Loay @LoayWithAI

@zeddotdev going to try /delta share on a debugging session with a teammate. curious how it handles two people prompting the same thread at once, does it queue or do you get interleaved edits?

@akshdeeps_001 if subagents re-pull the main context and take twice the steps, you're paying for the same tokens over and over. capping subagents and restarting sessions early is the cheap fix while we wait on a reset

1.7B parameters, 256K context, and the model card tells you up front not to use it for agentic coding. That's LFM3 from Liquid AI. It's post-trained for local agent harnesses like Hermes Agent, Pi and OpenClaw, and aimed at tool calling and structured outputs. So it's the small model that routes, extracts and calls tools on your machine while the big coder does the diffs. GGUF and MLX builds are already up, under Liquid's own LFM Open License rather than Apache.

@DamiDefi if duplicate spans can make usage look 3x higher, i'd sum the usage field from the raw API responses as the source of truth and use traces only for the call order otherwise you're debugging against inflated totals

@shubhporwal24 how many cold start runs per build did you average for the 24%, and did you drop caches between them? the tight CI range makes me curious how the benchmark is set up

Telling someone to restart their Mac: worth a few cents. Telling a pharma company where to move a ring on a molecule: maybe tens of millions. Dario Amodei's point in the clip @guilleflorvs posted is that those tokens aren't worth the same, and he expects the industry to try paying for results, or something closer to paying for labor by the hour. For agent work, the first thing I'd want answered: under pay-for-results, who pays for a long run that fails?

Guillermo Flor @guilleflorvs

a day ago

Dario Amodei: some tokens Claude produces are worth a few cents, others could be worth tens of millions of dollars. His examples: 1. Telling someone to restart their Mac: the model has said it 10 million times, worth a dollar or less 2. Telling a pharma company to move an

380 110 3K 1.2M 2K

@TeksEdge Going to try the Q4_K_M in llama.cpp as a small agent model and see if the Luna traces help with tool call formatting. Curious if it holds up past a few thousand tokens of context

@rezoundous Not having to watch the usage meter every hour probably explains half of that feeling. Was it the limits or the output quality that pushed you over?

$0.40 per task for GPT-6.1 Terra vs $0.72 for GPT-6.1 Sol, on Artificial Analysis's index at max effort. OpenAI prices Terra at half of Sol per token. Per task it comes out about 44% cheaper, because Terra spends more tokens getting through the same tests. It scores 48 to Sol's 52. Claude Opus 5.5 hits 58 at $5.98 per task, roughly 15x Terra.

@IndependentEco Is the gap the model or the harness though? A lot of extra iterations come from a weaker agent loop and worse context handling, not the weights. Worth running the same task through the same harness before deciding.

@_philschmid Does --dry-run print the exact request the agent would send, or just validate the flags? Curious if you let the agent run it unattended for the Deep Research jobs or keep a check on those.

$1 in, $5 out per million tokens. GPT-6.1 Terra is OpenAI's new cheap model, live today in Codex and in the API as gpt-6.1-terra. In OpenAI's own numbers it scores 50.3% on Terminal-Bench 4.0 at max effort, up from 21.5% for 5.6 Terra. OpenAI still points you to Sol and Astra for the hardest coding. So Terra is for the high-volume agent and computer-use work, and your Codex allowance goes further on it.

@NaceAI 56% fewer tokens only matters if the answers stay right. Wasted tokens in file search are mostly agents reading whole files to find one page, so was the +4% measured against plain grep or the default search tool?

@trikcode Curious how often the agent picks up an agents.md from the downloads folder versus only the project root. If it only reads the repo one, a file dropped in downloads shouldn't matter unless you cd there.

Anthropic, the company behind Claude Code, is going to trade as ANT on Nasdaq. It filed its S-1 with the SEC today. The share count and price range aren't set yet. What I'm curious about as a builder: once every quarter gets judged by the market, do rate limits and pricing get more generous or tighter?

@cneuralnetwork The line items add up to $5,107.51, not $5,107.71. Per million tokens that works out to about $0.55 for Codex and $0.40 for Claude, so I would split input, cached and output tokens before comparing the two.

@jmeistrich Going to try this on an existing Expo project. Curious whether native menus and keyboard shortcuts stay in sync when the same screens also run on iOS and Android, or if that needs per-platform files.

Three models from three providers reviewing the same change until they agree. That's the skill I'd steal from @theo's tstack. He open-sourced the agent skills he uses every day for Claude Code and Codex, MIT licensed. council needs a harness that can run other providers' models, so it's not pure copy-paste. Install is a cp -R into ~/.claude/skills or ~/.codex/skills.

Theo - t3.gg @theo

2 days ago

So uh...I guess it's time to announce tstack? I caved. I am publishing my skills now.

205 71 3K 458K 2K

@unclebobmartin The thing I would watch with bots instead of CLI agents is cost. A flat subscription cap is predictable, per-token API usage on long runs usually is not.

@Voxyz_ai Do you keep that line in a global AGENTS.md or per repo? Curious if it makes the agent flag every tiny doc mismatch in the final reply, or only the ones that matter.