
👋 I’m @allglenn 🌍 Paris • London • US 🧑🏿‍💻 Tech builder ✍🏿 https://nitter.cf/t.co/I9Xxc3sSSp
Paris, France
Joined July 2013
- Tweets3.1K
- Following232
- Followers110
- Likes285
Fable 5.2's token price dropped 20%. Its cost per task doubled: $26.54 at max effort vs $12.99 for 5.1, since it writes ~1.7x the output tokens. Gemini 4 Carbon at a quarter of the price wins at max effort. Change the effort setting and most of the gap disappears.
1
27
Price per million tokens is the wrong number. Compare cost per finished task, on your workload, at the effort you ship. Full numbers: pub.towardsai.net/gemini-4-c…
29
Amazon's 98% AI cost cut is real but flattering. Sonnet 4.6 was writing ~170 tokens of JSON per ticket to pick 1 of 20 labels. Versus Haiku 5.5 on the same counts, the self-hosted model is only ~1.5x cheaper. The win was not using an essay writer to tick a box.
1
1
29
Bonus bug: JSON key order shifted label probabilities up to 0.19, and flipped a label in my own CPU test. Sort keys, gate on 0.9 confidence, audit every fixed-choice step. Full math: medium.com/data-science-coll…
16
7 popular MCP servers expose 112 tools: about 29,000 tokens of definitions before the agent reads your request. Roughly 15% of a 200k window. Worse: look-alike tools make the agent pick the wrong source while every call returns 200. Each new server is a small loan.
2
24
Cloudflare fit 2,500+ endpoints into 2 tools and about 1,000 tokens, so the context cost is fixable. The security reviews and tracing aren't. Full breakdown: levelup.gitconnected.com/you…
1
35
Anthropic's lead in Ramp's data is real. Its moat isn't. The 8% monthly provider churn that filled the funnel can empty it. Models are cheap to copy, Google and Microsoft own the front door, and the lead gets re-bought every quarter in compute. That's a rented lead.
1
31
Builders: make multi-model the default, keep evals and routing in your own code, and put the value in the workflow layer. Full breakdown: pub.towardsai.net/anthropics…
31
Anthropic's S-1 doesn't say Claude is losing to cheap models. It says your agent workflow rides on someone else's pricing. And "we'll just swap the endpoint" is a fake hedge: an API can accept valid requests while your agent can't finish one task.
1
32
Reported case: local Qwen behind Ollama returned tool calls as plain-text JSON, 4 times. A fallback only counts once it finishes the work. Measure cost per accepted task, not price per token.
Full breakdown: levelup.gitconnected.com/ant…
48
"Rewrite your agent stack in Rust" misses where agents break. Payment accepted, connection drops, worker times out, orchestrator retries, customer charged twice. Ownership can't know if a remote API finished before the network failed. That needs idempotency and durable state.
1
1
44
Speed? In an illustrative run of 4 model calls (~800 ms), 6 tool calls (~150 ms) and 100 ms of orchestration, deleting orchestration saves ~2.4%. Rust wins on concurrent sessions per worker in a fixed memory budget.
Full breakdown: pub.towardsai.net/why-rust-m…
30
The next agent upgrade may be the harness, not the model. Google's Harness² froze the weights and let agents rewrite prompt, guardrails, memory, tools and skills. GLM-5 went from 64.2 to 75.7 on WorkBuddy-Bench, past base Gemini at 72.4. Cheaper model, better harness, won.
1
51
The catch: no answer key, so gains are noisy, and the repo warns about open executor network access and a disabled sandbox. Self-editing harnesses need limits in code. The model is swappable. The harness is the moat.
Full breakdown: medium.com/towards-artificia…
28
Proactive agents get compared on plan price. The real bill is cost per accepted outcome. Worked example: $100 plan, 40 briefs, 32 accepted. $3.13 each on paper, about $13 each once review and fix time at $60/h goes in.
1
31
Before model quality matters, permissions decide where the work stops: Duet blocks transfers even with approval, Dots hands them back, Muse has an allow/ask/deny layer. Full comparison: pub.towardsai.net/proactive-…
1
66
Cheap model releases get read as price stories. They're time stories. At ~12 cents of tokens per SWE task, the real bill is wall clock: sandboxes, CI runners, concurrency slots. Same DeepSeek weights run 24 to 290 tok/s depending on host.
1
48
7.6 min vs 23.5 min per task. At 1,000 tasks/hour that's ~127 vs ~392 slots in flight. Compare cost per successful task, not per token. Full breakdown: medium.com/towards-artificia…
58
Langfuse, Phoenix, LangSmith all take OpenTelemetry spans over OTLP. Tracing is a config line. The lock-in is scores, datasets and prompts, because no exporter setting moves them. Pick where your eval loop lives.
2
1
48
Then pick on license, ops (six containers vs one process) and billing shape (per-unit is cheap at 4 spans per trace, not at 40). Full breakdown: pub.towardsai.net/langfuse-v…
38