Cheap Models, Expensive Token Bills

Three weeks after launching TIN, its full-text search index for Postgres, PlanetScale has shipped v1.0.4 and v1.0.6. The releases answer customer requests and a ParadeDB benchmark that showed ParadeDB 0.26 beating TIN on read-only, BM25-ranked top-10 queries.

What changed

Stemming maps word variants to one stem, so "runs" also matches "running". Stems aren't always real words: the stem of "database" is "databas". TIN uses the rust_stemmers crate, which runs Snowball algorithms for 18 languages, including Arabic, Russian, Tamil and Turkish. Stemming changes what goes into the index, so you have to rebuild existing indexes (WITH (stemmer = 'en')) after updating to at least v1.0.4.

Tiebreakers. A query like ORDER BY tin.score(ctid) DESC, created_at DESC, id used to score every matching document. Now it uses the same block-max top-k plan as a score-only sort. The post credits TIN's on-disk format and SIMD for the general top-k speedups but doesn't explain them.

The benchmark

The setup is the same as ParadeDB's: 85 GB of Stack Exchange data (150M documents) and 1,254 queries, each run as a conjunction, a disjunction and a phrase query. The server ran in a CPU-pinned 8-vCPU, 64 GB container on an i7i.8xlarge (Xeon), with an i8g (Graviton4) for ARM. Data was read-only and freshly vacuumed, which Reynolds calls a best case. The index was about twice the size of the 24 GB shared_buffers.

One caveat: by default, TIN's scores ignore terms that appear in more than 10% of documents. The terms still count for matching. TIN_FULL scores every term and is the fairer comparison. On the mixed workload on x86:

Anthropic released Claude Haiku 5.5 on Oct 7, just under a year after Haiku 4.5. It is priced exactly like OpenAI's GPT-6 Luna. Anthropic is pitching it as a worker model that runs under Opus 5.5 or Sonnet 5.5, handling high-volume jobs such as summaries, context compaction and database queries.

Pricing and specs

The price depends on prompt length. Under 100K tokens, input costs $0.10, output $0.50 and cache reads $0.01 per 1M tokens. Above 100K, each of those is 5x higher ($0.50 / $2.50 / $0.05). Anthropic says Haiku 5.5 costs about 75% less to run than Haiku 4.5 on average. Cursor says it is "10x less" on short requests. The context window grows from 200K to 1M tokens, max output is about 12K, and it takes text and images as input. It is the first Haiku with effort settings and adaptive thinking.

On the same day, Anthropic halved Sonnet 5.5 cache reads from $0.20 to $0.10 per 1M tokens, which it says cuts about 20% off agentic workloads. Max and Team subscriptions now include monthly API credits ($100 to $500). The Python and TypeScript SDKs now run the computer-use and browser-use action loop themselves and send actions to drivers such as Browserbase, E2B or Daytona.

Independent numbers

Artificial Analysis scores it 43 on its Intelligence Index at max effort. That is 26 points above the previous Haiku and ahead of GLM-5.3 Flash (42), Gemini 3.8 Flash (41) and GPT-6 Luna (38), but 13 points behind Sonnet 5.5. It scores 33% on Terminal-Bench 4.0, up from 0%, compared with 13% for Luna. Anthropic reports 39% on TerminalBench and says OSWorld rose from 15% to 72%.

LegalOn Technologies sells AI tools for legal and business work. It first gave its developers unlimited Codex access to GPT‑5.5 in Fast mode, for design, implementation and everyday work. Spending was on track to exceed the annual budget. An internal group, the AI-powered Development CoE (AID CoE), then wrote guidelines for picking a model, tested and monitored the models, and passed what it learned to managers. Engineers now choose a model for each task themselves.

The routing scheme

The default changed from "most capable model for everything" to "start light, move up as complexity rises." After internal testing, the tiers settled like this:

• GPT‑6 Luna (lightest): code implementation when requirements are clear, everyday automations and simple analysis. It runs mostly as a subagent.
• GPT‑6.1 Sol: standard and fairly simple software design, routine numerical analysis and documents. It is also the pick for tasks that need to finish faster than Luna would.
• GPT‑6 Astra (most capable): complex design and analysis, architecture decisions, and acting as the orchestrator that coordinates other agents.

This is the familiar planner/executor split: an expensive model plans and delegates, and cheap models write the code.

Controls on top of routing

The company added three more levers:

1. Fast mode is off by default. Engineers have to request it for individual tasks. Teams worried about losing speed and say they made up for it partly by running tasks in parallel.

Hugging Face's ML Intern is an agent mode in HuggingChat. You describe a model you want, and it plans the work, asks for a budget, runs a small test, then trains, evaluates and publishes to the Hub on Hugging Face hardware. In this post, the author describes six models built with it over a few days, at a total compute cost of about USD 103. Each model card includes its evaluation.

How the prompts are written

The author says the first prompt is where the real work happens. Prompts grew from about 450 words to about 2,000 over six projects, and all of them are published on GitHub. Each one names the exact dataset, base model and training script. A section headed "Verified facts, do not re-derive" stops the agent from spending budget on things the author already knows. Two lines matter most:

• A zero-shot baseline before training, so there is a measured gain to report.
• A smoke test with a check. For example, run 50 steps and confirm the saved weights actually changed before paying for the full run.

The agent starts every session with a zero-dollar budget and needs permission for paid jobs. That lets an instruction like "cap at USD 12 and ask before exceeding" act as a hard limit.

Cloudflare Radar launched in 2020 and publishes real-time Internet trends drawn from Cloudflare's network, which the company says sits in front of roughly 20% of web traffic. Its users have mostly been specialists such as network operators and academic researchers. This redesign, described by Sophia Alfred, aims to make Radar easier for journalists, human rights advocates, policymakers and general users without removing the depth experts rely on.

What was wrong with the old page

The team names three problems with the previous landing page. Its bento-box layout gave every widget roughly the same visual weight. Its vertical cards broke natural reading order. And its styling didn't match the rest of Cloudflare's brand. One external user said the site was hard to show to his students, which the team took as a sign that Radar intimidated newcomers. Alfred argues that a consistent look builds trust: if Radar feels like part of Cloudflare, it reads as authoritative, and familiar layouts make complex data easier to move through.