Skip to content
wildcard

InsightsAI news4 min read

AI weekly: price steps, stable harnesses, week of 2026-10-02

GPT-6.1 Sol at a fifth of Astra's price, Sonnet 5.5 up to 30% cheaper, Pi 1.0 stable, sandboxes per cloud. What a DACH engineering leader does with it.

Cliff des Ligneris

The price of near-frontier work dropped a step this week, twice. GPT-6.1 Sol at a fifth of Astra’s price and Sonnet 5.5 at up to 30% less are not research news; they are a line in your budget that is now wrong. If I ran engineering at a DACH product company, Monday would have one job: re-run the eval set on the cheaper models and move every task that passes.

Top highlights at a glance

  • OpenAI’s DevDay brought GPT-6.1 Sol, which OpenAI prices at one-fifth of GPT-6 Astra’s standard API token prices for what it calls near-Astra results.
  • Anthropic released Claude Sonnet 5.5: 30% faster and up to 30% cheaper than Sonnet 5 for most work, by Anthropic’s count. ThursdAI reports it beats Opus on Terminal-Bench.
  • The agent harness is settling into infrastructure. Pi reached a stable 1.0, and OpenAI’s Cookbook added MicroVM sandbox examples for the Agents API.

This is the first briefing in the series, so the three themes below are baselines I will track week over week.

  1. Near-frontier capability drops a price step a week. GPT-6.1 Sol at a fifth of Astra, Sonnet 5.5 up to 30% cheaper, and ThursdAI’s 24 September issue already counted Opus 5.5 at 40% cheaper and GPT-6 at half price. A cost model older than a month is wrong.
  2. Harnesses are becoming infrastructure. Pi 1.0 and Pi Durable reach a stable release, OpenAI publishes sandbox examples per cloud provider, and Anthropic’s Claude Code team talks about mods, plugins and projects. The loop around the model is where the engineering happens now.
  3. Narrow models for narrow decisions. TypeSafe’s Jev returns typed values instead of text, Sebastian Raschka measures it against the history of text classifiers, and practitioners route small decisions to it while the coding agent keeps the hard work.

Models, harnesses and technical frameworks

OpenAI DevDay 2026 and GPT-6.1 Sol

Source: https://openai.com/index/introducing-gpt-6-1-sol and https://openai.com/index/devday-2026-recap. Category: Models.

OpenAI announced more than 20 items at DevDay on 29 September. GPT-6.1 Sol is positioned for coding, computer use and professional work at one-fifth of Astra’s standard API input and output token prices, “near-Astra” in OpenAI’s own words. The recap also lists an Agents API and Dots, proactive assistants that keep working across projects and everyday tasks.

Why it matters: if your default model is Astra-class, a fifth of the price for most tasks is a budget decision, not a research question. Re-run your evals on Sol before the next sprint.

Claude Sonnet 5.5

Source: https://www.anthropic.com/claude-sonnet-5-5 and https://sub.thursdai.news/p/thursdai-oct-1-openai-joins-the-assistant. Category: Models.

Anthropic released Sonnet 5.5 on 28 September and describes it as a clear upgrade over Sonnet 5 that runs 30% faster and costs up to 30% less for most work. ThursdAI’s 1 October issue reports Sonnet 5.5 beating Opus on Terminal-Bench.

Why it matters: the mid-tier model now does what the top tier did a quarter ago. Teams paying Opus prices for routine agent steps should check whether Sonnet 5.5 passes their evals.

Pi 1.0 and Pi Durable

Source: https://www.latent.space/p/ainews-pi-10-pi-durable-and-aie-nyc. Category: Agent harnesses.

Latent Space reports that Pi, the minimalist agent harness, reached a stable 1.0 release with a TypeScript implementation, alongside Pi Durable.

Why it matters: a stable, small harness is something a team can read in an afternoon and own. That beats a framework nobody on the team understands.

Agents API sandboxes in MicroVMs

Source: https://github.com/openai/openai-cookbook/commit/86ff2cc0c4b2417feb36bc76ff0f24c5c4e785be and https://github.com/openai/openai-cookbook/commit/6dc6324fb9ed780b32b787f23fad336e9f1eff15. Category: Agent harnesses.

On 29 September OpenAI’s Cookbook added AWS Lambda MicroVM sandbox examples for the Agents API, three days after reorganising its sandbox examples by cloud provider.

Why it matters: agent tool execution belongs in an isolated sandbox, and the vendor now ships reference code per cloud. Nobody needs to invent this for a pilot.

Jev: typed outputs as a classification layer

Source: https://magazine.sebastianraschka.com/p/classifier-history-and-jev and https://www.lennysnewsletter.com/p/jev-for-beginners-how-to-use-it-and. Category: Frameworks.

TypeSafe’s Jev returns type-safe structured values instead of text. Sebastian Raschka places it at the end of a history of text classifiers, from bag-of-words through RNNs, CNNs and transformers, with hands-on experiments on accuracy, calibration and efficiency. Lenny’s Newsletter calls it the fastest and cheapest classification layer its author has put in a pipeline, with uses from deduplication to agent routing.

Why it matters: not every decision in a workflow needs a frontier model. A typed small model for routing and classification cuts cost and makes outputs testable.

Product and engineering strategy

DHH declares the end of coding by hand

Source: https://blog.pragmaticengineer.com/the-pulse-ror-creator-sparks-new-death-of-coding-by-hand-debate/.

At Rails World, David Heinemeier Hansson declared the end of writing code by hand for professional work, at 37signals at least. Key takeaway: agree or not, your hiring, review and onboarding processes assume the opposite. Decide which of them to change this quarter.

OpenAI’s model guide for the GPT-6 family

Source: https://openai.com/index/practical-guide-building-gpt-6.

OpenAI’s own guidance is to choose a model per task, tune reasoning effort, invest in prompts and skills, coordinate tools, and prepare workflows for production. Key takeaway: model routing is now vendor-recommended practice, not an optimisation your team has to argue for.

Back to insights

Talk to us.

Thirty minutes. We talk about your situation and you leave with a next step or an honest no.

Book a 30-minute intro