Issue 02  ·  14–18 November 2026

Last issue, I wanted to see a cheaper model handle real work without handing the savings back in rework, and a robot do something useful beyond a demonstration. This week brought a partial answer to both. It also gave us a good reason not to make “always use the cheapest model” company policy.

When is Astra worth four times the price of Sol?

On 17 November, OpenAI released GPT-6.2 Astra and GPT-6.2 Sol in its API and began rolling them out in ChatGPT Work and Codex. Look in the Work model picker, not ordinary Chat. Workspace permissions can affect what appears. Astra is the interesting choice when a deliverable has to match a template, reconcile several sources and survive someone else’s review. Sol gets the same published token rates as its predecessor; it’s where I’d start for routine work.

One identical workload, two model bills
GPT-6.2 Sol
$0.18
$2 input / $10 output per million tokens
GPT-6.2 Astra
$0.72
$8 input / $40 output per million tokens

Illustration, not a measured task: 40,000 uncached input tokens and 10,000 billed output tokens each. US dollars, standard short-context API rates; no tool fees or retries.

That 54-cent difference buys 32 seconds of a reviewer’s time at an illustrative $60 an hour. If Astra saves an analyst one trip back to a spreadsheet to fix a chart, it can be the cheaper choice. If both outputs go straight through, Sol wins. ClickUp offers an early data point: in a blind comparison against its own production model on apps, dashboards and slides, it reported about 30% fewer review rounds with Astra. ClickUp didn’t identify that model as Sol, so this isn’t a head-to-head result for the two prices above.

One detail in OpenAI’s launch evaluation would change how I commissioned a report. In two tests the new models were more likely than GPT-6 Astra to leave out a blocker, such as a missing file. OpenAI says new Codex instructions reduced that behaviour. For a board pack, ask for the workbook and the cells behind the headline number. A well-formatted slide is a poor way to discover that the source never opened.

Microsoft’s specialist has a good case. Read the small print.

Microsoft shipped MAI Code 2 and Code 2 Flash in GitHub Copilot on 17 November. Flash is available even on Copilot Free; the larger model starts on Pro. The split makes sense for a maintainer. Explaining an unfamiliar function or fixing a label should not command the same budget as tracing a regression across three services.

Microsoft’s four-test comparison puts Code 2 at 98% of GPT-6.1 Sol’s average performance for 40% of the cost. That’s a relative score, not 98% of tickets fixed. Microsoft ran the models at maximum reasoning effort, compared against 6.1 Sol, and used a 50% promotional price for Code 2. I’d try Flash on a queue of small fixes before changing the default for a whole engineering team. The useful number is what it costs to get a change a maintainer will merge, including the patches they send back.

There’s a better illustration of why specialisation can work. In a Microsoft-reported evaluation at Alaska Airlines, tuning Code 2 Flash improved its accuracy in classifying aircraft maintenance reports from 81.2% to 91.1%. Qualified staff still make the decisions. A team with years of labelled reports has something useful to teach a smaller model. Microsoft’s proposed licence would let customers run and adapt models in their own Azure environment, controlling when they upgrade. That offer is listed as coming soon through customer engagements; it isn’t an open-weight download.

At CVS, the useful agent knew when to stop improvising

On 16 November, OpenAI described two different kinds of healthcare evidence. Its demonstration used synthetic patient records: an agent prepared an MRI authorisation request in an insurer’s portal for a specialist to review. At CVS Health, by contrast, a production agent reportedly cleared a months-old queue of documents waiting to be sent in a single weekend. The first is evidence that the steps can be connected. The second describes work that actually left a queue.

CVS’s published starter kit explains a choice other teams can borrow. An agent explores an interface to find the steps; a developer can then turn the successful sequence into code that repeats it. Screenshots and action records let staff check what happened. Where a referral has already been approved and merely needs recording in a second system, paying a model to reason afresh about each button buys little. Keep people on the questions that can change a patient’s care.

The economics are promising, but use the right denominator. On Stanford’s HealthAdminBench, OpenAI reports that GPT-6 Luna earned 91.8% of the available rubric points for 8.5 US cents per task, versus 94.3% for GPT-6 Astra at the same maximum reasoning setting. These were the earlier GPT-6 models, not the new 6.2 release. A form can earn most of its points and still be missing the attachment an insurer requires. I’d want the cost per complete submission and the number returned to staff before buying on that result.

Two robot stories, and two very different stopwatches

Google’s new Robotics plugin for Antigravity puts hardware setup, simulation and robot control in one workspace for Google AI Pro and Ultra users. Gemini 4 Hydrogen was announced last week; the plugin is the 17 November news. In Google’s demonstration, a developer went from a sketch of a snack-sorting conveyor to a real arm sorting snacks in about an hour. They rehearsed the picks in simulation and adjusted the gripper on the real setup. For a small team, that shortens the distance between a promising idea and finding out where it fails. The stopwatch measured time to a working prototype.

GlobalFoundries and Boston Dynamics reported a different measure on the same day. Since April, two Spot robots inspecting the service areas beneath a Singapore chip fab travelled almost 900 kilometres and completed more than 90,000 inspections. The companies report 99% mission success. That tells us about completing routes; it doesn’t tell us what fraction of equipment faults the robots caught. The new agreement expands Spot deployments across four sites, including Singapore. Testing of the Atlas humanoid is planned for early 2027. The robot GlobalFoundries is expanding today is the four-legged inspector.

Reporting window: after 06:21 UTC on 14 November through 06:57 UTC on 18 November 2026.

Share this post
The link has been copied!