens took about 6.7 hours.
Using a 250 watt GPU and Michigan electricity at around 21 cents per kilowatt-hour, the speaker estimated home electricity at about 37 cents per million output tokens.
Core Reef FP8 was listed at $3 per million output tokens and about 17 tokens per second.
Home electricity was almost eight times cheaper by the speaker’s estimate. But that calculation excluded the cost of buying the hardware. It compared electricity with the stated service price.
Xiaomi’s AI cube also ran locally, according to @Cloud-Codes, with no cloud behind it and everything running on the box. No specifications were provided, so its performance cannot be judged.
Agents Are Becoming Shared Team Systems
DevRev’s Computer lets multiple people work in the same conversation and see the same chat. It can show where information came from, explain what it knows, and retain useful context from people and teams.
It connects to Granola, Google Drive, Slack, and Notion. The agent can work around shared applications, conversations, and deadlines instead of staying inside one person’s private chat.
Codex CLI 0.150.0 added task-to-task controls. Agents can read, create, message, wait, fork, or interrupt other Codex tasks.
Codex CLI 0.150.1 then made retained images count against the token budget during remote compaction.
Coordination is expanding along with the amount of state these systems must track.
Verification Is Becoming Part of the Workflow
Cursor engineer Lauren Tan’s reported workflow uses browsers, traces, and simulators to check agent output. Failures are turned into reusable skills.
The post credited that workflow with about 1,000 PRs per month.
Other reports described:
- A 20-person GTM company using Claude Code for about 80% of execution after connecting SOPs, Slack, HubSpot, GitHub, wikis, and skills
- A pipeline that sends Claude Code a raw video clip and returns a finished edit
- A trading agent that records why strategies succeeded or failed, then reads those lessons before its next build
- A setup routing Claude Code, Codex, and Cursor through Grok Bot in about 4 minutes
A 28-minute tutorial covered CLAUDE.md files, memory shortcuts, and parallel sessions. A 10-minute talk from the head of Claude Code focused on graphs, loops, and agents building other agents.
These were reported workflows and tutorials, not controlled comparisons.
Competition is also moving beyond the model. A community post described xAI’s “Grok Build” as a free Apache-2.0 Rust rewrite of the open-source Codex CLI.
User reactions focused on PR review, memory use, coordination, terminal interfaces, and interface friction. Those reactions compared the system around the model, not benchmark results.
Reliability Failures Can Put Work in the Wrong Place
Installing Codex or ChatGPT on the Windows D: drive reportedly broke CLI detection after an update. Moving the app to C: fixed it.
A Cursor worktree bug on macOS 3.17.19 reportedly forked a running agent’s chat instead of transferring it. That produced duplicate parents and subagents on the same checkout.
Cursor CVE-2026-63093, involving code execution from an untrusted repository, was reported fixed.
These tools edit files and run commands. Their failures can duplicate work or execute code in the wrong place.
As agents coordinate with other agents, are teams counting completed work, and are verification and task isolation keeping pace?